first artificial intelligence research paper

• History & Society
• Science & Tech
• Biographies
• Animals & Nature
• Geography & Travel
• Arts & Culture
• Games & Quizzes
• On This Day
• One Good Fact
• New Articles
• Lifestyles & Social Issues
• Philosophy & Religion
• Politics, Law & Government
• World History
• Health & Medicine
• Browse Biographies
• Birds, Reptiles & Other Vertebrates
• Bugs, Mollusks & Other Invertebrates
• Environment
• Fossils & Geologic Time
• Entertainment & Pop Culture
• Sports & Recreation
• Visual Arts
• Demystified
• Image Galleries
• Infographics
• Top Questions
• Britannica Kids
• Saving Earth
• Space Next 50
• Student Center
• Introduction

Theoretical work

The turing test, the first ai programs, evolutionary computing, logical reasoning and problem solving, english dialogue, ai programming languages, microworld programs, knowledge and inference, the cyc project.

• Creating an artificial neural network
• Perceptrons
• Conjugating verbs
• Other neural networks
• New foundations
• The situated approach

• Is Internet technology "making us stupid"?
• What is the impact of artificial intelligence (AI) technology on society?
• What is a computer?
• Who invented the computer?
• What can computers do?

history of artificial intelligence (AI)

Our editors will review what you’ve submitted and determine whether to revise the article.

• Table Of Contents

history of artificial intelligence (AI) , a survey of important events and people in the field of artificial intelligence (AI) from the early work of British logician Alan Turing in the 1930s to advancements at the turn of the 21st century. AI is the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. The term is frequently applied to the project of developing systems endowed with the intellectual processes characteristic of humans, such as the ability to reason, discover meaning, generalize, or learn from past experience. For modern developments in AI, see artificial intelligence .

Alan Turing and the beginning of AI

The earliest substantial work in the field of artificial intelligence was done in the mid-20th century by the British logician and computer pioneer Alan Mathison Turing . In 1935 Turing described an abstract computing machine consisting of a limitless memory and a scanner that moves back and forth through the memory, symbol by symbol, reading what it finds and writing further symbols. The actions of the scanner are dictated by a program of instructions that also is stored in the memory in the form of symbols. This is Turing’s stored-program concept , and implicit in it is the possibility of the machine operating on, and so modifying or improving, its own program. Turing’s conception is now known simply as the universal Turing machine . All modern computers are in essence universal Turing machines.

During World War II Turing was a leading cryptanalyst at the Government Code and Cypher School in Bletchley Park , Buckinghamshire, England. Turing could not turn to the project of building a stored-program electronic computing machine until the cessation of hostilities in Europe in 1945. Nevertheless, during the war he gave considerable thought to the issue of machine intelligence. One of Turing’s colleagues at Bletchley Park, Donald Michie (who later founded the Department of Machine Intelligence and Perception at the University of Edinburgh), later recalled that Turing often discussed how computers could learn from experience as well as solve new problems through the use of guiding principles—a process now known as heuristic problem solving .

Turing gave quite possibly the earliest public lecture (London, 1947) to mention computer intelligence, saying, “What we want is a machine that can learn from experience,” and that the “possibility of letting the machine alter its own instructions provides the mechanism for this.” In 1948 he introduced many of the central concepts of AI in a report entitled “Intelligent Machinery.” However, Turing did not publish this paper, and many of his ideas were later reinvented by others. For instance, one of Turing’s original ideas was to train a network of artificial neurons to perform specific tasks, an approach described in the section Connectionism .

At Bletchley Park Turing illustrated his ideas on machine intelligence by reference to chess —a useful source of challenging and clearly defined problems against which proposed methods for problem solving could be tested. In principle, a chess-playing computer could play by searching exhaustively through all the available moves, but in practice this is impossible because it would involve examining an astronomically large number of moves. Heuristics are necessary to guide a narrower, more discriminative search. Although Turing experimented with designing chess programs, he had to content himself with theory in the absence of a computer to run his chess program. The first true AI programs had to await the arrival of stored-program electronic digital computers .

In 1945 Turing predicted that computers would one day play very good chess, and just over 50 years later, in 1997, Deep Blue , a chess computer built by IBM (International Business Machines Corporation), beat the reigning world champion, Garry Kasparov , in a six-game match. While Turing’s prediction came true, his expectation that chess programming would contribute to the understanding of how human beings think did not. The huge improvement in computer chess since Turing’s day is attributable to advances in computer engineering rather than advances in AI: Deep Blue’s 256 parallel processors enabled it to examine 200 million possible moves per second and to look ahead as many as 14 turns of play. Many agree with Noam Chomsky , a linguist at the Massachusetts Institute of Technology (MIT) , who opined that a computer beating a grandmaster at chess is about as interesting as a bulldozer winning an Olympic weightlifting competition.

In 1950 Turing sidestepped the traditional debate concerning the definition of intelligence by introducing a practical test for computer intelligence that is now known simply as the Turing test . The Turing test involves three participants: a computer, a human interrogator, and a human foil. The interrogator attempts to determine, by asking questions of the other two participants, which is the computer. All communication is via keyboard and display screen. The interrogator may ask questions as penetrating and wide-ranging as necessary, and the computer is permitted to do everything possible to force a wrong identification. (For instance, the computer might answer “No” in response to “Are you a computer?” and might follow a request to multiply one large number by another with a long pause and an incorrect answer.) The foil must help the interrogator to make a correct identification. A number of different people play the roles of interrogator and foil, and, if a sufficient proportion of the interrogators are unable to distinguish the computer from the human being , then (according to proponents of Turing’s test) the computer is considered an intelligent, thinking entity.

In 1991 the American philanthropist Hugh Loebner started the annual Loebner Prize competition, promising $100,000 to the first computer to pass the Turing test and awarding $2,000 each year to the best effort. However, no AI program has come close to passing an undiluted Turing test. In late 2022 the advent of the large language model ChatGPT reignited conversation about the likelihood that the components of the Turing test had been met. BuzzFeed data scientist Max Woolf said that ChatGPT had passed the Turing test in December 2022, but some experts claim that ChatGPT did not pass a true Turing test, because, in ordinary usage, ChatGPT often states that it is a language model.

Early milestones in AI

The earliest successful AI program was written in 1951 by Christopher Strachey, later director of the Programming Research Group at the University of Oxford . Strachey’s checkers (draughts) program ran on the Ferranti Mark I computer at the University of Manchester , England. By the summer of 1952 this program could play a complete game of checkers at a reasonable speed.

Information about the earliest successful demonstration of machine learning was published in 1952. Shopper, written by Anthony Oettinger at the University of Cambridge , ran on the EDSAC computer. Shopper’s simulated world was a mall of eight shops. When instructed to purchase an item, Shopper would search for it, visiting shops at random until the item was found. While searching, Shopper would memorize a few of the items stocked in each shop visited (just as a human shopper might). The next time Shopper was sent out for the same item, or for some other item that it had already located, it would go to the right shop straight away. This simple form of learning is called rote learning.

The first AI program to run in the United States also was a checkers program, written in 1952 by Arthur Samuel for the prototype of the IBM 701. Samuel took over the essentials of Strachey’s checkers program and over a period of years considerably extended it. In 1955 he added features that enabled the program to learn from experience. Samuel included mechanisms for both rote learning and generalization, enhancements that eventually led to his program’s winning one game against a former Connecticut checkers champion in 1962.

Samuel’s checkers program was also notable for being one of the first efforts at evolutionary computing. (His program “evolved” by pitting a modified copy against the current best version of his program, with the winner becoming the new standard.) Evolutionary computing typically involves the use of some automatic method of generating and evaluating successive “generations” of a program, until a highly proficient solution evolves.

A leading proponent of evolutionary computing, John Holland, also wrote test software for the prototype of the IBM 701 computer. In particular, he helped design a neural-network virtual rat that could be trained to navigate through a maze. This work convinced Holland of the efficacy of the bottom-up approach to AI, which involves creating neural networks in imitation of the brain’s structure. While continuing to consult for IBM , Holland moved to the University of Michigan in 1952 to pursue a doctorate in mathematics . He soon switched, however, to a new interdisciplinary program in computers and information processing (later known as communications science) created by Arthur Burks, one of the builders of ENIAC and its successor EDVAC. In his 1959 dissertation, for what was likely the world’s first computer science Ph.D., Holland proposed a new type of computer—a multiprocessor computer—that would assign each artificial neuron in a network to a separate processor. (In 1985 Daniel Hillis solved the engineering difficulties to build the first such computer, the 65,536-processor Thinking Machines Corporation supercomputer .)

Holland joined the faculty at Michigan after graduation and over the next four decades directed much of the research into methods of automating evolutionary computing, a process now known by the term genetic algorithms . Systems implemented in Holland’s laboratory included a chess program, models of single- cell biological organisms, and a classifier system for controlling a simulated gas-pipeline network. Genetic algorithms are no longer restricted to academic demonstrations, however; in one important practical application, a genetic algorithm cooperates with a witness to a crime in order to generate a portrait of the perpetrator.

The ability to reason logically is an important aspect of intelligence and has always been a major focus of AI research. An important landmark in this area was a theorem-proving program written in 1955–56 by Allen Newell and J. Clifford Shaw of the RAND Corporation and Herbert Simon of Carnegie Mellon University . The Logic Theorist, as the program became known, was designed to prove theorems from Principia Mathematica (1910–13), a three-volume work by the British philosopher-mathematicians Alfred North Whitehead and Bertrand Russell . In one instance, a proof devised by the program was more elegant than the proof given in the books.

Newell, Simon, and Shaw went on to write a more powerful program, the General Problem Solver, or GPS. The first version of GPS ran in 1957, and work continued on the project for about a decade. GPS could solve an impressive variety of puzzles using a trial and error approach. However, one criticism of GPS, and similar programs that lack any learning capability, is that the program’s intelligence is entirely secondhand, coming from whatever information the programmer explicitly includes.

Two of the best-known early AI programs, Eliza and Parry, gave an eerie semblance of intelligent conversation. (Details of both were first published in 1966.) Eliza, written by Joseph Weizenbaum of MIT’s AI Laboratory, simulated a human therapist. Parry, written by Stanford University psychiatrist Kenneth Colby, simulated a human experiencing paranoia . Psychiatrists who were asked to decide whether they were communicating with Parry or a human experiencing paranoia were often unable to tell. Nevertheless, neither Parry nor Eliza could reasonably be described as intelligent. Parry’s contributions to the conversation were canned—constructed in advance by the programmer and stored away in the computer’s memory . Eliza, too, relied on canned sentences and simple programming tricks.

In the course of their work on the Logic Theorist and GPS, Newell, Simon, and Shaw developed their Information Processing Language (IPL), a computer language tailored for AI programming. At the heart of IPL was a highly flexible data structure that they called a list. A list is simply an ordered sequence of items of data. Some or all of the items in a list may themselves be lists. This scheme leads to richly branching structures.

In 1960 John McCarthy combined elements of IPL with the lambda calculus (a formal mathematical-logical system) to produce the programming language LISP (List Processor), which for decades was the principal language for AI work in the United States, before it was supplanted in the 21st century by such languages as Python , Java , and C++ . (The lambda calculus itself was invented in 1936 by Princeton logician Alonzo Church while he was investigating the abstract Entscheidungsproblem , or “decision problem,” for predicate logic—the same problem that Turing had been attacking when he invented the universal Turing machine .)

The logic programming language PROLOG (Programmation en Logique) was conceived by Alain Colmerauer at the University of Aix-Marseille, France, where the language was first implemented in 1973. PROLOG was further developed by the logician Robert Kowalski, a member of the AI group at the University of Edinburgh . This language makes use of a powerful theorem-proving technique known as resolution, invented in 1963 at the U.S. Atomic Energy Commission’s Argonne National Laboratory in Illinois by the British logician Alan Robinson. PROLOG can determine whether or not a given statement follows logically from other given statements. For example, given the statements “All logicians are rational” and “Robinson is a logician,” a PROLOG program responds in the affirmative to the query “Robinson is rational?” PROLOG was widely used for AI work, especially in Europe and Japan.

To cope with the bewildering complexity of the real world, scientists often ignore less relevant details; for instance, physicists often ignore friction and elasticity in their models. In 1970 Marvin Minsky and Seymour Papert of the MIT AI Laboratory proposed that, likewise, AI research should focus on developing programs capable of intelligent behavior in simpler artificial environments known as microworlds. Much research has focused on the so-called blocks world, which consists of colored blocks of various shapes and sizes arrayed on a flat surface.

An early success of the microworld approach was SHRDLU, written by Terry Winograd of MIT. (Details of the program were published in 1972.) SHRDLU controlled a robot arm that operated above a flat surface strewn with play blocks. Both the arm and the blocks were virtual. SHRDLU would respond to commands typed in natural English, such as “Will you please stack up both of the red blocks and either a green cube or a pyramid.” The program could also answer questions about its own actions. Although SHRDLU was initially hailed as a major breakthrough, Winograd soon announced that the program was, in fact, a dead end. The techniques pioneered in the program proved unsuitable for application in wider, more interesting worlds. Moreover, the appearance that SHRDLU gave of understanding the blocks microworld, and English statements concerning it, was in fact an illusion . SHRDLU had no idea what a green block was.

Another product of the microworld approach was Shakey, a mobile robot developed at the Stanford Research Institute by Bertram Raphael, Nils Nilsson, and others during the period 1968–72. The robot occupied a specially built microworld consisting of walls, doorways, and a few simply shaped wooden blocks. Each wall had a carefully painted baseboard to enable the robot to “see” where the wall met the floor (a simplification of reality that is typical of the microworld approach). Shakey had about a dozen basic abilities, such as TURN, PUSH, and CLIMB-RAMP. Critics pointed out the highly simplified nature of Shakey’s environment and emphasized that, despite these simplifications, Shakey operated excruciatingly slowly; a series of actions that a human could plan out and execute in minutes took Shakey days.

The greatest success of the microworld approach is a type of program known as an expert system , described in the next section.

Expert systems

Expert systems occupy a type of microworld—for example, a model of a ship’s hold and its cargo—that is self-contained and relatively uncomplicated. For such AI systems every effort is made to incorporate all the information about some narrow field that an expert (or group of experts) would know, so that a good expert system can often outperform any single human expert. There are many commercial expert systems, including programs for medical diagnosis , chemical analysis , credit authorization, financial management, corporate planning, financial document routing, oil and mineral prospecting, genetic engineering , automobile design and manufacture, camera lens design, computer installation design, airline scheduling, cargo placement, and automatic help services for home computer owners.

The basic components of an expert system are a knowledge base, or KB, and an inference engine. The information to be stored in the KB is obtained by interviewing people who are expert in the area in question. The interviewer, or knowledge engineer, organizes the information elicited from the experts into a collection of rules, typically of an “if-then” structure. Rules of this type are called production rules. The inference engine enables the expert system to draw deductions from the rules in the KB. For example, if the KB contains the production rules “if x , then y ” and “if y , then z ,” the inference engine is able to deduce “if x , then z .” The expert system might then query its user, “Is x true in the situation that we are considering?” If the answer is affirmative, the system will proceed to infer z .

Some expert systems use fuzzy logic . In standard logic there are only two truth values, true and false. This absolute precision makes vague attributes or situations difficult to characterize. (For example, when, precisely, does a thinning head of hair become a bald head?) Often the rules that human experts use contain vague expressions, and so it is useful for an expert system’s inference engine to employ fuzzy logic.

In 1965 the AI researcher Edward Feigenbaum and the geneticist Joshua Lederberg , both of Stanford University, began work on Heuristic DENDRAL (later shortened to DENDRAL), a chemical-analysis expert system. The substance to be analyzed might, for example, be a complicated compound of carbon , hydrogen , and nitrogen . Starting from spectrographic data obtained from the substance, DENDRAL would hypothesize the substance’s molecular structure. DENDRAL’s performance rivaled that of chemists expert at this task, and the program was used in industry and in academia .

Work on MYCIN, an expert system for treating blood infections, began at Stanford University in 1972. MYCIN would attempt to diagnose patients based on reported symptoms and medical test results. The program could request further information concerning the patient, as well as suggest additional laboratory tests, to arrive at a probable diagnosis, after which it would recommend a course of treatment. If requested, MYCIN would explain the reasoning that led to its diagnosis and recommendation. Using about 500 production rules, MYCIN operated at roughly the same level of competence as human specialists in blood infections and rather better than general practitioners.

Nevertheless, expert systems have no common sense or understanding of the limits of their expertise. For instance, if MYCIN were told that a patient who had received a gunshot wound was bleeding to death, the program would attempt to diagnose a bacterial cause for the patient’s symptoms. Expert systems can also act on absurd clerical errors, such as prescribing an obviously incorrect dosage of a drug for a patient whose weight and age data were accidentally transposed.

CYC is a large experiment in symbolic AI. The project began in 1984 under the auspices of the Microelectronics and Computer Technology Corporation, a consortium of computer, semiconductor , and electronics manufacturers. In 1995 Douglas Lenat, the CYC project director, spun off the project as Cycorp, Inc., based in Austin , Texas. The most ambitious goal of Cycorp was to build a KB containing a significant percentage of the commonsense knowledge of a human being. Millions of commonsense assertions, or rules, were coded into CYC. The expectation was that this “critical mass” would allow the system itself to extract further rules directly from ordinary prose and eventually serve as the foundation for future generations of expert systems.

With only a fraction of its commonsense KB compiled, CYC could draw inferences that would defeat simpler systems. For example, CYC could infer, “Garcia is wet,” from the statement, “Garcia is finishing a marathon run,” by employing its rules that running a marathon entails high exertion, that people sweat at high levels of exertion, and that when something sweats, it is wet. Among the outstanding remaining problems are issues in searching and problem solving—for example, how to search the KB automatically for information that is relevant to a given problem. AI researchers call the problem of updating, searching, and otherwise manipulating a large structure of symbols in realistic amounts of time the frame problem. Some critics of symbolic AI believe that the frame problem is largely unsolvable and so maintain that the symbolic approach will never yield genuinely intelligent systems. It is possible that CYC, for example, will succumb to the frame problem long before the system achieves human levels of knowledge.

The brief history of artificial intelligence: the world has changed fast — what might be next?

Despite their brief history, computers and ai have fundamentally changed what we see, what we know, and what we do. little is as important for the world’s future and our own lives as how this history continues..

To see what the future might look like, it is often helpful to study our history. This is what I will do in this article. I retrace the brief history of computers and artificial intelligence to see what we can expect for the future.

How did we get here?

How rapidly the world has changed becomes clear by how even quite recent computer technology feels ancient today. Mobile phones in the ‘90s were big bricks with tiny green displays. Two decades before that, the main storage for computers was punch cards.

In a short period, computers evolved so quickly and became such an integral part of our daily lives that it is easy to forget how recent this technology is. The first digital computers were only invented about eight decades ago, as the timeline shows.

Since the early days of this history, some computer scientists have strived to make machines as intelligent as humans. The next timeline shows some of the notable artificial intelligence (AI) systems and describes what they were capable of.

The first system I mention is the Theseus. It was built by Claude Shannon in 1950 and was a remote-controlled mouse that was able to find its way out of a labyrinth and could remember its course. 1 In seven decades, the abilities of artificial intelligence have come a long way.

The language and image recognition capabilities of AI systems have developed very rapidly

The chart shows how we got here by zooming into the last two decades of AI development. The plotted data stems from a number of tests in which human and AI performance were evaluated in different domains, from handwriting recognition to language understanding.

Within each of the domains, the initial performance of the AI system is set to –100, and human performance in these tests is used as a baseline set to zero. This means that when the model’s performance crosses the zero line is when the AI system scored more points in the relevant test than the humans who did the same test. 2

Just 10 years ago, no machine could reliably provide language or image recognition at a human level. But, as the chart shows, AI systems have become steadily more capable and are now beating humans in tests in all these domains. 3

Outside of these standardized tests, the performance of these AIs is mixed. In some real-world cases, these systems are still performing much worse than humans. On the other hand, some implementations of such AI systems are already so cheap that they are available on the phone in your pocket: image recognition categorizes your photos and speech recognition transcribes what you dictate.

From image recognition to image generation

The previous chart showed the rapid advances in the perceptive abilities of artificial intelligence. AI systems have also become much more capable of generating images.

This series of nine images shows the development over the last nine years. None of the people in these images exist; all were generated by an AI system.

The series begins with an image from 2014 in the top left, a primitive image of a pixelated face in black and white. As the first image in the second row shows, just three years later, AI systems were already able to generate images that were hard to differentiate from a photograph.

In recent years, the capability of AI systems has become much more impressive still. While the early systems focused on generating images of faces, these newer models broadened their capabilities to text-to-image generation based on almost any prompt. The image in the bottom right shows that even the most challenging prompts — such as “A Pomeranian is sitting on the King’s throne wearing a crown. Two tiger soldiers are standing next to the throne” — are turned into photorealistic images within seconds. 5

Timeline of images generated by artificial intelligence 4

Language recognition and production is developing fast

Just as striking as the advances of image-generating AIs is the rapid development of systems that parse and respond to human language.

The image shows examples of an AI system developed by Google called PaLM. In these six examples, the system was asked to explain six different jokes. I find the explanation in the bottom right particularly remarkable: the AI explains an anti-joke specifically meant to confuse the listener.

AIs that produce language have entered our world in many ways over the last few years. Emails get auto-completed, massive amounts of online texts get translated, videos get automatically transcribed, school children use language models to do their homework, reports get auto-generated, and media outlets publish AI-generated journalism.

AI systems are not yet able to produce long, coherent texts. In the future, we will see whether the recent developments will slow down — or even end — or whether we will one day read a bestselling novel written by an AI.

Output of the AI system PaLM after being asked to interpret six different jokes 6

Where we are now: AI is here

These rapid advances in AI capabilities have made it possible to use machines in a wide range of new domains:

When you book a flight, it is often an artificial intelligence, no longer a human, that decides what you pay. When you get to the airport, it is an AI system that monitors what you do at the airport. And once you are on the plane, an AI system assists the pilot in flying you to your destination.

AI systems also increasingly determine whether you get a loan , are eligible for welfare, or get hired for a particular job. Increasingly, they help determine who is released from jail .

Several governments have purchased autonomous weapons systems for warfare, and some use AI systems for surveillance and oppression .

AI systems help to program the software you use and translate the texts you read. Virtual assistants , operated by speech recognition, have entered many households over the last decade. Now self-driving cars are becoming a reality.

In the last few years, AI systems have helped to make progress on some of the hardest problems in science.

Large AIs called recommender systems determine what you see on social media, which products are shown to you in online shops, and what gets recommended to you on YouTube. Increasingly they are not just recommending the media we consume, but based on their capacity to generate images and texts, they are also creating the media we consume.

Artificial intelligence is no longer a technology of the future; AI is here, and much of what is reality now would have looked like sci-fi just recently. It is a technology that already impacts all of us, and the list above includes just a few of its many applications .

The wide range of listed applications makes clear that this is a very general technology that can be used by people for some extremely good goals — and some extraordinarily bad ones, too. For such “dual-use technologies”, it is important that all of us develop an understanding of what is happening and how we want the technology to be used.

Just two decades ago, the world was very different. What might AI technology be capable of in the future?

What is next?

The AI systems that we just considered are the result of decades of steady advances in AI technology.

The big chart below brings this history over the last eight decades into perspective. It is based on the dataset produced by Jaime Sevilla and colleagues. 7

The rise of artificial intelligence over the last 8 decades: As training computation has increased, AI systems have become more powerful 8

Each small circle in this chart represents one AI system. The circle’s position on the horizontal axis indicates when the AI system was built, and its position on the vertical axis shows the amount of computation used to train the particular AI system.

Training computation is measured in floating point operations , or FLOP for short. One FLOP is equivalent to one addition, subtraction, multiplication, or division of two decimal numbers.

All AI systems that rely on machine learning need to be trained, and in these systems, training computation is one of the three fundamental factors that are driving the capabilities of the system. The other two factors are the algorithms and the input data used for the training. The visualization shows that as training computation has increased, AI systems have become more and more powerful.

The timeline goes back to the 1940s when electronic computers were first invented. The first shown AI system is ‘Theseus’, Claude Shannon’s robotic mouse from 1950 that I mentioned at the beginning. Towards the other end of the timeline, you find AI systems like DALL-E and PaLM; we just discussed their abilities to produce photorealistic images and interpret and generate language. They are among the AI systems that used the largest amount of training computation to date.

The training computation is plotted on a logarithmic scale so that from each grid line to the next, it shows a 100-fold increase. This long-run perspective shows a continuous increase. For the first six decades, training computation increased in line with Moore’s Law , doubling roughly every 20 months. Since about 2010, this exponential growth has sped up further, to a doubling time of just about 6 months. That is an astonishingly fast rate of growth. 9

The fast doubling times have accrued to large increases. PaLM’s training computation was 2.5 billion petaFLOP, more than 5 million times larger than AlexNet, the AI with the largest training computation just 10 years earlier. 10

Scale-up was already exponential and has sped up substantially over the past decade. What can we learn from this historical development for the future of AI?

Studying the long-run trends to predict the future of AI

AI researchers study these long-term trends to see what is possible in the future. 11

Perhaps the most widely discussed study of this kind was published by AI researcher Ajeya Cotra. She studied the increase in training computation to ask at what point the computation to train an AI system could match that of the human brain. The idea is that, at this point, the AI system would match the capabilities of a human brain. In her latest update, Cotra estimated a 50% probability that such “transformative AI” will be developed by the year 2040, less than two decades from now. 12

In a related article , I discuss what transformative AI would mean for the world. In short, the idea is that such an AI system would be powerful enough to bring the world into a ‘qualitatively different future’. It could lead to a change at the scale of the two earlier major transformations in human history, the agricultural and industrial revolutions. It would certainly represent the most important global change in our lifetimes.

Cotra’s work is particularly relevant in this context as she based her forecast on the kind of historical long-run trend of training computation that we just studied. But it is worth noting that other forecasters who rely on different considerations arrive at broadly similar conclusions. As I show in my article on AI timelines , many AI experts believe that there is a real chance that human-level artificial intelligence will be developed within the next decades, and some believe that it will exist much sooner.

Building a public resource to enable the necessary public conversation

Computers and artificial intelligence have changed our world immensely, but we are still in the early stages of this history. Because this technology feels so familiar, it is easy to forget that all of these technologies we interact with are very recent innovations and that the most profound changes are yet to come.

Artificial intelligence has already changed what we see, what we know, and what we do. This is despite the fact that this technology has had only a brief history.

There are no signs that these trends are hitting any limits anytime soon. On the contrary, particularly over the course of the last decade, the fundamental trends have accelerated: investments in AI technology have rapidly increased , and the doubling time of training computation has shortened to just six months.

All major technological innovations lead to a range of positive and negative consequences. This is already true of artificial intelligence. As this technology becomes more and more powerful, we should expect its impact to still increase.

Because of the importance of AI, we should all be able to form an opinion on where this technology is heading and understand how this development is changing our world. For this purpose, we are building a repository of AI-related metrics, which you can find on OurWorldinData.org/artificial-intelligence .

We are still in the early stages of this history, and much of what will become possible is yet to come. A technological development as powerful as this should be at the center of our attention. Little might be as important for how the future of our world — and the future of our lives — will play out.

Acknowledgments: I would like to thank my colleagues Natasha Ahuja, Daniel Bachler, Julia Broden, Charlie Giattino, Bastian Herre, Edouard Mathieu, and Ike Saunders for their helpful comments to drafts of this essay and their contributions in preparing the visualizations.

On the Theseus see Daniel Klein (2019) — Mighty mouse , Published in MIT Technology Review. And this video on YouTube of a presentation by its inventor Claude Shannon.

The chart shows that the speed at which these AI technologies developed increased over time. Systems for which development was started early — handwriting and speech recognition — took more than a decade to approach human-level performance, while more recent AI developments led to systems that overtook humans in only a few years. However, one should not overstate this point. To some extent, this is dependent on when the researchers started to compare machine and human performance. One could have started evaluating the system for language understanding much earlier, and its development would appear much slower in this presentation of the data.

It is important to remember that while these are remarkable achievements — and show very rapid gains — these are the results from specific benchmarking tests. Outside of tests, AI models can fail in surprising ways and do not reliably achieve performance that is comparable with human capabilities.

The relevant publications are the following:

2014: Goodfellow et al.: Generative Adversarial Networks

2015: Radford, Metz, and Chintala: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

2016: Liu and Tuzel: Coupled Generative Adversarial Networks

2017: Karras et al.: Progressive Growing of GANs for Improved Quality, Stability, and Variation

2018: Karras, Laine, and Aila: A Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN from NVIDIA)

2019: Karras et al.: Analyzing and Improving the Image Quality of StyleGAN

AI-generated faces generated by this technology can be found on thispersondoesnotexist.com .

2020: Ho, Jain, and Abbeel: Denoising Diffusion Probabilistic Models

2021: Ramesh et al: Zero-Shot Text-to-Image Generation (first DALL-E from OpenAI; blog post ). See also Ramesh et al. (2022) — Hierarchical Text-Conditional Image Generation with CLIP Latents (DALL-E 2 from OpenAI; blog post ).

2022: Saharia et al: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (Google’s Imagen; blog post )

Because these systems have become so powerful, the latest AI systems often don’t allow the user to generate images of human faces to prevent abuse.

From Chowdhery et al. (2022) —  PaLM: Scaling Language Modeling with Pathways . Published on arXiv on 7 Apr 2022.

See the footnote on the chart's title for the references and additional information.

The data is taken from Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, Pablo Villalobos (2022) — Compute Trends Across Three eras of Machine Learning . Published in arXiv on March 9, 2022. See also their post on the Alignment Forum .

The authors regularly update and extend their dataset, a helpful service to the AI research community. At Our World in Data, my colleague Charlie Giattino regularly updates the interactive version of this chart with the latest data made available by Sevilla and coauthors.

See also these two related charts:

Number of parameters in notable artificial intelligence systems

Number of datapoints used to train notable artificial intelligence systems

At some point in the future, training computation is expected to slow down to the exponential growth rate of Moore's Law. Tamay Besiroglu, Lennart Heim, and Jaime Sevilla of the Epoch team estimate in their report that the highest probability for this reversion occurring is in the early 2030s.

The training computation of PaLM, developed in 2022, was 2,700,000,000 petaFLOP. The training computation of AlexNet, the AI with the largest training computation up to 2012, was 470 petaFLOP. 2,500,000,000 petaFLOP / 470 petaFLOP = 5,319,148.9. At the same time, the amount of training computation required to achieve a given performance has been falling exponentially.

The costs have also increased quickly. The cost to train PaLM is estimated to be $9–$23 million, according to Lennart Heim, a researcher in the Epoch team. See Lennart Heim (2022) — Estimating PaLM's training cost .

Scaling up the size of neural networks — in terms of the number of parameters and the amount of training data and computation — has led to surprising increases in the capabilities of AI systems. This realization motivated the “scaling hypothesis.” See Gwern Branwen (2020) — The Scaling Hypothesis ⁠.

Her research was announced in various places, including in the AI Alignment Forum here: Ajeya Cotra (2020) —  Draft report on AI timelines . As far as I know, the report always remained a “draft report” and was published here on Google Docs .

The cited estimate stems from Cotra’s Two-year update on my personal AI timelines , in which she shortened her median timeline by 10 years.

Cotra emphasizes that there are substantial uncertainties around her estimates and therefore communicates her findings in a range of scenarios. She published her big study in 2020, and her median estimate at the time was that around the year 2050, there will be a 50%-probability that the computation required to train such a model may become affordable. In her “most conservative plausible”-scenario, this point in time is pushed back to around 2090, and in her “most aggressive plausible”-scenario, this point is reached in 2040.

The same is true for most other forecasters: all emphasize the large uncertainty associated with their forecasts .

It is worth emphasizing that the computation of the human brain is highly uncertain. See Joseph Carlsmith's New Report on How Much Computational Power It Takes to Match the Human Brain from 2020.

Cite this work

Our articles and data visualizations rely on work from many different people and organizations. When citing this article, please also cite the underlying data sources. This article can be cited as:

BibTeX citation

Reuse this work freely

All visualizations, data, and code produced by Our World in Data are completely open access under the Creative Commons BY license . You have the permission to use, distribute, and reproduce these in any medium, provided the source and authors are credited.

The data produced by third parties and made available by Our World in Data is subject to the license terms from the original third-party authors. We will always indicate the original source of the data in our documentation, so you should always check the license of any such third-party data before use and redistribution.

All of our charts can be embedded in any site.

Our World in Data is free and accessible for everyone.

Help us do this work by making a donation.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Published: 02 August 2023

Scientific discovery in the age of artificial intelligence

• Hanchen Wang   ORCID: orcid.org/0000-0002-1691-024X 1 , 2   na1   nAff37   nAff38 ,
• Tianfan Fu 3   na1 ,
• Yuanqi Du 4   na1 ,
• Wenhao Gao 5 ,
• Kexin Huang 6 ,
• Ziming Liu 7 ,
• Payal Chandak   ORCID: orcid.org/0000-0003-1097-803X 8 ,
• Shengchao Liu   ORCID: orcid.org/0000-0003-2030-2367 9 , 10 ,
• Peter Van Katwyk   ORCID: orcid.org/0000-0002-3512-0665 11 , 12 ,
• Andreea Deac 9 , 10 ,
• Anima Anandkumar 2 , 13 ,
• Karianne Bergen 11 , 12 ,
• Carla P. Gomes   ORCID: orcid.org/0000-0002-4441-7225 4 ,
• Shirley Ho 14 , 15 , 16 , 17 ,
• Pushmeet Kohli   ORCID: orcid.org/0000-0002-7466-7997 18 ,
• Joan Lasenby 1 ,
• Jure Leskovec   ORCID: orcid.org/0000-0002-5411-923X 6 ,
• Tie-Yan Liu 19 ,
• Arjun Manrai 20 ,
• Debora Marks   ORCID: orcid.org/0000-0001-9388-2281 21 , 22 ,
• Bharath Ramsundar 23 ,
• Le Song 24 , 25 ,
• Jimeng Sun 26 ,
• Jian Tang 9 , 27 , 28 ,
• Petar Veličković 18 , 29 ,
• Max Welling 30 , 31 ,
• Linfeng Zhang 32 , 33 ,
• Connor W. Coley   ORCID: orcid.org/0000-0002-8271-8723 5 , 34 ,
• Yoshua Bengio   ORCID: orcid.org/0000-0002-9322-3515 9 , 10 &
• Marinka Zitnik   ORCID: orcid.org/0000-0001-8530-7228 20 , 22 , 35 , 36  

Nature volume  620 ,  pages 47–60 ( 2023 ) Cite this article

114k Accesses

298 Citations

602 Altmetric

Metrics details

• Computer science
• Machine learning
• Scientific community

A Publisher Correction to this article was published on 30 August 2023

This article has been updated

Artificial intelligence (AI) is being increasingly integrated into scientific discovery to augment and accelerate research, helping scientists to generate hypotheses, design experiments, collect and interpret large datasets, and gain insights that might not have been possible using traditional scientific methods alone. Here we examine breakthroughs over the past decade that include self-supervised learning, which allows models to be trained on vast amounts of unlabelled data, and geometric deep learning, which leverages knowledge about the structure of scientific data to enhance model accuracy and efficiency. Generative AI methods can create designs, such as small-molecule drugs and proteins, by analysing diverse data modalities, including images and sequences. We discuss how these methods can help scientists throughout the scientific process and the central issues that remain despite such advances. Both developers and users of AI tools need a better understanding of when such approaches need improvement, and challenges posed by poor data quality and stewardship remain. These issues cut across scientific disciplines and require developing foundational algorithmic approaches that can contribute to scientific understanding or acquire it autonomously, making them critical areas of focus for AI innovation.

This is a preview of subscription content, access via your institution

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 51 print issues and online access

185,98 € per year

only 3,65 € per issue

Buy this article

• Purchase on SpringerLink
• Instant access to full article PDF

Prices may be subject to local taxes which are calculated during checkout

Similar content being viewed by others

Accelerating science with human-aware artificial intelligence

Accelerating material design with the generative toolkit for scientific discovery

A data science roadmap for open science organizations engaged in early-stage drug discovery

Change history, 30 august 2023.

A Correction to this paper has been published: https://doi.org/10.1038/s41586-023-06559-7

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521 , 436–444 (2015). This survey summarizes key elements of deep learning and its development in speech recognition, computer vision and and natural language processing .

Article   ADS   CAS   PubMed   Google Scholar  

de Regt, H. W. Understanding, values, and the aims of science. Phil. Sci. 87 , 921–932 (2020).

Article   MathSciNet   Google Scholar  

Pickstone, J. V. Ways of Knowing: A New History of Science, Technology, and Medicine (Univ. Chicago Press, 2001).

Han, J. et al. Deep potential: a general representation of a many-body potential energy surface. Commun. Comput. Phys. 23 , 629–639 (2018). This paper introduced a deep neural network architecture that learns the potential energy surface of many-body systems while respecting the underlying symmetries of the system by incorporating group theory.

Akiyama, K. et al. First M87 Event Horizon Telescope results. IV. Imaging the central supermassive black hole. Astrophys. J. Lett. 875 , L4 (2019).

Article   ADS   CAS   Google Scholar  

Wagner, A. Z. Constructions in combinatorics via neural networks. Preprint at https://arxiv.org/abs/2104.14516 (2021).

Coley, C. W. et al. A robotic platform for flow synthesis of organic compounds informed by AI planning. Science 365 , eaax1566 (2019).

Article   CAS   PubMed   Google Scholar  

Bommasani, R. et al. On the opportunities and risks of foundation models. Preprint at https://arxiv.org/abs/2108.07258 (2021).

Davies, A. et al. Advancing mathematics by guiding human intuition with AI. Nature 600 , 70–74 (2021). This paper explores how AI can aid the development of pure mathematics by guiding mathematical intuition.

Article   ADS   CAS   PubMed   PubMed Central   MATH   Google Scholar  

Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596 , 583–589 (2021). This study was the first to demonstrate the ability to predict protein folding structures using AI methods with a high degree of accuracy, achieving results that are at or near the experimental resolution. This accomplishment is particularly noteworthy, as predicting protein folding has been a grand challenge in the field of molecular biology for over 50 years.

Article   ADS   CAS   PubMed   PubMed Central   Google Scholar  

Stokes, J. M. et al. A deep learning approach to antibiotic discovery. Cell 180 , 688–702 (2020).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Bohacek, R. S., McMartin, C. & Guida, W. C. The art and practice of structure-based drug design: a molecular modeling perspective. Med. Res. Rev. 16 , 3–50 (1996).

Bileschi, M. L. et al. Using deep learning to annotate the protein universe. Nat. Biotechnol. 40 , 932–937 (2022).

Bellemare, M. G. et al. Autonomous navigation of stratospheric balloons using reinforcement learning. Nature 588 , 77–82 (2020). This paper describes a reinforcement-learning algorithm for navigating a super-pressure balloon in the stratosphere, making real-time decisions in the changing environment.

Tshitoyan, V. et al. Unsupervised word embeddings capture latent knowledge from materials science literature. Nature 571 , 95–98 (2019).

Zhang, L. et al. Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics. Phys. Rev. Lett. 120 , 143001 (2018).

Deiana, A. M. et al. Applications and techniques for fast machine learning in science. Front. Big Data 5 , 787421 (2022).

Karagiorgi, G. et al. Machine learning in the search for new fundamental physics. Nat. Rev. Phys. 4 , 399–412 (2022).

Zhou, C. & Paffenroth, R. C. Anomaly detection with robust deep autoencoders. In ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 665–674 (2017).

Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science 313 , 504–507 (2006).

Article   ADS   MathSciNet   CAS   PubMed   MATH   Google Scholar  

Kasieczka, G. et al. The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics. Rep. Prog. Phys. 84 , 124201 (2021).

Govorkova, E. et al. Autoencoders on field-programmable gate arrays for real-time, unsupervised new physics detection at 40 MHz at the Large Hadron Collider. Nat. Mach. Intell. 4 , 154–161 (2022).

Article   Google Scholar  

Chamberland, M. et al. Detecting microstructural deviations in individuals with deep diffusion MRI tractometry. Nat. Comput. Sci. 1 , 598–606 (2021).

Article   PubMed   PubMed Central   Google Scholar  

Rafique, M. et al. Delegated regressor, a robust approach for automated anomaly detection in the soil radon time series data. Sci. Rep. 10 , 3004 (2020).

Pastore, V. P. et al. Annotation-free learning of plankton for classification and anomaly detection. Sci. Rep. 10 , 12142 (2020).

Naul, B. et al. A recurrent neural network for classification of unevenly sampled variable stars. Nat. Astron. 2 , 151–155 (2018).

Article   ADS   Google Scholar  

Lee, D.-H. et al. Pseudo-label: the simple and efficient semi-supervised learning method for deep neural networks. In ICML Workshop on Challenges in Representation Learning (2013).

Zhou, D. et al. Learning with local and global consistency. In Advances in Neural Information Processing Systems 16 , 321–328 (2003).

Radivojac, P. et al. A large-scale evaluation of computational protein function prediction. Nat. Methods 10 , 221–227 (2013).

Barkas, N. et al. Joint analysis of heterogeneous single-cell RNA-seq dataset collections. Nat. Methods 16 , 695–698 (2019).

Tran, K. & Ulissi, Z. W. Active learning across intermetallics to guide discovery of electrocatalysts for CO 2 reduction and H 2 evolution. Nat. Catal. 1 , 696–703 (2018).

Article   CAS   Google Scholar  

Jablonka, K. M. et al. Bias free multiobjective active learning for materials design and discovery. Nat. Commun. 12 , 2312 (2021).

Roussel, R. et al. Turn-key constrained parameter space exploration for particle accelerators using Bayesian active learning. Nat. Commun. 12 , 5612 (2021).

Ratner, A. J. et al. Data programming: creating large training sets, quickly. In Advances in Neural Information Processing Systems 29 , 3567–3575 (2016).

Ratner, A. et al. Snorkel: rapid training data creation with weak supervision. In International Conference on Very Large Data Bases 11 , 269–282 (2017). This paper presents a weakly-supervised AI system designed to annotate massive amounts of data using labeling functions.

Butter, A. et al. GANplifying event samples. SciPost Phys. 10 , 139 (2021).

Brown, T. et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 , 1877–1901 (2020).

Ramesh, A. et al. Zero-shot text-to-image generation. In International Conference on Machine Learning 139 , 8821–8831 (2021).

Littman, M. L. Reinforcement learning improves behaviour from evaluative feedback. Nature 521 , 445–451 (2015).

Cubuk, E. D. et al. Autoaugment: learning augmentation strategies from data. In IEEE Conference on Computer Vision and Pattern Recognition 113–123 (2019).

Reed, C. J. et al. Selfaugment: automatic augmentation policies for self-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition 2674–2683 (2021).

ATLAS Collaboration et al. Deep generative models for fast photon shower simulation in ATLAS. Preprint at https://arxiv.org/abs/2210.06204 (2022).

Mahmood, F. et al. Deep adversarial training for multi-organ nuclei segmentation in histopathology images. IEEE Trans. Med. Imaging 39 , 3257–3267 (2019).

Teixeira, B. et al. Generating synthetic X-ray images of a person from the surface geometry. In IEEE Conference on Computer Vision and Pattern Recognition 9059–9067 (2018).

Lee, D., Moon, W.-J. & Ye, J. C. Assessing the importance of magnetic resonance contrasts using collaborative generative adversarial networks. Nat. Mach. Intell. 2 , 34–42 (2020).

Kench, S. & Cooper, S. J. Generating three-dimensional structures from a two-dimensional slice with generative adversarial network-based dimensionality expansion. Nat. Mach. Intell. 3 , 299–305 (2021).

Wan, C. & Jones, D. T. Protein function prediction is improved by creating synthetic feature samples with generative adversarial networks. Nat. Mach. Intell. 2 , 540–550 (2020).

Repecka, D. et al. Expanding functional protein sequence spaces using generative adversarial networks. Nat. Mach. Intell. 3 , 324–333 (2021).

Marouf, M. et al. Realistic in silico generation and augmentation of single-cell RNA-seq data using generative adversarial networks. Nat. Commun. 11 , 166 (2020).

Ghahramani, Z. Probabilistic machine learning and artificial intelligence. Nature 521 , 452–459 (2015). This survey provides an introduction to probabilistic machine learning, which involves the representation and manipulation of uncertainty in models and predictions, playing a central role in scientific data analysis.

Cogan, J. et al. Jet-images: computer vision inspired techniques for jet tagging. J. High Energy Phys. 2015 , 118 (2015).

Zhao, W. et al. Sparse deconvolution improves the resolution of live-cell super-resolution fluorescence microscopy. Nat. Biotechnol. 40 , 606–617 (2022).

Brbić, M. et al. MARS: discovering novel cell types across heterogeneous single-cell experiments. Nat. Methods 17 , 1200–1206 (2020).

Article   PubMed   Google Scholar  

Qiao, C. et al. Evaluation and development of deep neural networks for image super-resolution in optical microscopy. Nat. Methods 18 , 194–202 (2021).

Andreassen, A. et al. OmniFold: a method to simultaneously unfold all observables. Phys. Rev. Lett. 124 , 182001 (2020).

Bergenstråhle, L. et al. Super-resolved spatial transcriptomics by deep data fusion. Nat. Biotechnol. 40 , 476–479 (2021).

Vincent, P. et al. Extracting and composing robust features with denoising autoencoders. In International Conference on Machine Learning 1096–1103 (2008).

Kingma, D. P. & Welling, M. Auto-encoding variational Bayes. In International Conference on Learning Representations (2014).

Eraslan, G. et al. Single-cell RNA-seq denoising using a deep count autoencoder. Nat. Commun. 10 , 390 (2019).

Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning (MIT Press, 2016).

Olshausen, B. A. & Field, D. J. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381 , 607–609 (1996).

Bengio, Y. Deep learning of representations for unsupervised and transfer learning. In ICML Workshop on Unsupervised and Transfer Learning (2012).

Detlefsen, N. S., Hauberg, S. & Boomsma, W. Learning meaningful representations of protein sequences. Nat. Commun. 13 , 1914 (2022).

Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. Nat. Biotechnol. 37 , 38–44 (2019).

Bronstein, M. M. et al. Geometric deep learning: going beyond euclidean data. IEEE Signal Process Mag. 34 , 18–42 (2017).

Anderson, P. W. More is different: broken symmetry and the nature of the hierarchical structure of science. Science 177 , 393–396 (1972).

Qiao, Z. et al. Informing geometric deep learning with electronic interactions to accelerate quantum chemistry. Proc. Natl Acad. Sci. USA 119 , e2205221119 (2022).

Bogatskiy, A. et al. Symmetry group equivariant architectures for physics. Preprint at https://arxiv.org/abs/2203.06153 (2022).

Bronstein, M. M. et al. Geometric deep learning: grids, groups, graphs, geodesics, and gauges. Preprint at https://arxiv.org/abs/2104.13478 (2021).

Townshend, R. J. L. et al. Geometric deep learning of RNA structure. Science 373 , 1047–1051 (2021).

Wicky, B. I. M. et al. Hallucinating symmetric protein assemblies. Science 378 , 56–61 (2022).

Kipf, T. N. & Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (2017).

Veličković, P. et al. Graph attention networks. In International Conference on Learning Representations (2018).

Hamilton, W. L., Ying, Z. & Leskovec, J. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems 30 , 1024–1034 (2017).

Gilmer, J. et al. Neural message passing for quantum chemistry. In International Conference on Machine Learning 1263–1272 (2017).

Li, M. M., Huang, K. & Zitnik, M. Graph representation learning in biomedicine and healthcare. Nat. Biomed. Eng. 6 , 1353–1369 (2022).

Satorras, V. G., Hoogeboom, E. & Welling, M. E( n ) equivariant graph neural networks. In International Conference on Machine Learning 9323–9332 (2021). This study incorporates principles of physics into the design of neural models, advancing the field of equivariant machine learning .

Thomas, N. et al. Tensor field networks: rotation-and translation-equivariant neural networks for 3D point clouds. Preprint at https://arxiv.org/abs/1802.08219 (2018).

Finzi, M. et al. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning 3165–3176 (2020).

Fuchs, F. et al. SE(3)-transformers: 3D roto-translation equivariant attention networks. In Advances in Neural Information Processing Systems 33 , 1970-1981 (2020).

Zaheer, M. et al. Deep sets. In Advances in Neural Information Processing Systems 30 , 3391–3401 (2017). This paper is an early study that explores the use of deep neural architectures on set data, which consists of an unordered list of elements.

Cohen, T. S. et al. Spherical CNNs. In International Conference on Learning Representations (2018).

Gordon, J. et al. Permutation equivariant models for compositional generalization in language. In International Conference on Learning Representations (2019).

Finzi, M., Welling, M. & Wilson, A. G. A practical method for constructing equivariant multilayer perceptrons for arbitrary matrix groups. In International Conference on Machine Learning 3318–3328 (2021).

Dijk, D. V. et al. Recovering gene interactions from single-cell data using data diffusion. Cell 174 , 716–729 (2018).

Gainza, P. et al. Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning. Nat. Methods 17 , 184–192 (2020).

Hatfield, P. W. et al. The data-driven future of high-energy-density physics. Nature 593 , 351–361 (2021).

Bapst, V. et al. Unveiling the predictive power of static structure in glassy systems. Nat. Phys. 16 , 448–454 (2020).

Zhang, R., Zhou, T. & Ma, J. Multiscale and integrative single-cell Hi-C analysis with Higashi. Nat. Biotechnol. 40 , 254–261 (2022).

Sammut, S.-J. et al. Multi-omic machine learning predictor of breast cancer therapy response. Nature 601 , 623–629 (2022).

DeZoort, G. et al. Graph neural networks at the Large Hadron Collider. Nat. Rev. Phys . 5 , 281–303 (2023).

Liu, S. et al. Pre-training molecular graph representation with 3D geometry. In International Conference on Learning Representations (2022).

The LIGO Scientific Collaboration. et al. A gravitational-wave standard siren measurement of the Hubble constant. Nature 551 , 85–88 (2017).

Reichstein, M. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566 , 195–204 (2019).

Goenka, S. D. et al. Accelerated identification of disease-causing variants with ultra-rapid nanopore genome sequencing. Nat. Biotechnol. 40 , 1035–1041 (2022).

Bengio, Y. et al. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems 19 , 153–160 (2006).

Hinton, G. E., Osindero, S. & Teh, Y.-W. A fast learning algorithm for deep belief nets. Neural Comput. 18 , 1527–1554 (2006).

Article   MathSciNet   PubMed   MATH   Google Scholar  

Jordan, M. I. & Mitchell, T. M. Machine learning: trends, perspectives, and prospects. Science 349 , 255–260 (2015).

Devlin, J. et al. BERT: pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics 4171–4186 (2019).

Rives, A. et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl Acad. Sci. USA 118 , e2016239118 (2021).

Elnaggar, A. et al. ProtTrans: rowards cracking the language of lifes code through self-supervised deep learning and high performance computing. In IEEE Transactions on Pattern Analysis and Machine Intelligence (2021).

Hie, B. et al. Learning the language of viral evolution and escape. Science 371 , 284–288 (2021). This paper modeled viral escape with machine learning algorithms originally developed for human natural language.

Biswas, S. et al. Low- N protein engineering with data-efficient deep learning. Nat. Methods 18 , 389–396 (2021).

Ferruz, N. & Höcker, B. Controllable protein design with language models. Nat. Mach. Intell. 4 , 521–532 (2022).

Hsu, C. et al. Learning inverse folding from millions of predicted structures. In International Conference on Machine Learning 8946–8970 (2022).

Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373 , 871–876 (2021). Inspired by AlphaFold2, this study reported RoseTTAFold, a novel three-track neural module capable of simultaneously processing protein’s sequence, distance and coordinates.

Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 28 , 31–36 (1988).

Lin, T.-S. et al. BigSMILES: a structurally-based line notation for describing macromolecules. ACS Cent. Sci. 5 , 1523–1531 (2019).

Krenn, M. et al. SELFIES and the future of molecular string representations. Patterns 3 , 100588 (2022).

Flam-Shepherd, D., Zhu, K. & Aspuru-Guzik, A. Language models can learn complex molecular distributions. Nat. Commun. 13 , 3293 (2022).

Skinnider, M. A. et al. Chemical language models enable navigation in sparsely populated chemical space. Nat. Mach. Intell. 3 , 759–770 (2021).

Chithrananda, S., Grand, G. & Ramsundar, B. ChemBERTa: large-scale self-supervised pretraining for molecular property prediction. In Machine Learning for Molecules Workshop at NeurIPS (2020).

Schwaller, P. et al. Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy. Chem. Sci. 11 , 3316–3325 (2020).

Tetko, I. V. et al. State-of-the-art augmented NLP transformer models for direct and single-step retrosynthesis. Nat. Commun. 11 , 5575 (2020).

Schwaller, P. et al. Mapping the space of chemical reactions using attention-based neural networks. Nat. Mach. Intell. 3 , 144–152 (2021).

Kovács, D. P., McCorkindale, W. & Lee, A. A. Quantitative interpretation explains machine learning models for chemical reaction prediction and uncovers bias. Nat. Commun. 12 , 1695 (2021).

Article   ADS   PubMed   PubMed Central   Google Scholar  

Pesciullesi, G. et al. Transfer learning enables the molecular transformer to predict regio-and stereoselective reactions on carbohydrates. Nat. Commun. 11 , 4874 (2020).

Vaswani, A. et al. Attention is all you need. In Advances in Neural Information Processing Systems 30 , 5998–6008 (2017). This paper introduced the transformer, a modern neural network architecture that can process sequential data in parallel, revolutionizing natural language processing and sequence modeling.

Mousavi, S. M. et al. Earthquake transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking. Nat. Commun. 11 , 3952 (2020).

Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18 , 1196–1203 (2021).

Meier, J. et al. Language models enable zero-shot prediction of the effects of mutations on protein function. In Advances in Neural Information Processing Systems 34 , 29287–29303 (2021).

Kamienny, P.-A. et al. End-to-end symbolic regression with transformers. In Advances in Neural Information Processing Systems 35 , 10269–10281 (2022).

Jaegle, A. et al. Perceiver: general perception with iterative attention. In International Conference on Machine Learning 4651–4664 (2021).

Chen, L. et al. Decision transformer: reinforcement learning via sequence modeling. In Advances in Neural Information Processing Systems 34 , 15084–15097 (2021).

Dosovitskiy, A. et al. An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations (2020).

Choromanski, K. et al. Rethinking attention with performers. In International Conference on Learning Representations (2021).

Li, Z. et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations (2021).

Kovachki, N. et al. Neural operator: learning maps between function spaces. J. Mach. Learn. Res. 24 , 1–97 (2023).

Russell, J. L. Kepler’s laws of planetary motion: 1609–1666. Br. J. Hist. Sci. 2 , 1–24 (1964).

Huang, K. et al. Artificial intelligence foundation for therapeutic science. Nat. Chem. Biol. 18 , 1033–1036 (2022).

Guimerà, R. et al. A Bayesian machine scientist to aid in the solution of challenging scientific problems. Sci. Adv. 6 , eaav6971 (2020).

Liu, G. et al. Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nat. Chem. Biol. https://doi.org/10.1038/s41589-023-01349-8 (2023).

Gómez-Bombarelli, R. et al. Design of efficient molecular organic light-emitting diodes by a high-throughput virtual screening and experimental approach. Nat. Mater. 15 , 1120–1127 (2016). This paper proposes using a black-box AI predictor to accelerate high-throughput screening of molecules in materials science.

Article   ADS   PubMed   Google Scholar  

Sadybekov, A. A. et al. Synthon-based ligand discovery in virtual libraries of over 11 billion compounds. Nature 601 , 452–459 (2022).

The NNPDF Collaboration Evidence for intrinsic charm quarks in the proton. Nature 606 , 483–487 (2022).

Graff, D. E., Shakhnovich, E. I. & Coley, C. W. Accelerating high-throughput virtual screening through molecular pool-based active learning. Chem. Sci. 12 , 7866–7881 (2021).

Janet, J. P. et al. Accurate multiobjective design in a space of millions of transition metal complexes with neural-network-driven efficient global optimization. ACS Cent. Sci. 6 , 513–524 (2020).

Bacon, F. Novum Organon Vol. 1620 (2000).

Schmidt, M. & Lipson, H. Distilling free-form natural laws from experimental data. Science 324 , 81–85 (2009).

Petersen, B. K. et al. Deep symbolic regression: recovering mathematical expressions from data via risk-seeking policy gradients. In International Conference on Learning Representations (2020).

Zhavoronkov, A. et al. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat. Biotechnol. 37 , 1038–1040 (2019). This paper describes a reinforcement-learning algorithm for navigating molecular combinatorial spaces, and it validates generated molecules using wet-lab experiments.

Zhou, Z. et al. Optimization of molecules via deep reinforcement learning. Sci. Rep. 9 , 10752 (2019).

You, J. et al. Graph convolutional policy network for goal-directed molecular graph generation. In Advances in Neural Information Processing Systems 31 , 6412–6422 (2018).

Bengio, Y. et al. GFlowNet foundations. Preprint at https://arxiv.org/abs/2111.09266 (2021). This paper describes a generative flow network that generates objects by sampling them from a distribution optimized for drug design.

Jain, M. et al. Biological sequence design with GFlowNets. In International Conference on Machine Learning 9786–9801 (2022).

Malkin, N. et al. Trajectory balance: improved credit assignment in GFlowNets. In Advances in Neural Information Processing Systems 35 , 5955–5967 (2022).

Borkowski, O. et al. Large scale active-learning-guided exploration for in vitro protein production optimization. Nat. Commun. 11 , 1872 (2020). This study introduced a dynamic programming approach to determine the optimal locations and capacities of hydropower dams in the Amazon Basin, balancing between energy production and environmental impact .

Flecker, A. S. et al. Reducing adverse impacts of Amazon hydropower expansion. Science 375 , 753–760 (2022). This study introduced a dynamic programming approach to determine the optimal locations and capacities of hydropower dams in the Amazon basin, achieving a balance between the benefits of energy production and the potential environmental impacts.

Pion-Tonachini, L. et al. Learning from learning machines: a new generation of AI technology to meet the needs of science. Preprint at https://arxiv.org/abs/2111.13786 (2021).

Kusner, M. J., Paige, B. & Hernández-Lobato, J. M. Grammar variational autoencoder. In International Conference on Machine Learning 1945–1954 (2017). This paper describes a grammar variational autoencoder that generates novel symbolic laws and drug molecules.

Brunton, S. L., Proctor, J. L. & Kutz, J. N. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl Acad. Sci. USA 113 , 3932–3937 (2016).

Article   ADS   MathSciNet   CAS   PubMed   PubMed Central   MATH   Google Scholar  

Liu, Z. & Tegmark, M. Machine learning hidden symmetries. Phys. Rev. Lett. 128 , 180201 (2022).

Article   ADS   MathSciNet   CAS   PubMed   Google Scholar  

Gabbard, H. et al. Bayesian parameter estimation using conditional variational autoencoders for gravitational-wave astronomy. Nat. Phys. 18 , 112–117 (2022).

Chen, D. et al. Automating crystal-structure phase mapping by combining deep learning with constraint reasoning. Nat. Mach. Intell. 3 , 812–822 (2021).

Gómez-Bombarelli, R. et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS Cent. Sci. 4 , 268–276 (2018).

Anishchenko, I. et al. De novo protein design by deep network hallucination. Nature 600 , 547–552 (2021).

Fu, T. et al. Differentiable scaffolding tree for molecular optimization. In International Conference on Learning Representations (2021).

Sanchez-Lengeling, B. & Aspuru-Guzik, A. Inverse molecular design using machine learning: generative models for matter engineering. Science 361 , 360–365 (2018).

Huang, K. et al. Therapeutics Data Commons: machine learning datasets and tasks for drug discovery and development. In NeurIPS Datasets and Benchmarks (2021). This study describes an initiative with open AI models, datasets and education programmes to facilitate advances in therapeutic science across all stages of drug discovery and development.

Dance, A. Lab hazard. Nature 458 , 664–665 (2009).

Segler, M. H. S., Preuss, M. & Waller, M. P. Planning chemical syntheses with deep neural networks and symbolic AI. Nature 555 , 604–610 (2018). This paper describes an approach that combines deep neural networks with Monte Carlo tree search to plan chemical synthesis.

Gao, W., Raghavan, P. & Coley, C. W. Autonomous platforms for data-driven organic synthesis. Nat. Commun. 13 , 1075 (2022).

Kusne, A. G. et al. On-the-fly closed-loop materials discovery via Bayesian active learning. Nat. Commun. 11 , 5966 (2020).

Gormley,A. J. & Webb, M. A. Machine learning in combinatorial polymer chemistry. Nat. Rev. Mater. 6 , 642–644 (2021).

Ament, S. et al. Autonomous materials synthesis via hierarchical active learning of nonequilibrium phase diagrams. Sci. Adv. 7 , eabg4930 (2021).

Degrave, J. et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602 , 414–419 (2022). This paper describes an approach for controlling tokamak plasmas, using a reinforcement-learning agent to command-control coils and satisfy physical and operational constraints.

Melnikov, A. A. et al. Active learning machine learns to create new quantum experiments. Proc. Natl Acad. Sci. USA 115 , 1221–1226 (2018).

Smith, J. S., Isayev, O. & Roitberg, A. E. ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost. Chem. Sci. 8 , 3192–3203 (2017).

Wang, D. et al. Efficient sampling of high-dimensional free energy landscapes using adaptive reinforced dynamics. Nat. Comput. Sci. 2 , 20–29 (2022). This paper describes a neural network for reliable uncertainty estimations in molecular dynamics, enabling efficient sampling of high-dimensional free energy landscapes.

Wang, W. & Gómez-Bombarelli, R. Coarse-graining auto-encoders for molecular dynamics. npj Comput. Mater. 5 , 125 (2019).

Hermann, J., Schätzle, Z. & Noé, F. Deep-neural-network solution of the electronic Schrödinger equation. Nat. Chem. 12 , 891–897 (2020). This paper describes a method to learn the wavefunction of quantum systems using deep neural networks in conjunction with variational quantum Monte Carlo.

Carleo, G. & Troyer, M. Solving the quantum many-body problem with artificial neural networks. Science 355 , 602–606 (2017).

Em Karniadakis, G. et al. Physics-informed machine learning. Nat. Rev. Phys. 3 , 422–440 (2021).

Li, Z. et al. Physics-informed neural operator for learning partial differential equations. Preprint at https://arxiv.org/abs/2111.03794 (2021).

Kochkov, D. et al. Machine learning–accelerated computational fluid dynamics. Proc. Natl Acad. Sci. USA 118 , e2101784118 (2021). This paper describes an approach to accelerating computational fluid dynamics by training a neural network to interpolate from coarse to fine grids and generalize to varying forcing functions and Reynolds numbers.

Ji, W. et al. Stiff-PINN: physics-informed neural network for stiff chemical kinetics. J. Phys. Chem. A 125 , 8098–8106 (2021).

Smith, J. D., Azizzadenesheli, K. & Ross, Z. E. EikoNet: solving the Eikonal equation with deep neural networks. IEEE Trans. Geosci. Remote Sens. 59 , 10685–10696 (2020).

Waheed, U. B. et al. PINNeik: Eikonal solution using physics-informed neural networks. Comput. Geosci. 155 , 104833 (2021).

Chen, R. T. Q. et al. Neural ordinary differential equations. In Advances in Neural Information Processing Systems 31 , 6572–6583 (2018). This paper established a connection between neural networks and differential equations by introducing the adjoint method to learn continuous-time dynamical systems from data, replacing backpropagation.

Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378 , 686–707 (2019). This paper describes a deep-learning approach for solving forwards and inverse problems in nonlinear partial differential equations and can find solutions to differential equations from data.

Article   ADS   MathSciNet   MATH   Google Scholar  

Lu, L. et al. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nat. Mach. Intell. 3 , 218–229 (2021).

Brandstetter, J., Worrall, D. & Welling, M. Message passing neural PDE solvers. In International Conference on Learning Representations (2022).

Noé, F. et al. Boltzmann generators: sampling equilibrium states of many-body systems with deep learning. Science 365 , eaaw1147 (2019). This paper presents an efficient sampling algorithm using normalizing flows to simulate equilibrium states in many-body systems.

Rezende, D. & Mohamed, S. Variational inference with normalizing flows. In International Conference on Machine Learning 37 , 1530–1538, (2015).

Dinh, L., Sohl-Dickstein, J. & Bengio, S. Density estimation using real NVP. In International Conference on Learning Representations (2017).

Nicoli, K. A. et al. Estimation of thermodynamic observables in lattice field theories with deep generative models. Phys. Rev. Lett. 126 , 032001 (2021).

Kanwar, G. et al. Equivariant flow-based sampling for lattice gauge theory. Phys. Rev. Lett. 125 , 121601 (2020).

Gabrié, M., Rotskoff, G. M. & Vanden-Eijnden, E. Adaptive Monte Carlo augmented with normalizing flows. Proc. Natl Acad. Sci. USA 119 , e2109420119 (2022).

Article   MathSciNet   PubMed   PubMed Central   Google Scholar  

Jasra, A., Holmes, C. C. & Stephens, D. A. Markov chain Monte Carlo methods and the label switching problem in Bayesian mixture modeling. Stat. Sci. 20 , 50–67 (2005).

Bengio, Y. et al. Better mixing via deep representations. In International Conference on Machine Learning 552–560 (2013).

Pompe, E., Holmes, C. & Łatuszyński, K. A framework for adaptive MCMC targeting multimodal distributions. Ann. Stat. 48 , 2930–2952 (2020).

Article   MathSciNet   MATH   Google Scholar  

Townshend, R. J. L. et al. ATOM3D: tasks on molecules in three dimensions. In NeurIPS Datasets and Benchmarks (2021).

Kearnes, S. M. et al. The open reaction database. J. Am. Chem. Soc. 143 , 18820–18826 (2021).

Chanussot, L. et al. Open Catalyst 2020 (OC20) dataset and community challenges. ACS Catal. 11 , 6059–6072 (2021).

Brown, N. et al. GuacaMol: benchmarking models for de novo molecular design. J. Chem. Inf. Model. 59 , 1096–1108 (2019).

Notin, P. et al. Tranception: protein fitness prediction with autoregressive transformers and inference-time retrieval. In International Conference on Machine Learning 16990–17017 (2022).

Mitchell, M. et al. Model cards for model reporting. In Conference on Fairness, Accountability, and Transparency 220–229 (2019).

Gebru, T. et al. Datasheets for datasets. Commun. ACM 64 , 86–92 (2021).

Bai, X. et al. Advancing COVID-19 diagnosis with privacy-preserving collaboration in artificial intelligence. Nat. Mach. Intell. 3 , 1081–1089 (2021).

Warnat-Herresthal, S. et al. Swarm learning for decentralized and confidential clinical machine learning. Nature 594 , 265–270 (2021).

Hie, B., Cho, H. & Berger, B. Realizing private and practical pharmacological collaboration. Science 362 , 347–350 (2018).

Rohrbach, S. et al. Digitization and validation of a chemical synthesis literature database in the ChemPU. Science 377 , 172–180 (2022).

Gysi, D. M. et al. Network medicine framework for identifying drug-repurposing opportunities for COVID-19. Proc. Natl Acad. Sci. USA 118 , e2025581118 (2021).

King, R. D. et al. The automation of science. Science 324 , 85–89 (2009).

Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nat. Methods 19 , 679–682 (2022).

Doerr, S. et al. TorchMD: a deep learning framework for molecular simulations. J. Chem. Theory Comput. 17 , 2355–2363 (2021).

Schoenholz, S. S. & Cubuk, E. D. JAX MD: a framework for differentiable physics. In Advances in Neural Information Processing Systems 33 , 11428–11441 (2020).

Peters, J., Janzing, D. & Schölkopf, B. Elements of Causal Inference: Foundations and Learning Algorithms (MIT Press, 2017).

Bengio, Y. et al. A meta-transfer objective for learning to disentangle causal mechanisms. In International Conference on Learning Representations (2020).

Schölkopf, B. et al. Toward causal representation learning. Proc. IEEE 109 , 612–634 (2021).

Goyal, A. & Bengio, Y. Inductive biases for deep learning of higher-level cognition. Proc. R. Soc. A 478 , 20210068 (2022).

Deleu, T. et al. Bayesian structure learning with generative flow networks. In Conference on Uncertainty in Artificial Intelligence 518–528 (2022).

Geirhos, R. et al. Shortcut learning in deep neural networks. Nat. Mach. Intell. 2 , 665–673 (2020).

Koh, P. W. et al. WILDS: a benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning 5637–5664 (2021).

Luo, Z. et al. Label efficient learning of transferable representations across domains and tasks. In Advances in Neural Information Processing Systems 30 , 165–177 (2017).

Mahmood, R. et al. How much more data do I need? estimating requirements for downstream tasks. In IEEE Conference on Computer Vision and Pattern Recognition 275–284 (2022).

Coley, C. W., Eyke, N. S. & Jensen, K. F. Autonomous discovery in the chemical sciences part II: outlook. Angew. Chem. Int. Ed. 59 , 23414–23436 (2020).

Gao, W. & Coley, C. W. The synthesizability of molecules proposed by generative models. J. Chem. Inf. Model. 60 , 5714–5723 (2020).

Kogler, R. et al. Jet substructure at the Large Hadron Collider. Rev. Mod. Phys. 91 , 045003 (2019).

Acosta, J. N. et al. Multimodal biomedical AI. Nat. Med. 28 , 1773–1784 (2022).

Alayrac, J.-B. et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems 35 , 23716–23736 (2022).

Elmarakeby, H. A. et al. Biologically informed deep neural network for prostate cancer discovery. Nature 598 , 348–352 (2021).

Qin, Y. et al. A multi-scale map of cell structure fusing protein images and interactions. Nature 600 , 536–542 (2021).

Schaffer, L. V. & Ideker, T. Mapping the multiscale structure of biological systems. Cell Systems 12 , 622–635 (2021).

Stiglic, G. et al. Interpretability of machine learning-based prediction models in healthcare. Wiley Interdiscip. Rev. Data Min. Knowl. Discov. 10 , e1379 (2020).

Erion, G. et al. A cost-aware framework for the development of AI models for healthcare applications. Nat. Biomed. Eng. 6 , 1384–1398 (2022).

Lundberg, S. M. et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat. Biomed. Eng. 2 , 749–760 (2018).

Sanders, L. M. et al. Beyond low Earth orbit: biological research, artificial intelligence, and self-driving labs. Preprint at https://arxiv.org/abs/2112.12582 (2021).

Gagne, D. J. II et al. Interpretable deep learning for spatial analysis of severe hailstorms. Mon. Weather Rev. 147 , 2827–2845 (2019).

Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell. 1 , 206–215 (2019).

Koh, P. W. & Liang, P. Understanding black-box predictions via influence functions. In International Conference on Machine Learning 1885–1894 (2017).

Mirzasoleiman, B., Bilmes, J. & Leskovec, J. Coresets for data-efficient training of machine learning models. In International Conference on Machine Learning 6950–6960 (2020).

Kim, B. et al. Interpretability beyond feature attribution: quantitative testing with concept activation vectors (TCAV). In International Conference on Machine Learning 2668–2677 (2018).

Silver, D. et al. Mastering the game of go without human knowledge. Nature 550 , 354–359 (2017).

Baum, Z. J. et al. Artificial intelligence in chemistry: current trends and future directions. J. Chem. Inf. Model. 61 , 3197–3212 (2021).

Finlayson, S. G. et al. Adversarial attacks on medical machine learning. Science 363 , 1287–1289 (2019).

Urbina, F. et al. Dual use of artificial-intelligence-powered drug discovery. Nat. Mach. Intell. 4 , 189–191 (2022).

Norgeot, B. et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat. Med. 26 , 1320–1324 (2020).

Download references

Acknowledgements

M.Z. gratefully acknowledges the support of the National Institutes of Health under R01HD108794, U.S. Air Force under FA8702-15-D-0001, awards from Harvard Data Science Initiative, Amazon Faculty Research, Google Research Scholar Program, Bayer Early Excellence in Science, AstraZeneca Research, Roche Alliance with Distinguished Scientists, and Kempner Institute for the Study of Natural and Artificial Intelligence. C.P.G. and Y.D. acknowledge the support from the U.S. Air Force Office of Scientific Research under Multidisciplinary University Research Initiatives Program (MURI) FA9550-18-1-0136, Defense University Research Instrumentation Program (DURIP) FA9550-21-1-0316, and awards from Scientific Autonomous Reasoning Agent (SARA), and AI for Discovery Assistant (AIDA). Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funders. We thank D. Hassabis, A. Davies, S. Mohamed, Z. Li, K. Ma, Z. Qiao, E. Weinstein, A. V. Weller, Y. Zhong and A. M. Brandt for discussions on the paper.

Author information

Hanchen Wang

Present address: Department of Research and Early Development, Genentech Inc, South San Francisco, CA, USA

Present address: Department of Computer Science, Stanford University, Stanford, CA, USA

These authors contributed equally: Hanchen Wang, Tianfan Fu, Yuanqi Du

Authors and Affiliations

Department of Engineering, University of Cambridge, Cambridge, UK

Hanchen Wang & Joan Lasenby

Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA, USA

Hanchen Wang & Anima Anandkumar

Department of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Department of Computer Science, Cornell University, Ithaca, NY, USA

Yuanqi Du & Carla P. Gomes

Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA

Wenhao Gao & Connor W. Coley

Department of Computer Science, Stanford University, Stanford, CA, USA

Kexin Huang & Jure Leskovec

Department of Physics, Massachusetts Institute of Technology, Cambridge, MA, USA

Harvard-MIT Program in Health Sciences and Technology, Cambridge, MA, USA

Payal Chandak

Mila – Quebec AI Institute, Montreal, Quebec, Canada

Shengchao Liu, Andreea Deac, Jian Tang & Yoshua Bengio

Université de Montréal, Montreal, Quebec, Canada

Shengchao Liu, Andreea Deac & Yoshua Bengio

Department of Earth, Environmental and Planetary Sciences, Brown University, Providence, RI, USA

Peter Van Katwyk & Karianne Bergen

Data Science Institute, Brown University, Providence, RI, USA

NVIDIA, Santa Clara, CA, USA

Anima Anandkumar

Center for Computational Astrophysics, Flatiron Institute, New York, NY, USA

Department of Astrophysical Sciences, Princeton University, Princeton, NJ, USA

Department of Physics, Carnegie Mellon University, Pittsburgh, PA, USA

Department of Physics and Center for Data Science, New York University, New York, NY, USA

Google DeepMind, London, UK

Pushmeet Kohli & Petar Veličković

Microsoft Research, Beijing, China

Tie-Yan Liu

Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA

Arjun Manrai & Marinka Zitnik

Department of Systems Biology, Harvard Medical School, Boston, MA, USA

Debora Marks

Broad Institute of MIT and Harvard, Cambridge, MA, USA

Debora Marks & Marinka Zitnik

Deep Forest Sciences, Palo Alto, CA, USA

Bharath Ramsundar

BioMap, Beijing, China

Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates

University of Illinois at Urbana-Champaign, Champaign, IL, USA

HEC Montréal, Montreal, Quebec, Canada

CIFAR AI Chair, Toronto, Ontario, Canada

Department of Computer Science and Technology, University of Cambridge, Cambridge, UK

Petar Veličković

University of Amsterdam, Amsterdam, Netherlands

Max Welling

Microsoft Research Amsterdam, Amsterdam, Netherlands

DP Technology, Beijing, China

Linfeng Zhang

AI for Science Institute, Beijing, China

Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA

Connor W. Coley

Harvard Data Science Initiative, Cambridge, MA, USA

Marinka Zitnik

Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, Cambridge, MA, USA

You can also search for this author in PubMed   Google Scholar

Contributions

All authors contributed to the design and writing of the paper, helped shape the research, provided critical feedback, and commented on the paper and its revisions. H.W., T.F., Y.D. and M.Z conceived the study and were responsible for overall direction and planning. W.G., K.H. and Z.L. contributed equally to this work (equal second authorship) and are listed alphabetically.

Corresponding author

Correspondence to Marinka Zitnik .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Nature thanks Brian Gallagher and Benjamin Nachman for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Cite this article.

Wang, H., Fu, T., Du, Y. et al. Scientific discovery in the age of artificial intelligence. Nature 620 , 47–60 (2023). https://doi.org/10.1038/s41586-023-06221-2

Download citation

Received : 30 March 2022

Accepted : 16 May 2023

Published : 02 August 2023

Issue Date : 03 August 2023

DOI : https://doi.org/10.1038/s41586-023-06221-2

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Antimicrobial resistance crisis: could artificial intelligence be the solution.

• Guang-Yu Liu
• Xiao-Fen Liu

Military Medical Research (2024)

The advancement of artificial intelligence in biomedical research and health innovation: challenges and opportunities in emerging economies

• Renan Gonçalves Leonel da Silva

Globalization and Health (2024)

Embracing data science in catalysis research

• Manu Suvarna
• Javier Pérez-Ramírez

Nature Catalysis (2024)

Artificial intelligence to predict soil temperatures by development of novel model

• Lakindu Mampitiya
• Kenjabek Rozumbetov
• Upaka Rathnayake

Scientific Reports (2024)

Machine learning for micro- and nanorobots

• Lidong Yang
• Jialin Jiang

Nature Machine Intelligence (2024)

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

Artificial Intelligence: Definition and Background

• Open Access
• First Online: 31 January 2023

Cite this chapter

You have full access to this open access chapter

• Haroon Sheikh 6 ,
• Corien Prins 6 &
• Erik Schrijvers 6  

Part of the book series: Research for Policy ((RP))

61k Accesses

29 Citations

21 Altmetric

If we want to embed AI in society, we need to understand what it is. What do we mean by artificial intelligence? How has the technology developed? Where do we stand now?

You have full access to this open access chapter,  Download chapter PDF

1 Definitions of AI

Defining AI is not easy; in fact, there is no generally accepted definition of the concept. Footnote 1 Numerous different ones are used, and this can easily lead to confusion. It is therefore important to clarify our use of the term. We start by discussing various definitions of AI, then explain which we have settled on. The sheer variety of definitions in circulation is not due to carelessness, but inherent in the phenomenon of AI itself.

In its broadest definition, AI is equated with algorithms. However, this is not an especially useful approach for our analysis. Algorithms predate AI and have been widely used outside this field. The term ‘algorithm’ is derived from the name of the ninth-century Persian mathematician Mohammed ibn Musa al-Kharizmi and refers to a specific instruction for solving a problem or performing a calculation. If we were to define AI simply as the use of algorithms, it would include many other activities such as the operations of a pocket calculator or even the instructions in a cookbook.

In its strictest definition, AI stands for the imitation by computers of the intelligence inherent in humans. Purists point out that many current applications are still relatively simple and therefore not true AI. That makes this definition inappropriate for our report, too; to use it would be to imply that AI does not exist at present. We would effectively be defining the phenomenon out of existence.

A common definition of AI is that it is a technology that enables machines to imitate various complex human skills. This, however, does not give is much to go on. In fact, it does no more than render the term ‘artificial intelligence’ in different words. As long as those ‘complex human skills’ are not specified, it remains unclear exactly what AI is. The same applies to the definition of AI as the performance by computers of complex tasks in complex environments.

Other definitions go further in explaining these skills and tasks. For example, the computer scientist Nils John Nilsson describes a technology that “functions appropriately and with foresight in its environment”. Footnote 2 Others speak of the ability to perceive, to pursue goals, to initiate actions and to learn from a feedback loop. Footnote 3 A similar definition has been put forward by the High-Level Expert Group on Artificial Intelligence (AI HLEG) of the European Commission (EC): “Systems that display intelligent behaviour by analysing their environment and taking actions – with some degree of autonomy – to achieve specific goals.” Footnote 4

These task-based definitions go some way towards giving us a better understanding of what AI is. But they still have limitations. Concepts like “some degree of autonomy” remain somewhat vague. Moreover, these definitions still seem overly broad in that they describe phenomena that most of us would not be inclined to bundle under the term AI. For example, Nilsson’s definition also applies to a classic thermostat. This device is also able to perceive (measure the temperature of the room), pursue goals (the programmed temperature), initiate actions (regulate the thermostat) and learn from a feedback loop (stop once the programmed temperature has been reached). Even so, most people would not be inclined to regard a thermostat as AI.

It is not surprising that AI is so difficult to define clearly. It is, after all, an imitation or simulation of something we do not yet fully understand ourselves: human intelligence. This has long been the subject of research by psychologists, behavioural scientists and neurologists, amongst others. We know a lot about intelligence and the human brain, but that knowledge is far from complete and there is no consensus as to what exactly human intelligence is. Until that comes about, it is impossible to be precise about how that intelligence can be imitated artificially.

Moreover, there is a clear interface between research into human intelligence on the one hand and into artificial intelligence on the other, where our understanding of both is co-evolving. We can illustrate this using the example of chess, a game AI has been able to play extremely well since the 1990s. In the 1950s an expert predicted, “If one could devise a successful chess machine, one would seem to have penetrated to the core of human intellectual endeavour.” Footnote 5 In 1965 the Russian mathematician Alexander Kronrod called chess “the fruit fly of intelligence” – that is, the key to understanding it. Footnote 6 So people were amazed when a computer did finally manage to beat a chess grandmaster. In the Netherlands, research in this field led to the founding of the Dutch Computer Chess Association foundation (Computer Schaak Vereniging Nederland, CSVN) in 1980. Amongst its initiators were chess legend and former world champion Max Euwe and computer scientist Jaap van den Herik. Three years later Van den Herik would defend the first PhD thesis in the Netherlands on computer chess and artificial intelligence. In 1997, when Garry Kasparov was defeated by Deep Blue, IBM’s chess computer, the cover of Newsweek claimed that this was “The brain’s last stand.” Chess was considered the pinnacle of human intelligence. At first glance this is not surprising, because the game is difficult for people to learn and those who are good at it are often considered very clever. It was with this in mind that commentators declared Deep Blue’s victory a huge breakthrough for human intelligence in machines, stating that it must now be within the reach of computers to surpass humans in all sorts of activities we consider easier than chess.

Yet this did not happen. We have since revised our view of this form of intelligence. Chess is not the crowning glory of human intellectual endeavour; it is simply a mathematical problem with very clear rules and a finite set of alternatives. In this sense, a chess program is actually not very different from a pocket calculator, which can also do things too difficult even for very clever people. But they do not make it an artificial form of human intelligence.

Chess was long considered an extremely advanced game. However, years of research have revealed that something as apparently simple as recognizing a cat in a photograph – which AI has only learnt to do in recent years – is far more complex. This phenomenon has come to be known as Moravec’s paradox: certain things that are very difficult for humans, such as chess or advanced calculus, are quite easy for computers. Footnote 7 But things that are very simple for us humans, such as perceiving objects or using motor skills to do the washing up, turn out to be very difficult for computers: “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers [draughts], and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.” Footnote 8

This reflects a recurring pattern in the history of AI: people’s idea of what constitutes a complex form of human intelligence has evolved with the increasing skills of our computers. What used to be considered a fine example of artificial intelligence eventually degrades to a simple calculation that no longer deserves the name AI. Pamela McCorduck calls this the ‘AI effect’: as soon as a computer figures out how to do something, people declare that it is ‘just a calculation’ and not actual intelligence. According to Nick Bostrom, director of the Oxford Institute for Internet Governance, AI includes anything that impresses us at any given time. Once we are no longer impressed, we simply call it software. Footnote 9 A chess app on a smartphone is an example. The difficulties in defining AI are therefore not the result of some shortcoming or carelessness, but rather arise from the fact that we were long unable to determine precisely what intelligence we wanted to imitate artificially.

In this context, it is also claimed that the use of the term ‘intelligence’ is misleading in that it wrongly suggests that machines can do the same things as people. Some have therefore suggested adopting other terms. Agrawal, Gans and Goldfarb say that modern technology does not bring us intelligence, but only one of its components, predictions, and so they use the term ‘prediction machines’. Footnote 10 The philosopher Daniel Dennett goes even further and suggests that we should not model AI on humans at all. These are not artificial people, but a completely new type of entity – one he compares with oracles: entities that make predictions, but unlike humans have no personality, conscience or emotions. Footnote 11 In other words, AI appears to do what people do but in fact does something else. Edsger Dijkstra illustrated this through the question ‘Do submarines swim?’. Footnote 12 What these vessels do is similar to what humans call swimming, but to call it that would be a mistake. AI can certainly do things that look like the intelligent things we do, but in fact it does them very differently.

This perspective also sheds light on the Moravec paradox mentioned above. Recognizing faces is easy for humans, but difficult for computers. This is because recognizing others was critical for our evolutionary survival and so our brain has learned to do it without thinking.

Being able to play chess was not essential in evolution and is therefore more difficult to master. That is to say, it requires a certain level of computational skill. Computers have not evolved biologically, so their abilities are different from those of humans. One important aspect of this theory is that we should not try too hard to understand AI from the point of view of human intelligence. Nevertheless, the term ‘artificial intelligence’ has become so commonplace that there is no point trying to replace it now.

Finally, AI is also often equated with the latest technology. As we will see later, AI has gained huge momentum in recent years. One of the major drivers of this has been progress in a specific area of the field, ‘machine learning’ (ML), where the innovation has resulted in what is now called ‘deep learning’ (DL). It is this technology that has been behind recent milestones, such as computers able to recognize faces and play games like Go. By contrast with the more traditional approaches whereby computer systems apply fixed rules, ML and DL algorithms can recognize patterns in data. We also speak here of ‘self-learning algorithms’. Many people who talk about AI today are actually referring to these algorithms, and often specifically to DL. The focus on this technology is important because several pressing questions concerning AI are particularly relevant here (such as problems of explainability).

Given all the different definitions discussed here and elsewhere, we have settled on an open definition of AI. Two considerations are relevant in this respect. Firstly, it would be unwise for the purposes of this report to limit the definition of AI to a specific part of the technology. If, for example, we were to confine ourselves to ‘deep learning’ as discussed above, we would ignore the fact that many current issues also play a role in other AI domains, such as logical systems. One such example is the ‘black box’ question. Also, most applications of AI used by governments are not based on advanced techniques like DL and yet still have many important issues that need to be addressed in this report. Too narrow a definition would place them outside the scope of this study. While developments in DL have indeed resulted in a great leap forward, moreover, at the end of the next chapter we also point out several shortcomings of this technique. In fact, future advances in AI may well come from other fields. To allow for this, it is important to have an open definition of AI.

Secondly, as discussed above the nature of this scientific discipline necessarily means that our definition of AI will change over time. Instead of considering AI as a discipline that can be clearly delineated, with uncomplicated definitions and fixed methodologies, it is more useful to see it as a complex and diverse field focused on a certain horizon. The dot on that horizon is the understanding and simulation of all human intellectual skills. This goal is also called ‘artificial general intelligence’ or AGI (other names are ‘strong AI’ and ‘full AI’). However, it remains to be seen whether this dot, with such a generic definition of AI, will ever be reached. Most experts believe that this it at least several decades away – if it is ever attained at all. Footnote 13

A fixed definition of AI as the imitation of full human intelligence is of little use for the purposes of this report. We need a definition that captures the whole range of applications finding their way into practice today and in the near future. The definition from the AI HLEG provides the necessary freedom of scope. Describing AI as “systems that display intelligent behaviour by analysing their environment and taking actions – with some degree of autonomy – to achieve specific goals”, this encompasses all the applications we currently qualify as AI and at the same time provides scope for future changes to that qualification. Alongside advanced machine learning and deep learning technologies, this definition also allows for other technologies, including the more traditional approaches mentioned above, as used by many government bodies. In short, this definition is sufficiently strict to distinguish AI from algorithms and digital technology in general, while at the same time open enough to include future developments. Figure 2.1 provides an overview of the definitions discussed and the AI HLEG definition used in this report.

Various definitions of AI

It is worth emphasizing that the current applications considered as AI according to this definition all fall under the heading ‘narrow’ or ‘weak’ AI. Footnote 14 The AI that we are familiar with today focuses on specific skills, such as image or speech recognition, and has little to do with the full spectrum of human cognitive capabilities covered by AGI. This does not alter the fact that current AI applications can and do give rise to major issues, too. The American professor of Machine Learning Pedro Domingos has put this nicely; in his view we focus too much on a future AGI and too little on the narrow AI that is already all around us. “People worry that computers will get too smart and take over the world,” he says, “but the real problem is that they’re too stupid and they’ve already taken over the world.” Footnote 15

The fact that AI is difficult to define is linked to the evolution of this discipline. We now take a closer look at how that evolution took place. A short historical overview is not only relevant as a background for understanding AI, it is also the prelude to the next chapter in which we see that AI has reached a turning point.

2 AI Prior to the Lab

It is possible to date the birth of some disciplines very precisely. AI is one. Its conception in the laboratory is often dated to 1956, during a summer school at Dartmouth College in New Hampshire, USA. AI did not come out of the blue, however. The technology already had a long history before it was first seriously investigated as a scientific discipline.

This history can be divided roughly into three phases: early mythical representations of artificial forms of life and intelligence; speculations about thinking machines during the Enlightenment; and the establishment of the theoretical foundation for the computer (see Fig. 2.2 ). The latter was the springboard for the development of AI as a separate discipline. We now discuss these three phases in turn, but bearing in mind that in practice they have never been mutually exclusive. Myths have always existed and there has always been creative speculation about the future in parallel with the theoretical research into AI. Nevertheless, the phases reveal how the nature and focus of AI thinking have changed over time.

Three phases of AI prior to the lab

2.1 The Mythical Representation of AI

Myths and stories about what we would now call AI have been around for centuries (see Fig. 2.3 ). The ancient Greeks in particular celebrated a multitude of characters in their mythology who can be characterized as artificial forms of intelligence. Footnote 16 Take Talos, a robot created by the great inventor Daedalus to protect the island of Crete. Every day, Talos would run circles around the island and throw stones at any approaching ships he spotted. This is clearly a myth about a mechanical super-soldier. A robotic exoskeleton used by the US Army now bears the same name.

Ancient myths about AI

Daedalus, the ancient world’s great inventor, is famous for the wings that cost the life of his son Icarus, but he was also the inventor of all manner of artificial intelligence, such as moving statues as well as Talos. According to the myth, this robot was eventually defeated by the witch Medea, who tricked it into disabling itself. So, while Daedalus was an AI inventor, in the same legend Medea was able to magically control his AI. Moreover, her father was responsible for creating artificial soldiers who could fight without needing rest.

In addition to the two human characters of Daedalus and Medea, various Greek gods were also associated with artificial intelligence. Hephaistos, the blacksmith of the gods, was assisted in his workshop by mechanical helpers. He also built tools that moved independently and a heavenly gate that opened automatically. The titan Prometheus ‘built’ humans and stole fire from the gods for them. To punish humankind, Zeus created a kind of robot, the mechanical woman Pandora, who poured out all kinds of suffering on humans when she opened her jar (‘Pandora’s box’). A less grim example is the myth of Pygmalion. A sculptor, he fell in love with a statue he had made, upon which Aphrodite brought it to life and he made his creation, named Galatea, his wife. So the ancient Greeks were already imagining what we now would call killer robots, mechanical assistants and sex robots in their mythology.

There are also stories about forms of AI in other traditions, such as the Jewish golem and the mythical jinn (genies) of Arabia who can grant wishes. The Buddhist story Lokapannatti tells how the emperor Ashoka wanted to lay his hands on the relics of the Buddha, which were protected by dangerous mechanical guards made in Rome. Footnote 17 Norse mythology tells of the giant Hrungnir, built to battle Thor. The Liezi , an ancient Chinese text, relates the story of the craftsman Yan Shi, who built an automaton with leather for muscles and wood for bones. Footnote 18 Estonia has a legend about the Kratt, a magical creature made of hay and household items that did everything its owner asked. If the Kratt was not kept busy, it became a danger to its owner. The modern law in Estonia that governs liability for the use of algorithms is known there as the ‘Kratt Law’.

2.2 Speculation About Thinking Machines

The next phase was heralded by the ‘mechanization of the world’ Footnote 19 envisaged in the work of thinkers like Galileo Galilei, Isaac Newton and René Descartes. Their mechanical worldview was accompanied by the construction of all kinds of novel machines. Artificial intelligence was still far beyond the realm of possibility, but the new devices did lead to speculation about its creation (see Fig. 2.4 ) – speculation that was no longer mythical, but mechanical in nature.

Timeline of speculations about AI

In 1642 Blaise Pascal built a mechanical calculator which he said was “closer to thought than anything done by animals”. Footnote 20 Gottfried Leibniz constructed an instrument he called the ‘step reckoner’ in 1673, which could be used to perform arithmetical calculations. This laid the foundation for many future computers. Footnote 21 The philosophers of the time speculated about such devices using the term ‘automata’.

In 1769 Wolfgang von Kempelen built a highly sophisticated machine – or so people long thought. He gained worldwide fame after offering his mechanical ‘Turk’ to the Austrian Empress Maria Theresa. The huge device was an automatic chess machine, which toured the western world for 48 years and defeated opponents like Napoleon Bonaparte and Benjamin Franklin. It was not until the 1820s that it was discovered to be a total fake: there was a man inside the machine moving the pieces. Footnote 22 As an aside, the company Amazon has a platform called Mechanical Turk where people can arrange to have tasks done cheaply online. While more open than Von Kempen’s original, here too the work is done by people behind the scenes we do not see.

Speculation about AI could also take magical forms during this period. Goethe’s story of the sorcerer’s apprentice, made famous in Disney’s animated film Fantasia starring Mickey Mouse, is about an apprentice who uses a spell to make a broom fetch water. When it turns out he does not know the spell to make the process stop, and instead the broom begins to multiply itself, a disaster unfolds that only ends when the wizard returns. Footnote 23 Other magical stories about phenomena similar to AI include Pinocchio and the horror story by W. W. Jacobs about a monkey’s paw that grants three wishes with terrible consequences.

Tales of magic have also spilled over into stories a little closer to scientific reality, in the form of science fiction. In 1816 a group of writers meeting near Geneva was forced to spend long periods indoors because of a volcanic eruption in what is now Indonesia. That caused the so-called ‘Year Without a Summer’, when abnormal rainfall kept people inside. Inspired by the magical stories of E. T. A. Hoffman, Lord Byron suggested that each member of the group write a supernatural story, upon which Mary Shelley penned the first version of her famous novel Frankenstein . Footnote 24

The story of a scientist who creates an artificial form of life that ultimately turns against its creator has become the archetype of the risks of modern technology. This motif lives on in countless films, including classics like Blade Runner (1982), The Terminator (1984) and The Matrix (1999).

Another important work of literary science fiction in the context of speculation about AI is R.U.R. by the Czech author Karel Capek. It is in this book that the writer introduces the term ‘robot’, a word derived from the Old Church Slavonic word ‘rabota’, meaning corvée or forced labour. This story also reveals a classic fear of AI; in it the artificial labourers (‘roboti’) created in a factory rebel against their creators and ultimately destroy humankind. Footnote 25 Capek’s book was published in 1920, by which time the next phase – much more concrete thinking about AI – had long since begun.

2.3 The Theory of AI

From the second half of the nineteenth century onwards, the idea of AI as ‘thinking computers’ became less fantastical and entered the realm of serious theoretical consideration (see Fig. 2.5 ). This development occurred in parallel with the theorization and construction of the first computers.

Timeline of theories of AI

Ada Lovelace – daughter of the poet Byron, instigator of the writing session that had produced Frankenstein – would play an important role in this field in the 1840s. She envisaged a machine that could play complex music based on logic, and also advance scientific research in general. Her acquaintance Charles Babbage designed such a device in 1834 and called it the ‘Analytical Engine’. Footnote 26 He had earlier failed in his efforts to build an enormously complex Difference Engine and so instead created the Analytical Engine as an alternative with which he hoped to construct mathematical and astronomical tables. Footnote 27 Lovelace, however, saw a much wider use for a ‘thinking machine’ that could reason about “all the subjects in the universe”. Footnote 28 She even wrote programs for the hypothetical device. However, science at that time was not advanced enough to actually build such computers.

That point would not be reached until the Second World War, when computing power was needed to defend against air raids. The use of fast-moving planes to drop bombs made it impossible for the human operators of anti-aircraft systems to respond quickly enough when relying on their eyesight alone. Instead, their targets’ trajectories needed to be calculated mathematically. Research in that field laid the foundations for the modern computer and for another discipline that would emerge in the 1950s, cybernetics. This work immediately raised questions about automation and human control that are still relevant today.

“The time factor has become so narrow for all operators,” a military spokesperson said at the time, “that the human link, which seems to be the only immutable factor in the whole problem and which is particularly fickle, has increasingly become the weakest link in the chain of operations, such that it has become clear that this link must be removed from the sequence.” Footnote 29

The development of the computer was given another boost during the war by the British research programme Colossus, which aimed to crack the Nazis’ secret communication system known as Enigma. One of the leading lights in this top-secret project at Bletchley Park was Alan Turing, often regarded as the father of both computers and AI. He went on to help develop the first truly modern computer in Manchester in 1948. Two years after that, in 1950, he wrote a paper proposing a thought experiment in the form of an ‘imitation game’ for a computer pretending to be a human being. Footnote 30 This has come to be known as the Turing test. A computer passes if a human is unable to establish that its written answers to their questions were provided by a person or a computer. Variants of this test are still used, for example, to compare AI systems with human abilities such as recognizing images or using language. Footnote 31

Another important theoretical contribution to this field was a paper by psychiatrist and neurologist Warren McCulloch and mathematician Walter Pitts. Footnote 32 In this they combined Turing’s work on computers with Bertrand Russell’s propositional logic and Charles Sherrington’s theory of neural synapses. Their most important contribution was that they demonstrated binary modalities (a situation with two options) in various domains and thus developed a common language for neurophysiology, logic and computation. The distinction between ‘true and false’ in logic was now linked to the ‘on or off’ state of neurons and the computer values ‘0 and 1’ in Turing machines. Footnote 33

John von Neumann continued to develop the basic concept of a computer with components such as the central processor, memory and input-output devices. Footnote 34 Another important founder of AI theory was Norbert Wiener. He coined the term ‘cybernetics’ in 1948 to describe “the study of control and communication in animals and machines”. Footnote 35 The key idea was that people, animals and machines could all be understood according to a number of basic principles. The first of these is control: all those entities strive to counter entropy and to control their environment using the principle of ‘feedback’, which is the “ability to adapt future behaviour to past experience”. Through the mechanism of continuous adjustment and feedback, organisms and machines ensure that equilibrium, or homeostasis, is achieved. Wiener used thermostats and servomechanisms as metaphors to explain these processes. Although cybernetics did not last long as a separate scientific field, its core concepts now permeate all manner of disciplines (Box 2.1 ). Footnote 36

Box 2.1: The Homeostat and Electronic Tortoises

In 1948 the Briton Ross Ashby unveiled his ‘homeostat’, a machine able to hold four electromagnets in a stable position. In that same year The Herald wrote of this ‘protobrain’ that “the clicking brain is cleverer than man’s”. Footnote 37 Another highlight of the cybernetics movement in the 1950s was William Grey Walter’s electronic tortoises. These small devices could walk around without bumping into obstacles and locate where in the room their charger was if their battery was weak. Moreover, they also exhibited complex social behaviour as a group. A later example of a cybernetic machine was the John Hopkins Beast, which in the early 1960s was able to trundle through corridors using sonar and a photocell eye to find a charging point. Footnote 38

Thanks to such advances, during this period scientists were ready to stop just dreaming and thinking about AI and start actually developing the technology and experimenting with it in the laboratory. The starting gun for this race was fired in 1956.

Key Points: AI Prior to the Lab

Mythical representations of AI have been around for centuries.

The most celebrated examples are the ancient Greek stories about Daedalus, Medea, Hephaistos, Prometheus and Pygmalion.

The mechanization of the world view from the seventeenth century onwards made the construction of all kinds of machines possible. This went hand in hand with speculation about mechanical brains.

Fictional stories about artificial intelligence appeared from the Industrial Revolution onwards, including Frankenstein and R.U.R .

The theoretical foundations for AI were laid when the first computers were built by people like Alan Turing.

3 AI in the Lab

3.1 the first wave.

As mentioned previously, the beginnings of AI as a discipline can be dated very precisely. Footnote 39 After all the myths, speculation and theorizing, artificial intelligence appeared in a lab for the first time in 1956 when a group of scientists made it the subject of a specific event: the Dartmouth Summer Research Project on Artificial Intelligence. This was a six-week brainstorming gathering attended by several of the discipline’s founders. The organizers were very optimistic about what they could achieve with this group in a few weeks, as is evident from the proposal they wrote to the Rockefeller Foundation.

We propose … a 2-month, 10-man study of artificial intelligence … The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer. Footnote 40

The proposal was overambitious, and research is still being carried out today in all the areas it mentioned. With this project, however, these scientists formulated a research agenda that launched AI as a discipline.

The summer project was organized by John McCarthy and Marvin Minsky. It was McCarthy who coined the term ‘artificial intelligence’ in 1956. Minsky was a leading figure in the history of AI and over the years came to be involved in many prominent high-tech projects around the world. The two men also established the Artificial Intelligence Lab at MIT. This was later renamed the MIT Media Lab and is still a centre for the creative use of new technology. Footnote 41 Among those present at the summer project were Herbert Simon (Nobel laureate in Economics and winner of the Turing Award, responsible for the idea of ‘bounded rationality’, amongst other things, and founder of the Carnegie Institute of Technology), John Nash (mathematician, game theorist and another Nobel laureate in Economics) and Arthur Samuel (pioneer of computer games and the man credited with popularizing the term ‘machine learning’). These leading scientists were responsible for bringing AI to the lab.

This landmark event heralded a period of great optimism and broad interest in the field of AI, which has come to be known as the first ‘AI spring’ (or ‘wave’). Various programs were developed that could play the board game draughts (checkers), although none was very good yet. The version developed by Samuel did eventually succeed in defeating its human creator, which caused a stir, although he was not known as a great player of the game. Wiener wrote in 1964 that, while Samuel was eventually able to beat the program again after some instruction, “the method of its learning was no different in principle from that of the human being who learns to play checkers”. He also expected that the same would happen with chess in ten to twenty-five years, and that people would lose interest in both games as a consequence. Footnote 42

Exciting breakthroughs followed when AI systems began focusing on a different category of challenges: logical and conceptual problems. For example, a ‘Logic Theory Machine’ was built to prove Bertrand Russell’s logical theorems. It not only succeeded in proving eighteen of them, it also developed a more elegant proof of one. This was important because, while Samuel was a mediocre draughts player, Bertrand Russell was a leading logician.

The next milestone was the ‘General Problem Solver’. This was a program that could, in principle, be applied to solve any problem – hence the name. By translating problems into goals, subgoals, actions and operators, the software could then reason what the right answer was. One example of a problem it solved is the classic logical puzzle of the river crossing. Footnote 43

By the mid-1960s the first students of the AI pioneers were working on programs that could prove geometric theorems and successfully complete intelligence tests, maths problems and calculus exams. So, the discipline was making progress, but its impact outside the lab was very limited. There were some interesting experiments with robots, as in the late 1960s at the Stanford Research Institute; its Shakey the Robot was able to find its way about through reasoning. Footnote 44 The American technology company General Electric built impressive robots such as the Beetle and an exoskeleton that enabled humans to lift heavy weights. Footnote 45 These robots were not very practical, though.

At the same time, there were grand expectations of AI. In 1965 Herbert Simon predicted that “machines will be capable, within twenty years, of doing any work a man can do”. Footnote 46 Meanwhile, the British mathematician Irving Jack Good foresaw a machine-induced ‘intelligence explosion’. This would also be the last invention of humankind, because machines would now be the most intelligent beings on earth and therefore do all the inventing. Footnote 47

AI caught the imagination of people outside science as well. In 1967 the computer program MacHack VI was made an honorary member of the American Chess Federation, despite having won very few matches. Footnote 48 A few years later the film Colossus: The Forbin Project was released. In this a computer program is handed control of the US military arsenal because it can make better decisions than humans and is unhindered by emotions. After the Soviets reveal a similar project, the two programs start communicating with one another – but in a way that is incomprehensible to their human creators – and subsequently take control of the entire world. Their pre-programmed goal of world peace is achieved, but the price is the freedom of the human race.

This gap between hopeful expectations and harsh reality did not go unnoticed, and from the second half of the 1960s onwards there was increasing criticism of AI research. The philosopher Hubert Dreyfus would remain critical of the potential of AI throughout his life. In 1965 he wrote a study called AI and Alchemy , commissioned by the Rand Corporation (the think tank of the American armed forces), in which he concluded that intelligent machines would not be developed any time in the near future. In a 1966 report to the US government, the Automatic Language Processing Advisory Committee concluded that little progress had been made. The National Research Council subsequently phased out its funding of AI. In the United Kingdom, Sir James Lighthill was commissioned in 1973 to conduct a survey of the topic; this brought to light considerable criticism of its failure to achieve the grandiose goals that had been promised. As a result, a lot of research funding was withdrawn in the UK as well. Footnote 49

One problem encountered by many AI systems at this time was the so-called ‘combinatorial explosion’. These systems solved problems by exploring all possible options, but they quickly reached the limits of their computing power when dealing with huge numbers of possible combinations. More heuristic approaches, based on rules of thumb, were needed to reduce the number of combinations. However, these did not yet exist. This and other problems – such as the lack of data to feed the systems and the limited capacity of the hardware – meant that progress with AI stalled.

Meanwhile, its practical applications were also proving unreliable. When an AI system was developed during the Cold War, in the 1960s, to translate Russian communications, the results proved less than impressive. One famous example was its translation of “the spirit is willing, but the flesh is weak” as “the vodka is good, but the meat is rotten”. Footnote 50 During the course of the 1970s, the earlier optimism turned to pessimism. There were too few breakthroughs, so criticism of AI grew, and funding dried up. The first ‘AI winter’ had set in and put an end to its first wave. Figure 2.6 provides an overview of the emergence of AI as a scientific discipline.

Timeline of the emergence of AI as a discipline (first wave)

3.2 Two Approaches

It is important to note that two distinct approaches to AI gained particular prominence during this first wave. While it is true that there were others as well (we will explain these later), these two still dominate the field to this day. The first is ‘rule-based’, also known as ‘symbolic’ or ‘logical’, AI (along with other names) and emerged in the 1970s in the form of so-called ‘expert systems’. Its core principle is that computers learn by encoding logical rules with formulas of the type ‘IF X, THEN Y’. The use of logic and rules is also why the term ‘symbolic AI’ is used, as this approach follows rules that can be expressed in human symbols.

The second approach uses artificial neural networks (ANNs) and is also called ‘connectionism’. This includes the deep learning and parallel distributed processing methods that have received a lot of attention in recent years. The central idea here is to simulate the functioning of neurons in the human brain. For this purpose, sets of ‘artificial neurons’ are built into networks that can receive and send information. These networks are then fed with large amounts of data and try to distil patterns from it. In this case the rules are not drawn up by humans in advance. Most ANNs are based on a principle formulated as early as 1949 by Donald Hebb, a Canadian psychologist, in his book The Organization of Behaviour : “Neurons that fire together, wire together”. Footnote 51 In other words, if two neurons are frequently activated at the same time, they become connected.

Both approaches to AI were there from the start. While many of the founding fathers at the 1956 summer school followed the rule-based approach, the first artificial neuron was also created around the same time at Cornell University. Footnote 52 The difference can be explained as follows. To be able to recognize a cat in a photo, in the first approach a series of ‘IF-THEN’ rules are established: the presence of certain colours, a given number of limbs, certain facial forms, whiskers, etc., means that it is a cat. With these rules, a program can ‘reason’ what the data means.

In the second approach, the program might be presented a large number of photos labelled as ‘cat’ and ‘non-cat’. The program distils patterns based on this data, which it then uses to recognize the presence of a cat in subsequent photos. Rather than using labels, another variant of this approach instead presents large numbers of images and then allows the program to come up with its own clustering of cats. In both variants, however, it is not the rules programmed by people, but the patterns identified by the program that determine the outcome.

As already noted, both approaches were explored during the first AI wave. One example of an application of neural networks was Frank Rosenblatt’s ‘perceptron’, an algorithm he invented which learned to recognize letters without these being pre-programmed. This was attracted much media interest in the 1960s. Symbolic AI, however, remained dominant. The Logical Theory Machine and General Problem Solver mentioned earlier were both examples of systems within this strand. For decades it would remain the dominant approach within AI.

The proponents of symbolic AI also expressed much criticism of neural networks. They considered that approach unreliable and of limited use due to its lack of rules. In 1969 Marvin Minsky, an ardent supporter of the symbolic approach, wrote a book called Perceptrons with Seymour Papert. This amounted to a painstaking critique of the neural network approach, backed by examples of mathematical proofs of problems it could not solve. To many this appeared to sound the death knell for that approach. Footnote 53 Such criticism not only marginalized the position of neural networks, it also contributed towards the onset of the first AI winter.

3.3 The Second Wave

In 1982 Time magazine named the personal computer its Man of the Year. This coincided with a revival of interest in AI, and the discipline entered a second spring. At the time, the programming language Prolog was used for many logical reasoning systems. In 1982 the Japanese government invested a huge sum in a Prolog-based AI system in the form of the Fifth-Generation Computer Systems Project. Footnote 54 This was a far-reaching, ten-year partnership between the government and industry and was intended to boost the discipline in Japan by establishing a ‘parallel computing architecture’. At a time when there was widespread fear of Japanese economic growth, several Western countries quickly followed suit with their own projects.

To keep up with the competition, the US established the Microelectronics and Computer Technology Corporation (MCC), a research consortium. In 1984 MCC’s principal scientist, Douglas Lenat, launched a huge project called Cyc. Initiated with the full support of Marvin Minsky, this is still running today and involves collecting vast amounts of human knowledge about how the world works. Footnote 55 In 1983 DARPA, the scientific arm of the US Department of Defense, announced a Strategic Computing Initiative (SCI) that would invest one billion dollars in the field over ten years. Footnote 56 Both the Japanese and the American research projects took a broad approach to AI, with hardware and human interfaces also playing an important role, for example. Footnote 57 In 1983 the United Kingdom announced its response to the Japanese plans in the form of the Alvey Programme.

One important development during this second wave was the emergence in the 1970s of expert systems within symbolic AI. These are a form of rule-based AI where human experts in a particular domain are asked to formulate the rules for a program. One example was MYCIN, a program trained by medical experts to help doctors identify infectious diseases and prescribe appropriate medication. The Dendral project involved the analysis of molecules in organic chemistry. Expert systems were also developed to plan manufacturing processes and solve complex mathematical problems; for example, the Macsyma project. Such systems thus found practical applications outside the lab.

Some were developed in the Netherlands, too, in the 1980s and tested in pilot projects. These addressed themes including the implementation of social security and criminal sentencing policies. Footnote 58 In part thanks to specific research programmes and funding provided by the Dutch Research Council (Nederlandse Organisatie voor Wetenschappelijk Onderzoek, NWO) and various universities, but also by a number of government departments, the Netherlands was even able to establish an international profile with a relatively large research community in the field of legal knowledge-based systems. An important early facilitator in this respect was JURIX, the Foundation for Legal Knowledge-Based Systems, an organization of ‘legal tech’ researchers from the Netherlands and Flanders. It has held annual international conferences since 1988; their proceedings – all available online – testify to the rich Dutch and Flemish academic history of research on and development of AI applications in the legal domain. Footnote 59 Another prominent platform is the Benelux Association for Artificial Intelligence (Benelux Vereniging voor Kunstmatige Intelligentie, BNVKI), originally formed in the Netherlands in 1981 (as the NVKI) but later connecting scientists from Belgium and Luxembourg as well. The US Office for Technology Assessment has called expert systems “the first real commercial products of about 25 years of AI research” Footnote 60 and in 1984 the front page of The New York Times reported that they held out “the prospect of computer-aided decisions based on more wisdom than any one person can contain”. Footnote 61

Nevertheless, the results of this second wave were ultimately disappointing. The big ambitions of the major national projects were never achieved, either in Japan, the US or Europe. Their poor results were why the US SCI drastically scaled down its funding. Among the problems to limit the potential of these projects were hardware issues. This period culminated with the bankruptcy of several specialized companies in the field in the late 1980s. Footnote 62 But the expert systems also had their own problems. They tended to be highly complex, so minor errors in the rules had disastrous consequences for the results and systems could fail when two rules contradicted each other. Footnote 63 The Cyc project is still ongoing but has failed to live up to expectations throughout almost four decades of existence. Footnote 64 By the late 1980s, therefore, another AI winter had set in: the second wave had run out of momentum.

3.4 The Third Wave

In the 1990s, however, AI again began to attract attention and eventually flourish anew. Initially, the logical systems approach had several successes. One of the most iconic of these was the victory of IBM’s Deep Blue program over chess grandmaster Garry Kasparov, in 1997. At the time this was considered a fundamental breakthrough. The successor to that program, named Watson, later participated in the US television quiz show Jeopardy! , in which contestants have to formulate questions to match given answers. In 2011 Watson defeated the game’s reigning human champions. This was seen as proof that AI was approaching mastery of human language, another major breakthrough. Both cases are examples of the use of symbolic AI, in which the lessons of chess masters and answers from previous players of Jeopardy! were fed to the programs as rules. At the same time, however, experts were becoming increasingly dissatisfied with this approach.

Although both events were huge landmarks in the eyes of the public, in reality the truth was more prosaic. Stuart Russell describes how the foundations of chess algorithms were laid by Claude Shannon in 1950, with further innovations following in the 1960s. Thereafter, these programs improved according to a predictable pattern, in parallel with the growth of computing power. This was easily measurable against the scores recorded by human chess players. The linear pattern predicted that the score of a grandmaster would be achieved in the 1990s – exactly when Deep Blue defeated Kasparov. So that was not so much a breakthrough as a milestone that had been anticipated as part of a predictable pattern. Footnote 65 Deep Blue won by brute force, thanks to its superior computing power. Moreover, various chess champions had fed heuristic principles into its software. Instead of the smart computer beating the human, this victory could also be seen as the triumph of a collective comprising a computer program and numerous human players over a single grandmaster. Footnote 66 It was man and machine together that were superior to a human opponent.

The computer’s victory in Jeopardy! is also questionable. It would be incorrect to claim that the program could understand the complex natural language of humans. The game has a very formalized question-and-answer design, and many of the questions can be found on a typical Wikipedia page. This makes them relatively easy to answer for a program that can rapidly search mountains of information for keywords; that does not require an in-depth understanding of language.

While these logical systems only began to attract attention in the 1990s, other forms of AI had been making progress for far longer and the momentum eventually shifted towards the neural network approach. This trend had already begun in the mid-1980s when fundamental research into the so-called ‘backpropagation algorithm’ (in which multiple layers of neural networks are trained) improved the process of pattern recognition. At about the same time the US Department of Defense recognized that its funding programme had been unfairly neglecting the neural networks approach. Under the banner of ‘parallel distributed processing’, neural networks returned to centre stage in 1986. In a book published the previous year, John Haugeland had introduced the term GOFAI (‘good old-fashioned AI’) – a phrase which has since become a pejorative term for symbolic AI. In the same period Judea Pearl began applying probability theory rather than logical reasoning to AI.

Breakthroughs below the radar were thus undermining the dominant rule-based approach. A paper on backpropagation was rejected for a leading AI conference in the early 1980s and, according to Yann LeCun, researchers at the time even used code words to mask the fact that they were working with neural networks. Footnote 67 It took time for the importance of this new approach to become recognized. For example, Jeff Hawkins said in 2004 that AI had fewer skills than a mouse when it came to image recognition. Footnote 68

At that time, it was thought it would take another century before a computer could beat a human in the Asian game go, which has many more combinations of moves than chess. Footnote 69 In fact, Google’s AlphaGo program defeated world champion Lee Sedol in 2016. This was made possible thanks to recent breakthroughs in the approach to neural networks, in which researchers such as Yann LeCun and Andrew Ng played an important role. But it is Geoffrey Hinton who is often seen as the father of those advances. Together with David Rumelhart and Ronald Williams, he had already popularized the use of the backpropagation algorithm in a paper published in Nature in 1986. That algorithm traces the contribution made by the output layer back to hidden layers behind it, where individual units are identified that need to be modified to make the algorithm work more effectively. For a long time, the ‘backprop’ had only a single hidden layer, but more have recently been distinguished. Backpropagation thus addresses a central problem of ANNs: the representation of hierarchy. Relationships can now be distinguished at different levels and the success factors of the algorithm are also determined at all levels (called ‘credit assignment’). Footnote 70 Such neural networks have since been used, for instance, to simulate the price of shares on the stock exchange. Figure 2.7 shows the historical development of the two approaches to AI.

The transition from a symbolic to a connectionist AI

In 1989 Yann LeCun applied backprop to train neural networks to recognize handwritten postcodes. He used convolutional neural networks (CNNs), where complex images are broken down into smaller parts to make image recognition more efficient. This was another important contribution to contemporary AI programs. Footnote 71

In another paper, written in 2012, Hinton introduced the idea of ‘dropout’, which addresses the specific problem of ‘overfitting’ in neural network training. That occurs when a model focuses so strongly on training with existing data that it cannot effectively process new information. Hinton’s work gave an enormous boost to the applicability of neural networks in the field of machine learning. The use of multiple layers in the training process is why it is called ‘deep’ learning; each layer provides a more complex representation of the input based on the previous one. For example, while the first layer may be able to identify corners and dots, the second one can distinguish parts of a face such as the tip of a nose or the iris of an eye. The third layer can recognize whole noses and eyes, and so it goes on until you reach a layer that recognizes the face of an individual person (Box 2.2 ). Footnote 72

Box 2.2: Three Forms of Machine Learning

ML can be subdivided into three different forms: supervised, unsupervised and reinforcement learning. In supervised learning, a program is fed data with labels as in our earlier example of ‘cat’ versus ‘non-cat’. The algorithm is trained on that input and then tested to see if it can correctly apply the labels to new data.

Unsupervised learning has no training step and so the algorithm needs to search for patterns within the data by itself. It is fed large amounts of unlabelled data, in which it starts to recognize patterns of its own accord. The starting point here is that clusters of characteristics in the data will also form clusters in the future. Supervised learning is ideal when it is clear what is being searched for. If the researchers themselves are not yet sure what patterns are hidden within data and are curious to know what they are, then unsupervised learning is the more appropriate method.

The third form is applicable in other contexts, such as playing a game. Here it is not about giving a right or wrong answer or clustering data, but about strategies that can ultimately lead to winning or losing. In these cases, the reinforcement learning approach is more suitable. The algorithm is trained by rewarding it for following certain strategies. In recent years reinforcement learning has been applied to various classic computer games such as Pacman and the Atari portfolio, as well as to ‘normal’ card games and poker. The algorithm is given the goal of optimizing the value of the score and then correlates all kinds of actions with that score to develop an optimum strategy.

In 2012 Hinton’s team won an international competition in the field of ‘computer vision’ – image processing using AI. They achieved a margin of error of 16%, whereas no team before them had ever managed less than 25%. A few years earlier the same team had been successful in using neural networks for speech recognition after a demonstration by two students in Toronto. But the gains in computer vision in 2012 were the real revelation for many researchers. Footnote 73 Deep learning proliferated, and in 2017 almost all the teams in the competition could boast margins of error lower than 5% – comparable with human scores. That improvement continues to this day. The application of DL has since gained momentum, with the scientific breakthroughs using neural networks prompting an explosion of activity in this approach to AI. We are currently at the height of this latest AI summer. In the next chapter we look in more detail at the developments that has set in motion outside the lab: in the market and in wider society.

It is clear that the rapid expansion of AI in recent years has its origins in fundamental scientific research. Big companies like Google have subsequently rushed to hire talented researchers in this field, but it is scientists at universities who have been responsible for the most important breakthroughs.

In addition to these academic milestones, two other factors underlie the recent rise and application of AI. The first is the growth in processing power, as encapsulated in Moore’s Law. This pattern, that the number of transistors on a chip roughly doubles every two years, has been observed consistently in the computer industry for decades. It means that more and more computing power is becoming available while prices continue to fall. Hence the fact that the smartphones of today surpass the computing power of the very best computers of only a few decades ago. We noted earlier how the first ‘AI winter’ was caused in part by the combinatorial explosion. The increase in computing power provided the solution to this problem. A further leap in that power came from the chip industry, using graphic processing units (GPUs) rather than the classic central processing units (CPUs). GPUs were originally developed for complex graphics in the gaming industry but were subsequently found to enable many more parallel calculations in AI systems as well. Footnote 74 Since 2015, tensor processing units (TPUs) specifically designed for ML applications have also come into use.

The other factor that has contributed to the current AI wave is the increase in the amount of data. This is closely linked to the rise of the internet. In the past algorithms could only be applied to a limited range of data sources. In recent decades, however, as people have started to use the internet more and more, and directly and indirectly to generate a lot more digital information, the amount of data available for AI systems to analyse has increased significantly.

The ‘digital breadcrumbs’ we leave behind on the internet are now food for training AI algorithms. But we are helping with this training in other ways, too. By tagging personal names in photos on Facebook, for example, people provide algorithms with labels that can be used to train facial recognition software. One specific dataset that is very important for this kind of training is ImageNet, an open database of more than 14 million hand-labelled images. The ‘internet of things’ (the growing number of sensors and connections in the physical environment) is also contributing to the growth in data.

The triad of scientific breakthroughs, greater computing power and more data has allowed AI to take off in a big way recently (see Fig. 2.8 ). As mentioned, this expansion has been driven mostly by the application of machine learning as part of the neural network approach, and within ML by the development of deep learning.

Three drivers of progress in AI

Key Points: AI in the Lab

In the lab AI has ridden three waves of development. Between these were two ‘winters’ when scientific progress ground to a halt, hardware capacity was inadequate, and expectations were not met.

The first wave began with the Dartmouth Summer Research Project in 1956. At that time AI was used mainly for games such as draughts, in early robots and to solve mathematical problems. Two further waves, dominated by progress in symbolic AI and then neural networks, would follow.

The second wave began in the 1980s, driven in part by the international competition between Japan, the US and Europe. This produced expert systems, the first major commercial applications of AI.

The third wave began in the 1990s with major achievements in symbolic AI, but only properly gained momentum some years later due to advances in the field of machine learning and its subfield of deep learning. The scientific breakthroughs in this area, together with increases in computing power and data volumes, are the driving force behind this wave, which continues to this day.

Russell & Norvig, 2020 .

Nilsson, 2009 : 13.

See, for example, DenkWerk, 2018 .

High-Level Expert Group on Artificial Intelligence, 2019 . At the end of this document the authors expand on their initial definition with a detailed explanation of its various elements.

Bostrom, 2016 : 14.

Floridi, 2014 : 139.

Moravec, 1988 . The AI scientist Donald Knuth formulated it differently. He noticed that AI could do things that humans need to think about but failed at tasks humans do without thinking, like recognizing objects, analysing images and moving an arm (Bostrom, 2016 : 17).

Moravec, 1988 : 15.

Bostrom, 2016 .

Agrawal et al., 2018 : 2, 39. Drawing on the work of Jeff Hawkins, these authors believe that the foundation of intelligence is ‘prediction’.

Dennett, 2019 .

Dignum, 2019 .

Martin Ford ( 2018 ) interviewed 23 experts for his book Architects of Intelligence: The Truth about AI from the People Building It and asked them, ‘What year do you think human-level AI might be achieved, with a 50% probability?’ Most were only willing to respond anonymously and the year they suggested, on average, was 2099 – so almost 80 years from now. We will return to the potential of AGI in later chapters.

With regard to the terms ‘narrow AI’ and ‘weak AI’, we prefer the former. The latter obviously suggests that this type of AI lacks strength, whereas that may well not be the case. In fact, it is simply limited to a well-defined (read: ‘narrow’) domain. For example, a computer program may be very good at translating texts but still ‘narrow’ because it cannot be used for image recognition.

Domingos, 2017 : 286.

In Gods and Robots – Myths, Machines, and Ancient Dreams of Technology , Mayor ( 2018 ) examines the phenomenon of ‘made, not born’ in antiquity.

Zarkadakis, 2015 : 34.

Brynjolfsson & McAfee, 2014 : 250.

Described by Dijksterhuis in De mechanisering van het wereldbeeld (‘The mechanization of the world view’, 1950).

Russell, 2019 : 40.

Broussard, 2019 : 76.

Zarkadakis, 2015 : 37.

Wiener, 1964 : 57.

Zarkadakis, 2015 : 60–63.

Rid, 2016 : 83.

Boden, 2018 : 6.

Freeman & Louçã, 2001 : 309.

Rid, 2016 : 37–38.

Turing, 2009 [1950].

There has also been criticism of the use of language in the Turing test. Yann LeCun, a prominent AI scientist, suggested in an interview that there are forms of intelligence that have nothing to do with language (Ford, 2018 : 129). Some animals, for example, use less complex language than humans but still form good models of the world and can employ tools.

McCulloch & Pitts, 1943 .

In a lecture at Yale in the 1950s, the scientist John von Neumann described the similarity between the computer and the brain as follows: “The nervous pulses can clearly be viewed as (two-valued) markers, in the sense discussed previously: the absence of a pulse represents one value (say, the binary digit 0), and the presence of one represents the other (say, the binary digit 1).” von Neumann, 2012 [1958]: 43.

Freeman & Louçã, 2001 : 310.

Wiener, 2019 [1965].

Rid, 2016 : 47–52. Famous cyberneticians in various disciplines include the neurophysiologist Warren McCulloch, the physicist Heinz von Foerster, the management theorist Stafford Beer, the philosopher Humberto Maturana, the political scientist Karl Deutsch, the anthropologist Gregory Bateson and the sociologist Talcott Parsons.

Rid, 2016 : 53–55.

Moravec, 1988 : 7.

The history of a scientific discipline can be written in several ways. It can focus on the fundamental science, for instance, or on practical inventions and applications. One example is the difference between the development of the natural sciences and the inventions of the Industrial Revolution. In this chapter we combine both perspectives, but the idea of waves in AI is rooted mainly in that of inventions and applications.

Bostrom, 2016 : 6.

Broussard, 2019 : 69–70.

Wiener, 1964 : 22–24. It would eventually take thirty years for a computer to defeat a chess grandmaster, as we shall see shortly. In any case, people have not lost their interest in these games since sophisticated programs have learned to play them.

Boden, 2018 : 10. In this logical problem, three entities all have to cross a river. Only two can cross at the same time. Each entity threatens to harm one of the others, so not every duo can cross together. The problem is: which combinations can be formed to convey everyone to the other side unharmed?

Russell, 2019 : 52.

Rid, 2016 : 136.

Brynjolfsson & McAfee, 2014 : 141.

Rid, 2016 : 148. The writer Vernor Vinge would later coin the term ‘singularity’ for this scenario.

Bakker & Korsten, 2021 : 24.

Leung, 2019 : 253.

Russell & Norvig, 2021 : 21.

Domingos, 2017 : 93.

Greenfield, 2017 : 214.

From an interview with Geoffrey Hinton (Ford, 2018 : 83).

Russell, 2019 : 271.

Domingos, 2017 : 35.

Leung, 2019 : 254.

Russell & Norvig, 2020 : 24.

Hage & Verheij, 1999 .

www.jurix.nl/proceedings/

Leung, 2019 : 259.

Dreyfus & Dreyfus, 1986 : ix.

Leung, 2019 : 255.

The idea of expressing the limits of human behaviour and language in rules had been explored earlier by philosophers such as Ludwig Wittgenstein (Wittgenstein, 1984 ).

According to Ray Kurzweil, a proponent of neural networks, Cyc has actually achieved almost nothing (Ford, 2018 : 233). That, however, is an oversimplification. Such projects form the foundations of techniques such as knowledge graphs, which are now important for the functioning of search engines like Google. This also demonstrates why the two approaches are not mutually exclusive and in practice often go hand in hand.

Russell, 2019 : 62–63.

Ihde, 2010 .

From an interview with Yann LeCun (Ford, 2018 : 122).

Tegmark, 2017 : 79.

Tonin, 2019 : 1.

In the book Perceptrons , which was highly critical of the neural networks approach, Minsky and Papert demonstrated that it was unable to solve the problem of the ‘exclusive OR’ (XOR). But Rumelhart, Hinton and Williams showed that backpropagation could learn XOR.

Marcus & Davi, 2019 : 52.

Domingos, 2017 : 117.

From an interview with Geoffrey Hinton (Ford, 2018 : 77).

Kelly, 2017 : 38.

Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction machines: The simple economics of Artificial Intelligence . Harvard Business Press.

Google Scholar  

Bakker, S., & Korsten, P. (2021). Artificiële Intelligentie Als Een general purpose technology: Strategische Belangen Van Verantwoorde Inzet In Historisch Perspectief (WRR Working Paper nr. 41). Wetenschappelijke Raad voor het Regeringsbeleid. Available at: https://www.wrr.nl/publicaties/working-papers/2021/02/16/artificiele-intelligentie-als-een-general-purpose-technology

Boden, M. (2018). Artificial Intelligence: A very short introduction . Oxford University Press.

Book   Google Scholar  

Bostrom, N. (2016). Superintelligence: Paths, Dangers, Strategies . Oxford University Press.

Broussard, M. (2019). Artificial Unintelligence: How computers misunderstand the world . MIT Press.

Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies . WW Norton and Company.

DenkWerk. (2018). Artificial Intelligence in Nederland: Zelf Aan Het Stuur . Available at: https://denkwerk.online/media/1029/artitificial_intelligence_in_nederland_juli_2018.pdf

Dennett, D. (2019). What can we do? In J. Brockman (red.), Possible minds: Twenty-five ways of looking at AI (pp. 41–53). Penguin.

Dignum, V. (2019). Responsible Artificial Intelligence: How to develop and use AI in a responsible way . Springer.

Domingos, P. (2017). The master algorithm: How the Quest for the ultimate learning machine will remake our world . Penguin Random House.

Dreyfus, H., & Dreyfus, S. (1986). Mind over Machine . The Free Press.

Floridi, L. (2014). The Fourth Revolution: How the Infosphere is reshaping human reality . Oxford University Press.

Ford, M. (2018). Architects of Intelligence . Packt Publishing.

Freeman, C., & Louçã, F. (2001). As time Goes By: From the industrial revolutions to the information revolution . Oxford University Press.

Greenfield, A. (2017). Radical technologies: The design of everyday life . Verso Books.

Hage, J., & Verheij, B. (1999). Rechtsinformatica: De Stand Van Zaken In De Wetenschap. In A. Oskamp and A. Lodder (reds.), Informatietechnologie voor juristen. Handboek voor de jurist in de 21e eeuw (pp. 65–92). Kluwer.

High-Level Expert Group on Artificial Intelligence. (2019). A definition of AI: Main capabilities and scientific disciplines . European Commission. Available at: https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=56341

Ihde, D. (2010). Embodied technics . Automatic Press/vip.

Kelly, K. (2017). The Inevitable: Understanding the 12 technological forces that will shape our future . Penguin.

Leung, J. (2019). Who will govern Artificial Intelligence? Learning from the history of strategic politics in emerging technologies . Dissertation, Oxford University. Available at: https://ora.ox.ac.uk/objects/uuid:ea3c7cb8-2464-45f1-a47c-c7b568f27665

Marcus, G., & Davi, E. (2019). Rebooting AI: Building Artificial Intelligence we can trust . Vintage.

Mayor, A. (2018). Gods and Robots: Myths, machines, and ancient dreams of technology . Princeton University Press.

McCulloch, W., & Pitts, W. (1943). A Logical Calculus of the ideas immanent in nervous activity. The Bulletin of Mathematical Biophysics, 5 (4), 115–133.

Article   Google Scholar  

Moravec, H. (1988). Mind Children: The future of robot and human intelligence . Harvard University Press.

Nilsson, N. (2009). The Quest for Artificial Intelligence . Cambridge University Press.

Rid, T. (2016). Rise of the machines: A cybernetic history . WW Norton & Company.

Russell, S. (2019). Human compatible: Artificial Intelligence and the problem of control . Penguin.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A modern approach (4th ed.). Pearson.

Russell, S., & Norvig, P. (2021). Artificial Intelligence: A modern approach . Pearson.

Tegmark, M. (2017). Life 3.0: Being Human in the age of Artificial Intelligence . Penguin.

Tonin, M. (2019). Artificial Intelligence: Implications for NATO’s Armed Forces. 149 stctts 19 E rev. 1 fin.

Turing, A. (2009 [1950]). Computing machinery and Intelligence. In R. Epstein, G. Roberts, and G. Beber (reds.), Parsing the turing test . Springer.

von Neumann, J. (2012 [1958]). The Computer and the Brain. Yale University Press.

Wiener, N. (1964). God and Golem, Inc.: A comment on certain points where cybernetics impinges on religion . MIT Press.

Wiener (2019 [1965]). Cybernetics: Or control and communication in the animal and the machine. MIT Press.

Wittgenstein, L. (1984). Tractatus logico-philosophicus. Tagebücher 1914–1916. Philosophische Untersuchungen . Suhrkamp.

Zarkadakis, G. (2015). In our own image: Will artificial intelligence save or destroy us? Ebury Publishing.

Download references

Author information

Authors and affiliations.

Wetenschappelijke Raad voor het Regeringsbeleid, Den Haag, The Netherlands

Haroon Sheikh, Corien Prins & Erik Schrijvers

You can also search for this author in PubMed   Google Scholar

Rights and permissions

Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

Reprints and permissions

Copyright information

© 2023 The Author(s)

About this chapter

Sheikh, H., Prins, C., Schrijvers, E. (2023). Artificial Intelligence: Definition and Background. In: Mission AI. Research for Policy. Springer, Cham. https://doi.org/10.1007/978-3-031-21448-6_2

Download citation

DOI : https://doi.org/10.1007/978-3-031-21448-6_2

Published : 31 January 2023

Publisher Name : Springer, Cham

Print ISBN : 978-3-031-21447-9

Online ISBN : 978-3-031-21448-6

eBook Packages : Social Sciences Social Sciences (R0)

Share this chapter

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

More From Forbes

A very short history of artificial intelligence (ai).

• Share to Facebook
• Share to Twitter
• Share to Linkedin

Shutterstock

1308                                  Catalan poet and theologian Ramon Llull publishes Ars generalis ultima (The Ultimate General Art), further perfecting his method of using paper-based mechanical means to create new knowledge from combinations of concepts.

1666                                  Mathematician and philosopher Gottfried Leibniz publishes Dissertatio de arte combinatoria (On the Combinatorial Art), following Ramon Llull in proposing an alphabet of human thought and arguing that all ideas are nothing but combinations of a relatively small number of simple concepts.

1726                                  Jonathan Swift publishes Gulliver's Travels , which includes a description of the Engine , a machine on the island of Laputa (and a parody of Llull's ideas): "a Project for improving speculative Knowledge by practical and mechanical Operations." By using this "Contrivance," "the most ignorant Person at a reasonable Charge, and with a little bodily Labour, may write Books in Philosophy, Poetry, Politicks, Law, Mathematicks, and Theology, with the least Assistance from Genius or study."

1763                                  Thomas Bayes develops a framework for reasoning about the probability of events. Bayesian inference will become a leading approach in machine learning.

1854                                  George Boole argues that logical reasoning could be performed systematically in the same manner as solving a system of equations.

1898                                  At an electrical exhibition in the recently completed Madison Square Garden, Nikola Tesla makes a demonstration of the world’s first radio-controlled vessel . The boat was equipped with, as Tesla described, “a borrowed mind.”

1914                                  The Spanish engineer Leonardo Torres y Quevedo demonstrates the first chess-playing machine, capable of king and rook against king endgames without any human intervention.

1921                                  Czech writer Karel Čapek introduces the word "robot" in his play R.U.R. (Rossum's Universal Robots). The word "robot" comes from the word "robota" (work).

1925                                  Houdina Radio Control releases a radio-controlled driverless car , travelling the streets of New York City.

1927                                  The science-fiction film Metropolis is released. It features a robot double of a peasant girl, Maria, which unleashes chaos in Berlin of 2026—it was the first robot depicted on film, inspiring the Art Deco look of C-3PO in Star Wars .

1929                                  Makoto Nishimura designs Gakutensoku , Japanese for "learning from the laws of nature," the first robot built in Japan. It could change its facial expression and move its head and hands via an air pressure mechanism.

1943                                  Warren S. McCulloch and Walter Pitts publish “A Logical Calculus of the Ideas Immanent in Nervous Activity” in the Bulletin of Mathematical Biophysics . This influential paper, in which they discussed networks of idealized and simplified artificial “neurons” and how they might perform simple logical functions, will become the inspiration for computer-based “neural networks” (and later “deep learning”) and their popular description as “mimicking the brain.”

1949                                  Edmund Berkeley publishes Giant Brains: Or Machines That Think in which he writes: “Recently there have been a good deal of news about strange giant machines that can handle information with vast speed and skill….These machines are similar to what a brain would be if it were made of hardware and wire instead of flesh and nerves… A machine can handle information; it can calculate, conclude, and choose; it can perform reasonable operations with information. A machine, therefore, can think.”

1949                                  Donald Hebb publishes Organization of Behavior: A Neuropsychological Theory in which he proposes a theory about learning based on conjectures regarding neural networks and the ability of synapses to strengthen or weaken over time.

1950                                  Claude Shannon’s “Programming a Computer for Playing Chess” is the first published article on developing a chess-playing computer program.

1950                                  Alan Turing publishes “Computing Machinery and Intelligence” in which he proposes “the imitation game” which will later become known as the “Turing Test.”

1951                                  Marvin Minsky and Dean Edmunds build SNARC (Stochastic Neural Analog Reinforcement Calculator), the first artificial neural network, using 3000 vacuum tubes to simulate a network of 40 neurons.

1952                                  Arthur Samuel develops the first computer checkers-playing program and the first computer program to learn on its own.

August 31, 1955              The term “artificial intelligence” is coined in a proposal for a “2 month, 10 man study of artificial intelligence” submitted by John McCarthy (Dartmouth College), Marvin Minsky (Harvard University), Nathaniel Rochester (IBM), and Claude Shannon (Bell Telephone Laboratories). The workshop, which took place a year later, in July and August 1956, is generally considered as the official birthdate of the new field.

December 1955               Herbert Simon and Allen Newell develop the Logic Theorist , the first artificial intelligence program, which eventually would prove 38 of the first 52 theorems in Whitehead and Russell's Principia Mathematica .

1957                                  Frank Rosenblatt develops the Perceptron, an early artificial neural network enabling pattern recognition based on a two-layer computer learning network. The New York Times reported the Perceptron to be "the embryo of an electronic computer that [the Navy] expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence." The New Yorker called it a “remarkable machine… capable of what amounts to thought.”

1958                                  John McCarthy develops programming language Lisp which becomes the most popular programming language used in artificial intelligence research.

1959                                  Arthur Samuel coins the term “machine learning,” reporting on programming a computer “so that it will learn to play a better game of checkers than can be played by the person who wrote the program.”

1959                                  Oliver Selfridge publishes “ Pandemonium: A paradigm for learning ” in the Proceedings of the Symposium on Mechanization of Thought Processes , in which he describes a model for a process by which computers could recognize patterns that have not been specified in advance.

1959                                  John McCarthy publishes “ Programs with Common Sense ” in the Proceedings of the Symposium on Mechanization of Thought Processes , in which he describes the Advice Taker, a program for solving problems by manipulating sentences in formal languages with the ultimate objective of making programs “that learn from their experience as effectively as humans do.”

1961                                  The first industrial robot, Unimate , starts working on an assembly line in a General Motors plant in New Jersey.

1961                                  James Slagle develops SAINT (Symbolic Automatic INTegrator), a heuristic program that solved symbolic integration problems in freshman calculus.

1964                                  Daniel Bobrow completes his MIT PhD dissertation titled “ Natural Language Input for a Computer Problem Solving System ” and develops STUDENT, a natural language understanding computer program.

1965                                  Herbert Simon predicts that "machines will be capable, within twenty years, of doing any work a man can do."

1965                                  Hubert Dreyfus publishes "Alchemy and AI," arguing that the mind is not like a computer and that there were limits beyond which AI would not progress.

1965                                  I.J. Good writes in "Speculations Concerning the First Ultraintelligent Machine" that “ the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”

1965                                  Joseph Weizenbaum develops ELIZA , an interactive program that carries on a dialogue in English language on any topic. Weizenbaum, who wanted to demonstrate the superficiality of communication between man and machine, was surprised by the number of people who attributed human-like feelings to the computer program.

1965                                  Edward Feigenbaum, Bruce G. Buchanan, Joshua Lederberg, and Carl Djerassi start working on DENDRAL at Stanford University. The first expert system, it automated the decision-making process and problem-solving behavior of organic chemists, with the general aim of studying hypothesis formation and constructing models of empirical induction in science.

1966                                  Shakey the robot is the first general-purpose mobile robot to be able to reason about its own actions. In a Life magazine 1970 article about this “first electronic person,” Marvin Minsky is quoted saying with “certitude”: “In from three to eight years we will have a machine with the general intelligence of an average human being.”

1968                                  The film 2001: Space Odyssey is released, featuring Hal, a sentient computer.

1968                                  Terry Winograd develops SHRDLU , an early natural language understanding computer program.

1969                                  Arthur Bryson and Yu-Chi Ho describe backpropagation as a multi-stage dynamic system optimization method. A learning algorithm for multi-layer artificial neural networks, it has contributed significantly to the success of deep learning in the 2000s and 2010s, once computing power has sufficiently advanced to accommodate the training of large networks.

1969                                  Marvin Minsky and Seymour Papert publish Perceptrons: An Introduction to Computational Geometry , highlighting the limitations of simple neural networks.  In an expanded edition published in 1988, they responded to claims that their 1969 conclusions significantly reduced funding for neural network research: “Our version is that progress had already come to a virtual halt because of the lack of adequate basic theories… by the mid-1960s there had been a great many experiments with perceptrons, but no one had been able to explain why they were able to recognize certain kinds of patterns and not others.”

1970                                  The first anthropomorphic robot, the WABOT-1 , is built at Waseda University in Japan. It consisted of a limb-control system, a vision system and a conversation system.

1972                                  MYCIN , an early expert system for identifying bacteria causing severe infections and recommending antibiotics, is developed at Stanford University.

1973                                  James Lighthill reports to the British Science Research Council on the state artificial intelligence research, concluding that "in no part of the field have discoveries made so far produced the major impact that was then promised," leading to drastically reduced government support for AI research.

1976                                  Computer scientist Raj Reddy publishes “Speech Recognition by Machine: A Review” in the Proceedings of the IEEE , summarizing the early work on Natural Language Processing (NLP).

1978                                  The XCON (eXpert CONfigurer) program, a rule-based expert system assisting in the ordering of DEC's VAX computers by automatically selecting the components based on the customer's requirements, is developed at Carnegie Mellon University.

1979                                  The Stanford Cart successfully crosses a chair-filled room without human intervention in about five hours, becoming one of the earliest examples of an autonomous vehicle.

1980                                  Wabot-2 is built at Waseda University in Japan, a musician humanoid robot able to communicate with a person, read a musical score and play tunes of average difficulty on an electronic organ.

1981                                  The Japanese Ministry of International Trade and Industry budgets $850 million for the Fifth Generation Computer project. The project aimed to develop computers that could carry on conversations, translate languages, interpret pictures, and reason like human beings.

1984                                  ­Electric Dreams is released, a film about a love triangle between a man, a woman and a personal computer.

1984                                  At the annual meeting of AAAI, Roger Schank and Marvin Minsky warn of the coming “ AI Winter ,” predicting an immanent bursting of the AI bubble (which did happen three years later), similar to the reduction in AI investment and research funding in the mid-1970s.

1986                                  First driverless car, a Mercedes-Benz van equipped with cameras and sensors, built at Bundeswehr University in Munich under the direction of Ernst Dickmanns , drives up to 55 mph on empty streets.

October 1986                   David Rumelhart, Geoffrey Hinton, and Ronald Williams publish ” Learning representations by back-propagating errors ,” in which they describe “a new learning procedure, back-propagation, for networks of neurone-like units.”

1987                                  The video Knowledge Navigator , accompanying Apple CEO John Sculley’s keynote speech at Educom, envisions a future in which “knowledge applications would be accessed by smart agents working over networks connected to massive amounts of digitized information.”

1988                                  Judea Pearl publishes Probabilistic Reasoning in Intelligent Systems . His 2011 Turing Award citation reads: “Judea Pearl created the representational and computational foundation for the processing of information under uncertainty. He is credited with the invention of Bayesian networks, a mathematical formalism for defining complex probability models, as well as the principal algorithms used for inference in these models. This work not only revolutionized the field of artificial intelligence but also became an important tool for many other branches of engineering and the natural sciences.”

1988                                  Rollo Carpenter develops the chat-bot Jabberwacky to "simulate natural human chat in an interesting, entertaining and humorous manner." It is an early attempt at creating artificial intelligence through human interaction.

1988                                  Members of the IBM T.J. Watson Research Center publish “ A statistical approach to language translation ,” heralding the shift from rule-based to probabilistic methods of machine translation, and reflecting a broader shift to “machine learning” based on statistical analysis of known examples, not comprehension and “understanding” of the task at hand (IBM’s project Candide, successfully translating between English and French, was based on 2.2 million pairs of sentences, mostly from the bilingual proceedings of the Canadian parliament).

1988                                  Marvin Minsky and Seymour Papert publish an expanded edition of their 1969 book Perceptrons . In “Prologue: A View from 1988” they wrote: “One reason why progress has been so slow in this field is that researchers unfamiliar with its history have continued to make many of the same mistakes that others have made before them.”

1989                                  Yann LeCun and other researchers at AT&T Bell Labs successfully apply a backpropagation algorithm to a multi-layer neural network, recognizing handwritten ZIP codes. Given the hardware limitations at the time, it took about 3 days (still a significant improvement over earlier efforts) to train the network.

1990                                  Rodney Brooks publishes “ Elephants Don’t Play Chess ,” proposing a new approach to AI—building intelligent systems, specifically robots, from the ground up and on the basis of ongoing physical interaction with the environment: “The world is its own best model… The trick is to sense it appropriately and often enough.”

1993                                  Vernor Vinge publishes “ The Coming Technological Singularity ,” in which he predicts that “within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.”

1995                                  Richard Wallace develops the chatbot A.L.I.C.E (Artificial Linguistic Internet Computer Entity), inspired by Joseph Weizenbaum's ELIZA program, but with the addition of natural language sample data collection on an unprecedented scale, enabled by the advent of the Web.

1997                                  Sepp Hochreiter and Jürgen Schmidhuber propose Long Short-Term Memory (LSTM), a type of a recurrent neural network used today in handwriting recognition and speech recognition.

1997                                  Deep Blue becomes the first computer chess-playing program to beat a reigning world chess champion.

1998                                  Dave Hampton and Caleb Chung create Furby , the first domestic or pet robot.

1998                                  Yann LeCun, Yoshua Bengio and others publish papers on the application of neural networks to handwriting recognition and on optimizing backpropagation .

2000                                  MIT’s Cynthia Breazeal develops Kismet , a robot that could recognize and simulate emotions.

2000                                  Honda's ASIMO robot, an artificially intelligent humanoid robot, is able to walk as fast as a human, delivering trays to customers in a restaurant setting.

2001                                  A.I. Artificial Intelligence is released, a Steven Spielberg film about David, a childlike android uniquely programmed with the ability to love.

2004                                  The first DARPA Grand Challenge , a prize competition for autonomous vehicles, is held in the Mojave Desert. None of the autonomous vehicles finished the 150-mile route.

2006                                  Oren Etzioni, Michele Banko, and Michael Cafarella coin the term “ machine reading ,” defining it as an inherently unsupervised “autonomous understanding of text.”

2006                                  Geoffrey Hinton publishes “ Learning Multiple Layers of Representation , ” summarizing the ideas that have led to “multilayer neural networks that contain top-down connections and training them to generate sensory data rather than to classify it,” i.e., the new approaches to deep learning.

2007                                  Fei Fei Li and colleagues at Princeton University start to assemble ImageNet , a large database of annotated images designed to aid in visual object recognition software research.

2009                                  Rajat Raina, Anand Madhavan and Andrew Ng publish “ Large-scale Deep Unsupervised Learning using Graphics Processors ,” arguing that “modern graphics processors far surpass the computational capabilities of multicore CPUs, and have the potential to revolutionize the applicability of deep unsupervised learning methods.”

2009                                  Google starts developing, in secret, a driverless car . In 2014, it became the first to pass, in Nevada , a U.S. state self-driving test.

2009                                  Computer scientists at the Intelligent Information Laboratory at Northwestern University develop Stats Monkey , a program that writes sport news stories without human intervention.

2010                                  Launch of the ImageNet Large Scale Visual Recognition Challenge (ILSVCR), an annual AI object recognition competition.

2011                                  A convolutional neural network wins the German Traffic Sign Recognition competition with 99.46% accuracy (vs. humans at 99.22%).

2011                                  Watson, a natural language question answering computer, competes on Jeopardy! and defeats two former champions.

2011                                  Researchers at the IDSIA in Switzerland report a 0.27% error rate in handwriting recognition using convolutional neural networks, a significant improvement over the 0.35%-0.40% error rate in previous years.

June 2012                         Jeff Dean and Andrew Ng report on an experiment in which they showed a very large neural network 10 million unlabeled images randomly taken from YouTube videos, and “to our amusement, one of our artificial neurons learned to respond strongly to pictures of... cats.”

October 2012                   A convolutional neural network designed by researchers at the University of Toronto achieve an error rate of only 16% in the ImageNet Large Scale Visual Recognition Challenge , a significant improvement over the 25% error rate achieved by the best entry the year before.

March 2016                      Google DeepMind's AlphaGo defeats Go champion Lee Sedol.

The Web (especially Wikipedia) is a great source for the history of artificial intelligence. Other key sources include Nils Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements ; Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach ;  Pedro Domingo s, The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World ; and Artificial Intelligence and Life in 2030 . Please comment regarding inadvertent omissions and inaccuracies.

See also A Very Short History of Data Science , A Very Short History of Big Data , and A Very Short History of Information Technology (IT) .

• Editorial Standards
• Reprints & Permissions

This week: the arXiv Accessibility Forum

Help | Advanced Search

Computer Science > Computation and Language

Title: sparks of artificial general intelligence: early experiments with gpt-4.

Abstract: Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition. The latest model developed by OpenAI, GPT-4, was trained using an unprecedented scale of compute and data. In this paper, we report on our investigation of an early version of GPT-4, when it was still in active development by OpenAI. We contend that (this early version of) GPT-4 is part of a new cohort of LLMs (along with ChatGPT and Google's PaLM for example) that exhibit more general intelligence than previous AI models. We discuss the rising capabilities and implications of these models. We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting. Moreover, in all of these tasks, GPT-4's performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. In our exploration of GPT-4, we put special emphasis on discovering its limitations, and we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction. We conclude with reflections on societal influences of the recent technological leap and future research directions.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

10 blog links

Bibtex formatted citation.

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

The Impact of Artificial Intelligence on Academics: A Concise Overview

• 7(4):218-220

• University of Benghazi

Discover the world's research

• 25+ million members
• 160+ million publication pages
• 2.3+ billion citations

• SCI ENG ETHICS

• Maud Chassignol

• James C. Lester
• Ashraf Alam
• EDUC TECHNOL SOC

• Recruit researchers
• Join for free
• Login Email Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google Welcome back! Please log in. Email · Hint Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google No account? Sign up

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Tzu Chi Med J
• v.32(4); Oct-Dec 2020

The impact of artificial intelligence on human society and bioethics

Michael cheng-tek tai.

Department of Medical Sociology and Social Work, College of Medicine, Chung Shan Medical University, Taichung, Taiwan

Artificial intelligence (AI), known by some as the industrial revolution (IR) 4.0, is going to change not only the way we do things, how we relate to others, but also what we know about ourselves. This article will first examine what AI is, discuss its impact on industrial, social, and economic changes on humankind in the 21 st century, and then propose a set of principles for AI bioethics. The IR1.0, the IR of the 18 th century, impelled a huge social change without directly complicating human relationships. Modern AI, however, has a tremendous impact on how we do things and also the ways we relate to one another. Facing this challenge, new principles of AI bioethics must be considered and developed to provide guidelines for the AI technology to observe so that the world will be benefited by the progress of this new intelligence.

W HAT IS ARTIFICIAL INTELLIGENCE ?

Artificial intelligence (AI) has many different definitions; some see it as the created technology that allows computers and machines to function intelligently. Some see it as the machine that replaces human labor to work for men a more effective and speedier result. Others see it as “a system” with the ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation [ 1 ].

Despite the different definitions, the common understanding of AI is that it is associated with machines and computers to help humankind solve problems and facilitate working processes. In short, it is an intelligence designed by humans and demonstrated by machines. The term AI is used to describe these functions of human-made tool that emulates the “cognitive” abilities of the natural intelligence of human minds [ 2 ].

Along with the rapid development of cybernetic technology in recent years, AI has been seen almost in all our life circles, and some of that may no longer be regarded as AI because it is so common in daily life that we are much used to it such as optical character recognition or the Siri (speech interpretation and recognition interface) of information searching equipment on computer [ 3 ].

D IFFERENT TYPES OF ARTIFICIAL INTELLIGENCE

From the functions and abilities provided by AI, we can distinguish two different types. The first is weak AI, also known as narrow AI that is designed to perform a narrow task, such as facial recognition or Internet Siri search or self-driving car. Many currently existing systems that claim to use “AI” are likely operating as a weak AI focusing on a narrowly defined specific function. Although this weak AI seems to be helpful to human living, there are still some think weak AI could be dangerous because weak AI could cause disruptions in the electric grid or may damage nuclear power plants when malfunctioned.

The new development of the long-term goal of many researchers is to create strong AI or artificial general intelligence (AGI) which is the speculative intelligence of a machine that has the capacity to understand or learn any intelligent task human being can, thus assisting human to unravel the confronted problem. While narrow AI may outperform humans such as playing chess or solving equations, but its effect is still weak. AGI, however, could outperform humans at nearly every cognitive task.

Strong AI is a different perception of AI that it can be programmed to actually be a human mind, to be intelligent in whatever it is commanded to attempt, even to have perception, beliefs and other cognitive capacities that are normally only ascribed to humans [ 4 ].

In summary, we can see these different functions of AI [ 5 , 6 ]:

• Automation: What makes a system or process to function automatically
• Machine learning and vision: The science of getting a computer to act through deep learning to predict and analyze, and to see through a camera, analog-to-digital conversion and digital signal processing
• Natural language processing: The processing of human language by a computer program, such as spam detection and converting instantly a language to another to help humans communicate
• Robotics: A field of engineering focusing on the design and manufacturing of cyborgs, the so-called machine man. They are used to perform tasks for human's convenience or something too difficult or dangerous for human to perform and can operate without stopping such as in assembly lines
• Self-driving car: Use a combination of computer vision, image recognition amid deep learning to build automated control in a vehicle.

D O HUMAN-BEINGS REALLY NEED ARTIFICIAL INTELLIGENCE ?

Is AI really needed in human society? It depends. If human opts for a faster and effective way to complete their work and to work constantly without taking a break, yes, it is. However if humankind is satisfied with a natural way of living without excessive desires to conquer the order of nature, it is not. History tells us that human is always looking for something faster, easier, more effective, and convenient to finish the task they work on; therefore, the pressure for further development motivates humankind to look for a new and better way of doing things. Humankind as the homo-sapiens discovered that tools could facilitate many hardships for daily livings and through tools they invented, human could complete the work better, faster, smarter and more effectively. The invention to create new things becomes the incentive of human progress. We enjoy a much easier and more leisurely life today all because of the contribution of technology. The human society has been using the tools since the beginning of civilization, and human progress depends on it. The human kind living in the 21 st century did not have to work as hard as their forefathers in previous times because they have new machines to work for them. It is all good and should be all right for these AI but a warning came in early 20 th century as the human-technology kept developing that Aldous Huxley warned in his book Brave New World that human might step into a world in which we are creating a monster or a super human with the development of genetic technology.

Besides, up-to-dated AI is breaking into healthcare industry too by assisting doctors to diagnose, finding the sources of diseases, suggesting various ways of treatment performing surgery and also predicting if the illness is life-threatening [ 7 ]. A recent study by surgeons at the Children's National Medical Center in Washington successfully demonstrated surgery with an autonomous robot. The team supervised the robot to perform soft-tissue surgery, stitch together a pig's bowel, and the robot finished the job better than a human surgeon, the team claimed [ 8 , 9 ]. It demonstrates robotically-assisted surgery can overcome the limitations of pre-existing minimally-invasive surgical procedures and to enhance the capacities of surgeons performing open surgery.

Above all, we see the high-profile examples of AI including autonomous vehicles (such as drones and self-driving cars), medical diagnosis, creating art, playing games (such as Chess or Go), search engines (such as Google search), online assistants (such as Siri), image recognition in photographs, spam filtering, predicting flight delays…etc. All these have made human life much easier and convenient that we are so used to them and take them for granted. AI has become indispensable, although it is not absolutely needed without it our world will be in chaos in many ways today.

T HE IMPACT OF ARTIFICIAL INTELLIGENCE ON HUMAN SOCIETY

Negative impact.

Questions have been asked: With the progressive development of AI, human labor will no longer be needed as everything can be done mechanically. Will humans become lazier and eventually degrade to the stage that we return to our primitive form of being? The process of evolution takes eons to develop, so we will not notice the backsliding of humankind. However how about if the AI becomes so powerful that it can program itself to be in charge and disobey the order given by its master, the humankind?

Let us see the negative impact the AI will have on human society [ 10 , 11 ]:

• A huge social change that disrupts the way we live in the human community will occur. Humankind has to be industrious to make their living, but with the service of AI, we can just program the machine to do a thing for us without even lifting a tool. Human closeness will be gradually diminishing as AI will replace the need for people to meet face to face for idea exchange. AI will stand in between people as the personal gathering will no longer be needed for communication
• Unemployment is the next because many works will be replaced by machinery. Today, many automobile assembly lines have been filled with machineries and robots, forcing traditional workers to lose their jobs. Even in supermarket, the store clerks will not be needed anymore as the digital device can take over human labor
• Wealth inequality will be created as the investors of AI will take up the major share of the earnings. The gap between the rich and the poor will be widened. The so-called “M” shape wealth distribution will be more obvious
• New issues surface not only in a social sense but also in AI itself as the AI being trained and learned how to operate the given task can eventually take off to the stage that human has no control, thus creating un-anticipated problems and consequences. It refers to AI's capacity after being loaded with all needed algorithm may automatically function on its own course ignoring the command given by the human controller
• The human masters who create AI may invent something that is racial bias or egocentrically oriented to harm certain people or things. For instance, the United Nations has voted to limit the spread of nucleus power in fear of its indiscriminative use to destroying humankind or targeting on certain races or region to achieve the goal of domination. AI is possible to target certain race or some programmed objects to accomplish the command of destruction by the programmers, thus creating world disaster.

P OSITIVE IMPACT

There are, however, many positive impacts on humans as well, especially in the field of healthcare. AI gives computers the capacity to learn, reason, and apply logic. Scientists, medical researchers, clinicians, mathematicians, and engineers, when working together, can design an AI that is aimed at medical diagnosis and treatments, thus offering reliable and safe systems of health-care delivery. As health professors and medical researchers endeavor to find new and efficient ways of treating diseases, not only the digital computer can assist in analyzing, robotic systems can also be created to do some delicate medical procedures with precision. Here, we see the contribution of AI to health care [ 7 , 11 ]:

Fast and accurate diagnostics

IBM's Watson computer has been used to diagnose with the fascinating result. Loading the data to the computer will instantly get AI's diagnosis. AI can also provide various ways of treatment for physicians to consider. The procedure is something like this: To load the digital results of physical examination to the computer that will consider all possibilities and automatically diagnose whether or not the patient suffers from some deficiencies and illness and even suggest various kinds of available treatment.

Socially therapeutic robots

Pets are recommended to senior citizens to ease their tension and reduce blood pressure, anxiety, loneliness, and increase social interaction. Now cyborgs have been suggested to accompany those lonely old folks, even to help do some house chores. Therapeutic robots and the socially assistive robot technology help improve the quality of life for seniors and physically challenged [ 12 ].

Reduce errors related to human fatigue

Human error at workforce is inevitable and often costly, the greater the level of fatigue, the higher the risk of errors occurring. Al technology, however, does not suffer from fatigue or emotional distraction. It saves errors and can accomplish the duty faster and more accurately.

Artificial intelligence-based surgical contribution

AI-based surgical procedures have been available for people to choose. Although this AI still needs to be operated by the health professionals, it can complete the work with less damage to the body. The da Vinci surgical system, a robotic technology allowing surgeons to perform minimally invasive procedures, is available in most of the hospitals now. These systems enable a degree of precision and accuracy far greater than the procedures done manually. The less invasive the surgery, the less trauma it will occur and less blood loss, less anxiety of the patients.

Improved radiology

The first computed tomography scanners were introduced in 1971. The first magnetic resonance imaging (MRI) scan of the human body took place in 1977. By the early 2000s, cardiac MRI, body MRI, and fetal imaging, became routine. The search continues for new algorithms to detect specific diseases as well as to analyze the results of scans [ 9 ]. All those are the contribution of the technology of AI.

Virtual presence

The virtual presence technology can enable a distant diagnosis of the diseases. The patient does not have to leave his/her bed but using a remote presence robot, doctors can check the patients without actually being there. Health professionals can move around and interact almost as effectively as if they were present. This allows specialists to assist patients who are unable to travel.

S OME CAUTIONS TO BE REMINDED

Despite all the positive promises that AI provides, human experts, however, are still essential and necessary to design, program, and operate the AI from any unpredictable error from occurring. Beth Kindig, a San Francisco-based technology analyst with more than a decade of experience in analyzing private and public technology companies, published a free newsletter indicating that although AI has a potential promise for better medical diagnosis, human experts are still needed to avoid the misclassification of unknown diseases because AI is not omnipotent to solve all problems for human kinds. There are times when AI meets an impasse, and to carry on its mission, it may just proceed indiscriminately, ending in creating more problems. Thus vigilant watch of AI's function cannot be neglected. This reminder is known as physician-in-the-loop [ 13 ].

The question of an ethical AI consequently was brought up by Elizabeth Gibney in her article published in Nature to caution any bias and possible societal harm [ 14 ]. The Neural Information processing Systems (NeurIPS) conference in Vancouver Canada in 2020 brought up the ethical controversies of the application of AI technology, such as in predictive policing or facial recognition, that due to bias algorithms can result in hurting the vulnerable population [ 14 ]. For instance, the NeurIPS can be programmed to target certain race or decree as the probable suspect of crime or trouble makers.

T HE CHALLENGE OF ARTIFICIAL INTELLIGENCE TO BIOETHICS

Artificial intelligence ethics must be developed.

Bioethics is a discipline that focuses on the relationship among living beings. Bioethics accentuates the good and the right in biospheres and can be categorized into at least three areas, the bioethics in health settings that is the relationship between physicians and patients, the bioethics in social settings that is the relationship among humankind and the bioethics in environmental settings that is the relationship between man and nature including animal ethics, land ethics, ecological ethics…etc. All these are concerned about relationships within and among natural existences.

As AI arises, human has a new challenge in terms of establishing a relationship toward something that is not natural in its own right. Bioethics normally discusses the relationship within natural existences, either humankind or his environment, that are parts of natural phenomena. But now men have to deal with something that is human-made, artificial and unnatural, namely AI. Human has created many things yet never has human had to think of how to ethically relate to his own creation. AI by itself is without feeling or personality. AI engineers have realized the importance of giving the AI ability to discern so that it will avoid any deviated activities causing unintended harm. From this perspective, we understand that AI can have a negative impact on humans and society; thus, a bioethics of AI becomes important to make sure that AI will not take off on its own by deviating from its originally designated purpose.

Stephen Hawking warned early in 2014 that the development of full AI could spell the end of the human race. He said that once humans develop AI, it may take off on its own and redesign itself at an ever-increasing rate [ 15 ]. Humans, who are limited by slow biological evolution, could not compete and would be superseded. In his book Superintelligence, Nick Bostrom gives an argument that AI will pose a threat to humankind. He argues that sufficiently intelligent AI can exhibit convergent behavior such as acquiring resources or protecting itself from being shut down, and it might harm humanity [ 16 ].

The question is–do we have to think of bioethics for the human's own created product that bears no bio-vitality? Can a machine have a mind, consciousness, and mental state in exactly the same sense that human beings do? Can a machine be sentient and thus deserve certain rights? Can a machine intentionally cause harm? Regulations must be contemplated as a bioethical mandate for AI production.

Studies have shown that AI can reflect the very prejudices humans have tried to overcome. As AI becomes “truly ubiquitous,” it has a tremendous potential to positively impact all manner of life, from industry to employment to health care and even security. Addressing the risks associated with the technology, Janosch Delcker, Politico Europe's AI correspondent, said: “I don't think AI will ever be free of bias, at least not as long as we stick to machine learning as we know it today,”…. “What's crucially important, I believe, is to recognize that those biases exist and that policymakers try to mitigate them” [ 17 ]. The High-Level Expert Group on AI of the European Union presented Ethics Guidelines for Trustworthy AI in 2019 that suggested AI systems must be accountable, explainable, and unbiased. Three emphases are given:

• Lawful-respecting all applicable laws and regulations
• Ethical-respecting ethical principles and values
• Robust-being adaptive, reliable, fair, and trustworthy from a technical perspective while taking into account its social environment [ 18 ].

Seven requirements are recommended [ 18 ]:

• AI should not trample on human autonomy. People should not be manipulated or coerced by AI systems, and humans should be able to intervene or oversee every decision that the software makes
• AI should be secure and accurate. It should not be easily compromised by external attacks, and it should be reasonably reliable
• Personal data collected by AI systems should be secure and private. It should not be accessible to just anyone, and it should not be easily stolen
• Data and algorithms used to create an AI system should be accessible, and the decisions made by the software should be “understood and traced by human beings.” In other words, operators should be able to explain the decisions their AI systems make
• Services provided by AI should be available to all, regardless of age, gender, race, or other characteristics. Similarly, systems should not be biased along these lines
• AI systems should be sustainable (i.e., they should be ecologically responsible) and “enhance positive social change”
• AI systems should be auditable and covered by existing protections for corporate whistleblowers. The negative impacts of systems should be acknowledged and reported in advance.

From these guidelines, we can suggest that future AI must be equipped with human sensibility or “AI humanities.” To accomplish this, AI researchers, manufacturers, and all industries must bear in mind that technology is to serve not to manipulate humans and his society. Bostrom and Judkowsky listed responsibility, transparency, auditability, incorruptibility, and predictability [ 19 ] as criteria for the computerized society to think about.

S UGGESTED PRINCIPLES FOR ARTIFICIAL INTELLIGENCE BIOETHICS

Nathan Strout, a reporter at Space and Intelligence System at Easter University, USA, reported just recently that the intelligence community is developing its own AI ethics. The Pentagon made announced in February 2020 that it is in the process of adopting principles for using AI as the guidelines for the department to follow while developing new AI tools and AI-enabled technologies. Ben Huebner, chief of the Office of Director of National Intelligence's Civil Liberties, Privacy, and Transparency Office, said that “We're going to need to ensure that we have transparency and accountability in these structures as we use them. They have to be secure and resilient” [ 20 ]. Two themes have been suggested for the AI community to think more about: Explainability and interpretability. Explainability is the concept of understanding how the analytic works, while interpretability is being able to understand a particular result produced by an analytic [ 20 ].

All the principles suggested by scholars for AI bioethics are well-brought-up. I gather from different bioethical principles in all the related fields of bioethics to suggest four principles here for consideration to guide the future development of the AI technology. We however must bear in mind that the main attention should still be placed on human because AI after all has been designed and manufactured by human. AI proceeds to its work according to its algorithm. AI itself cannot empathize nor have the ability to discern good from evil and may commit mistakes in processes. All the ethical quality of AI depends on the human designers; therefore, it is an AI bioethics and at the same time, a trans-bioethics that abridge human and material worlds. Here are the principles:

• Beneficence: Beneficence means doing good, and here it refers to the purpose and functions of AI should benefit the whole human life, society and universe. Any AI that will perform any destructive work on bio-universe, including all life forms, must be avoided and forbidden. The AI scientists must understand that reason of developing this technology has no other purpose but to benefit human society as a whole not for any individual personal gain. It should be altruistic, not egocentric in nature
• Value-upholding: This refers to AI's congruence to social values, in other words, universal values that govern the order of the natural world must be observed. AI cannot elevate to the height above social and moral norms and must be bias-free. The scientific and technological developments must be for the enhancement of human well-being that is the chief value AI must hold dearly as it progresses further
• Lucidity: AI must be transparent without hiding any secret agenda. It has to be easily comprehensible, detectable, incorruptible, and perceivable. AI technology should be made available for public auditing, testing and review, and subject to accountability standards … In high-stakes settings like diagnosing cancer from radiologic images, an algorithm that can't “explain its work” may pose an unacceptable risk. Thus, explainability and interpretability are absolutely required
• Accountability: AI designers and developers must bear in mind they carry a heavy responsibility on their shoulders of the outcome and impact of AI on whole human society and the universe. They must be accountable for whatever they manufacture and create.

C ONCLUSION

AI is here to stay in our world and we must try to enforce the AI bioethics of beneficence, value upholding, lucidity and accountability. Since AI is without a soul as it is, its bioethics must be transcendental to bridge the shortcoming of AI's inability to empathize. AI is a reality of the world. We must take note of what Joseph Weizenbaum, a pioneer of AI, said that we must not let computers make important decisions for us because AI as a machine will never possess human qualities such as compassion and wisdom to morally discern and judge [ 10 ]. Bioethics is not a matter of calculation but a process of conscientization. Although AI designers can up-load all information, data, and programmed to AI to function as a human being, it is still a machine and a tool. AI will always remain as AI without having authentic human feelings and the capacity to commiserate. Therefore, AI technology must be progressed with extreme caution. As Von der Leyen said in White Paper on AI – A European approach to excellence and trust : “AI must serve people, and therefore, AI must always comply with people's rights…. High-risk AI. That potentially interferes with people's rights has to be tested and certified before it reaches our single market” [ 21 ].

Financial support and sponsorship

Conflicts of interest.

There are no conflicts of interest.

R EFERENCES

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Using artificial intelligence to advance the research and development of orphan drugs.

1. Introduction

2. overview of orphan drug development and of the applications of ai in medical research, 2.1. definitions, 2.2. regulation ec 141/2000: a turning point for orphan drug development within the eu regulatory framework, 2.3. artificial intelligence in medical research, 3. decreasing the barriers of complexity and financial risk: how ai systems can facilitate the development of a molecule, 3.1. using ai to understand the etiology of monogenic and complex diseases and drug repurposing, 3.2. using ai to design molecules from scratch, 3.3. can the barriers of complexity and financial risk really be decreased by ai, 4. decreasing the barriers of low trialability and complexity: how ai can facilitate clinical trials, 4.1. using ai to recruit patients, 4.2. using ai to ensure a smooth run of the clinical trial, 5. conclusions, author contributions, institutional review board statement, informed consent statement, data availability statement, acknowledgments, conflicts of interest.

• Nguengang Wakap, S.; Lambert, D.M.; Olry, A.; Rodwell, C.; Gueydan, C.; Lanneau, V.; Murphy, D.; Le Cam, Y.; Rath, A. Estimating Cumulative Point Prevalence of Rare Diseases: Analysis of the Orphanet Database. Eur. J. Hum. Genet. 2020 , 28 , 165–173. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• DGOS. Les Maladies Rares ; Ministère de la Santé et de la Prévention: Paris, France, 2023. [ Google Scholar ]
• Ferreira, C.R. The Burden of Rare Diseases. Am. J. Med. Genet. Part A 2019 , 179 , 885–892. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• National Organization for Rare Disorders. Orphan Drugs in the United States: An Examination of Patents and Orphan Drug Exclusivity ; National Organization for Rare Disorders: Danbury, CT, USA, 2021. [ Google Scholar ]
• European Medicine Agency. Annual Report on the Use of the Special Contribution for Orphan Medicinal Products ; EMA: London, UK, 2023.
• Moors, E.H.M.; Faber, J. Orphan Drugs: Unmet Societal Need for Non-Profitable Privately Supplied New Products ; Elsevier: Amsterdam, The Netherlands, 2007; Volume 19. [ Google Scholar ]
• Althobaiti, H.; Seoane-Vazquez, E.; Brown, L.M.; Fleming, M.L.; Rodriguez-Monguio, R. Disentangling the Cost of Orphan Drugs Marketed in the United States. Healthcare 2023 , 11 , 558. [ Google Scholar ] [ CrossRef ]
• Nuijten, M.; Capri, S. Valuation of Medical Innovation Handling with Uncertainty and Risk. J. Mark. Access Health Policy 2024 , 12 , 199–208. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Douglas, C.M.W.; Aith, F.; Boon, W.; de Neiva Borba, M.; Doganova, L.; Grunebaum, S.; Hagendijk, R.; Lynd, L.; Mallard, A.; Mohamed, F.A.; et al. Social Pharmaceutical Innovation and Alternative Forms of Research, Development and Deployment for Drugs for Rare Diseases. Orphanet J. Rare Dis. 2022 , 17 , 344. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Zimmermann, B.M.; Eichinger, J.; Baumgartner, M.R. A Systematic Review of Moral Reasons on Orphan Drug Reimbursement. Orphanet J. Rare Dis. 2021 , 16 , 292. [ Google Scholar ] [ CrossRef ]
• Belousova, O.A.; Groen, A.J.; Ouendag, A.M. Opportunities and Barriers for Innovation and Entrepreneurship in Orphan Drug Development. Technol. Forecast. Soc. Change 2020 , 161 , 120333. [ Google Scholar ] [ CrossRef ]
• Wang, F.; Preininger, A. AI in Health: State of the Art, Challenges, and Future Directions. Yearb. Med. Inform. 2019 , 28 , 16–26. [ Google Scholar ] [ CrossRef ]
• Yao, X.; Rushlow, D.R.; Inselman, J.W.; McCoy, R.G.; Thacher, T.D.; Behnken, E.M.; Bernard, M.E.; Rosas, S.L.; Akfaly, A.; Misra, A.; et al. Artificial Intelligence–Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial. Nat. Med. 2021 , 27 , 815–819. [ Google Scholar ] [ CrossRef ]
• European Parliament. Regulation 141/2000 Regulation (EU) No 141/2000 of the European Parliament and of the Council of the 16 December 1999 on Orphan Medicinal Products ; European Parliament: Strasbourg, France, 1999.
• Bell, S.A.; Tudur Smith, C. A Comparison of Interventional Clinical Trials in Rare versus Non-Rare Diseases: An Analysis of ClinicalTrials.Gov. Orphanet J. Rare Dis. 2014 , 9 , 170. [ Google Scholar ] [ CrossRef ]
• Wästfelt, M.; Fadeel, B.; Henter, J.-I. A Journey of Hope: Lessons Learned from Studies on Rare Diseases and Orphan Drugs. J. Intern. Med. 2006 , 260 , 1–10. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Aymé, S. État de La R&D Dans Les Maladies Rares, 10 Ans Après Le Règlement Européen Du Médicament Orphelin. Presse Méd. 2010 , 39 , 5–7. [ Google Scholar ] [ CrossRef ]
• Sprecher, F. Orphan Drugs and Paediatric Medicinal Products. Pharm. Policy Law 2013 , 15 , 173–194. [ Google Scholar ] [ CrossRef ]
• Maresova, P.; Klimova, B.; Kuca, K. Financial and Legislative Aspects of Drug Development of Orphan Diseases on the European Market—A Systematic Review. Appl. Econ. 2016 , 48 , 2562–2570. [ Google Scholar ] [ CrossRef ]
• Roche. What Is a Clinical Trial and How Does It Work? Available online: https://www.roche.com/innovation/process/clinical-trials/about (accessed on 13 March 2023).
• Microsoft Azure. What Is Machine Learning? Available online: https://azure.microsoft.com/en-us/resources/cloud-computing-dictionary/what-is-machine-learning-platform (accessed on 13 March 2023).
• Maresova, P.; Klimova, B.; Kuca, K. Legislation, Regulation and Policies Issues of Orphan Drugs in Developed Countries from 2010 to 2016. J. Appl. Biomed. 2018 , 16 , 175–179. [ Google Scholar ] [ CrossRef ]
• Rodwell, C.; Aymé, S. Rare Disease Policies to Improve Care for Patients in Europe. Biochim. Biophys. Acta (BBA) Mol. Basis Dis. 2015 , 1852 , 2329–2335. [ Google Scholar ] [ CrossRef ]
• Regulation 726/2004 Regulation (EU) No 726/2004 of the European Parliament and of the Council Laying Down Community Pro-Cedures for the Authorisation and Supervision of Medicinal Products for Human and Veterinary Use and Establishing a European Medicines Agency ; European Parliament: Strasbourg, France, 2004.
• Sanofi. Sanofi Invests $ 180 Million Equity in Owkin’s Artificial Intelligence and Federated Learning to Advance Oncology Pipeline ; Sanofi: Paris, France, 2021. [ Google Scholar ]
• Atomwise. Atomwise Announces Multi-Year, Multi-Target Agreement with Lilly for AI- Driven Drug Discovery Program ; Atomwise: San Francisco, CA, USA, 2019. [ Google Scholar ]
• Gulshan, V.; Peng, L.; Coram, M.; Stumpe, M.C.; Wu, D.; Narayanaswamy, A.; Venugopalan, S.; Widner, K.; Madams, T.; Cuadros, J.; et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. J. Am. Med. Assoc. 2016 , 316 , 2402–2410. [ Google Scholar ] [ CrossRef ]
• Shu, C. Regard Serves as a “Medical Co-Pilot” for Busy Physicians ; TechCrunch: San Francisco, CA, USA, 2022. [ Google Scholar ]
• Miozza, M.; Brunetta, F.; Appio, F.P. Digital Transformation of the Pharmaceutical Industry: A Future Research Agenda for Management Studies. Technol. Forecast. Soc. Change 2024 , 207 , 123580. [ Google Scholar ] [ CrossRef ]
• Iansiti, M.; Lakhani, K.R.; Mayer, H.; Herman, K. Moderna (A) ; Harvard Business School: Boston, MA, USA, 2020; Volume 9–621–032. [ Google Scholar ]
• Moderna COVID-19 Vaccines ; FDA: Silver Spring, MD, USA, 2023.
• Chen, Y.; Yang, O.; Sampat, C.; Bhalode, P.; Ramachandran, R.; Ierapetritou, M. Digital Twins in Pharmaceutical and Biopharmaceutical Manufacturing: A Literature Review. Processes 2020 , 8 , 1088. [ Google Scholar ] [ CrossRef ]
• Coito, T.; Martins, M.S.E.; Firme, B.; Figueiredo, J.; Vieira, S.M.; Sousa, J.M.C. Assessing the Impact of Automation in Pharmaceutical Quality Control Labs Using a Digital Twin. J. Manuf. Syst. 2022 , 62 , 270–285. [ Google Scholar ] [ CrossRef ]
• Garza Ramos, A.; Daim, T.; Gaats, L.; Hutmacher, D.W.; Hackenberger, D. Technology Roadmap for the Development of a 3D Cell Culture Workstation for a Biomedical Industry Startup. Technol. Forecast. Soc. Change 2022 , 174 , 121213. [ Google Scholar ] [ CrossRef ]
• Wu, Q.; He, Q. DIY Laboratories and Business Innovation Ecosystems: The Case of Pharmaceutical Industry. Technol. Forecast. Soc. Change 2020 , 161 , 120336. [ Google Scholar ] [ CrossRef ]
• Lee, Y.; Krishnan, A.; Oughtred, R.; Rust, J.; Chang, C.S.; Ryu, J.; Kristensen, V.N.; Dolinski, K.; Theesfeld, C.L.; Troyanskaya, O.G. A Computational Framework for Genome-Wide Characterization of the Human Disease Landscape. Cell Syst. 2019 , 8 , 152–162. [ Google Scholar ] [ CrossRef ]
• BROAD Institute. Connectivity Map (CMAP) ; Broad Institute: Cambridge, MA, USA, 2015. [ Google Scholar ]
• Mears, A.J.; Schock, S.C.; Hadwen, J.; Putos, S.; Dyment, D.; Boycott, K.M.; MacKenzie, A. Mining the Transcriptome for Rare Disease Therapies: A Comparison of the Efficiencies of Two Data Mining Approaches and a Targeted Cell-Based Drug Screen. Npj Genom. Med. 2017 , 2 , 1–8. [ Google Scholar ] [ CrossRef ]
• National Human Genome Research Institute. Complex Disease ; Genome.Gov: Bethesda, MD, USA, 2022.
• Deep Genomics Therapeutic Approach. Available online: https://www.deepgenomics.com/ (accessed on 2 September 2024).
• Owkin. Discover—Advancing Drug Discovery ; OWKIN: New York, NY, USA.
• BioXcel AI Platform—BioXcel Therapeutics Inc. Available online: https://www.bioxceltherapeutics.com/ (accessed on 2 September 2024).
• The Lancet. Neurology Rare Diseases: Maintaining Momentum. Lancet Neurol. 2022 , 21 , 203. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Sexton, J. Repurposing Generic Drugs Can Reduce Time and Cost to Develop New Treatments—But Low Profitability Remains a Barrier ; The Conversation: Melbourne, Australia, 2022. [ Google Scholar ]
• Iansiti, M.; Lakhani, K.R. Competing in the Age of AI. Harv. Bus. Rev. 2020 , 98 , 60–67. [ Google Scholar ]
• Broom, D. 5 Charts That Tell the Story of Vaccines Today ; World Economic Forum: Cologny, Switzerland, 2020. [ Google Scholar ]
• Merk, D.; Friedrich, L.; Grisoni, F.; Schneider, G. De Novo Design of Bioactive Small Molecules by Artificial Intelligence. Mol. Inform. 2018 , 37 , 1700153. [ Google Scholar ] [ CrossRef ]
• Reeves, M.; Levin, S.; Fink, T.; Levina, A. Taming Complexity. Making Sure the Benefits of Any Addition to an Organizaiton’s Systems Outweight Its Costs. Harv. Bus. Rev. 2020 , 113 , 1–11. [ Google Scholar ]
• Jayasundara, K.; Hollis, A.; Krahn, M.; Mamdani, M.; Hoch, J.S.; Grootendorst, P. Estimating the Clinical Cost of Drug Development for Orphan versus Non-Orphan Drugs. Orphanet J. Rare Dis. 2019 , 14 , 12. [ Google Scholar ] [ CrossRef ]
• Harrer, S.; Shah, P.; Antony, B.; Hu, J. Artificial Intelligence for Clinical Trial Design. Trends Pharmacol. Sci. 2019 , 40 , 577–591. [ Google Scholar ] [ CrossRef ]
• Ismail, A.; Al-Zoubi, T.; El Naqa, I.; Saeed, H. The Role of Artificial Intelligence in Hastening Time to Recruitment in Clinical Trials. BJR Open 2023 , 5 , 20220023. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Geva, A.; Gronsbell, J.L.; Cai, T.; Cai, T.; Murphy, S.N.; Lyons, J.C.; Heinz, M.M.; Natter, M.D.; Patibandla, N.; Bickel, J.; et al. A Computable Phenotype Improves Cohort Ascertainment in a Pediatric Pulmonary Hypertension Registry. J. Pediatr. 2017 , 188 , 224–231. [ Google Scholar ] [ CrossRef ]
• Leventakos, K.; Helgeson, J.; Mansfield, A.; Deering, E.; Schwecke, A.; Adjei, A.; Molina, J.; Hocum, C.; Parikh, K.; Halfdanarson, T.; et al. P1.16-14 Effects of an Artificial Intelligence (AI) System on Clinical Trial Enrollment in Lung Cancer. J. Thorac. Oncol. 2019 , 14 , S592. [ Google Scholar ] [ CrossRef ]
• IQVIA. Patient Recruitment. Available online: https://www.iqvia.com/solutions/research-and-development/clinical-trials/patient-recruitment (accessed on 15 June 2023).
• AiCure. Patient Connect—Remote Patient Monitoring Solution. AiCure. 2023. Available online: https://aicure.com/solutions/patient-connect (accessed on 1 January 2024).
• Labovitz, D.L.; Shafner, L.; Reyes Gil, M.; Virmani, D.; Hanina, A. Using Artificial Intelligence to Reduce the Risk of Nonadherence in Patients on Anticoagulation Therapy. Stroke 2017 , 48 , 1416–1419. [ Google Scholar ] [ CrossRef ]
• Regulation 2016/679 Regulation (EU) No 2016/679 of the European Parliament and of the Council of 27 April 2016 on the Protection of Natural Persons with Regard to the Processing of Personal Data and on the Free Movement of Such Data, and Repealing Directive 95/46/EC (General Data Protection Regulation) ; European Parliament: Strasbourg, France, 2016.
• Murdoch, B. Privacy and Artificial Intelligence: Challenges for Protecting Health Information in a New Era. BMC Med. Ethics 2021 , 22 , 122. [ Google Scholar ] [ CrossRef ]
• McGrath, R.G.; Nerkar, A. Real Options Reasoning and a New Look at the R&d Investment Strategies of Pharmaceutical Firms. Strateg. Manag. J. 2004 , 25 , 1–21. [ Google Scholar ] [ CrossRef ]
• Burger-Helmchen, T. Option Chain and Change Management: A Structural Equation Application. Eur. Manag. J. 2009 , 27 , 176–186. [ Google Scholar ] [ CrossRef ]
• Napolitano, G.; Has, C.; Schwerk, A.; Yuan, J.-H.; Ullrich, C. Potential of Artificial Intelligence to Accelerate Drug Development for Rare Diseases. Pharm. Med. 2024 , 38 , 79–86. [ Google Scholar ] [ CrossRef ]
• Heraud, J.-A.; Kerr, F.; Burger-Helmchen, T. Creative Management of Complex Systems ; Wiley-ISTE: Hoboken, NJ, USA, 2019; ISBN 978-1-84821-957-1. [ Google Scholar ]
• Neukam, M.; Bollinger, S. Encouraging Creative Teams to Integrate a Sustainable Approach to Technology. J. Bus. Res. 2022 , 150 , 354–364. [ Google Scholar ] [ CrossRef ]
• Stefani, U.; Schiavone, F.; Laperche, B.; Burger-Helmchen, T. New Tools and Practices for Financing Novelty: A Research Agenda. Eur. J. Innov. Manag. 2019 , 23 , 314–328. [ Google Scholar ] [ CrossRef ]
• Sinha, M.S.; Stern, A.D.; Rai, A.K. Four Decades of Orphan Drugs and Priorities for the Future. N. Engl. J. Med. 2024 , 391 , 100–102. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Costa, E.; Isgrò, A.; de Montalembert, M.; Leufkens, H.G.M.; Ware, R.E.; De Franceschi, L. Successes and Pitfalls in Orphan Drug Development for Sickle Cell Disease. Blood Adv. 2024 , 8 , 2455–2465. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Agyralides, G. The Future of Medicine: An Outline Attempt Using State-of-the-Art Business and Scientific Trends. Front. Med. 2024 , 11 , 1391727. [ Google Scholar ] [ CrossRef ] [ PubMed ]

Click here to enlarge figure

Share and Cite

Irissarry, C.; Burger-Helmchen, T. Using Artificial Intelligence to Advance the Research and Development of Orphan Drugs. Businesses 2024 , 4 , 453-472. https://doi.org/10.3390/businesses4030028

Irissarry C, Burger-Helmchen T. Using Artificial Intelligence to Advance the Research and Development of Orphan Drugs. Businesses . 2024; 4(3):453-472. https://doi.org/10.3390/businesses4030028

Irissarry, Carla, and Thierry Burger-Helmchen. 2024. "Using Artificial Intelligence to Advance the Research and Development of Orphan Drugs" Businesses 4, no. 3: 453-472. https://doi.org/10.3390/businesses4030028

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

Artificial intelligence and investor behaviour

• Share to Twitter
• Share to Facebook
• Share to LinkedIn
• Share to Reddit
• Share via Email

Canadians are using artificial intelligence (AI) to access financial information, advice and recommendations. Find out more about an Ontario Securities Commission Ontario Securities Commission An independent Crown corporation that is responsible for regulating the capital markets in Ontario. Its… + read full definition ( OSC OSC See Ontario Securities Commission. + read full definition ) study looking at retail investor decision-making and the risks and benefits of advice from AI systems.

On this page you’ll find

Why study ai and retail investor decision-making, how was the experiment conducted and what did it find, how is ai currently being used in retail investing, what are the benefits and risks of using ai in retail investing.

Many Canadians are starting to use artificial intelligence (AI) tools in retail investing. Using AI presents a range of both new opportunities and potential risks for investors. As the use of AI increases, regulators around the world, including the OSC, are examining the risks AI could pose to investors while continuing to support innovation.

Improving the investor experience is a key priority for the OSC. That’s why the OSC conducted research to understand how investors react when they think advice is being provided by AI. A behavioural science experiment was conducted to explore the impact of AI on investor decision-making. And researchers identified common uses of AI that impact investors.

Read the full report: Artificial Intelligence and Retail Investing: Use Cases and Experimental Research

The experiment looked at how the source of an investment Investment An item of value you buy to get income or to grow in value. + read full definition suggestion — AI, human, or a blend of the two — impacts whether investors follow that suggestion.

Participants were given a hypothetical $20,000 to invest Invest To use money for the purpose of making more money by making an investment. Often… + read full definition in an online simulation. Participants who were not in the control condition were introduced to WealthTogether — a fictitious financial services firm, to help them decide how to allocate their funds. They were then introduced to one of three WealthTogether financial services providers:

• A person named Alex.
• A person named Alex who is using an AI tool to inform their suggestions.
• An AI tool named Kai.

After participants received the suggestion, they allocated the full $20,000 across any combination of equities Equities Another word for investments in the stock market. + read full definition , fixed income Fixed income An investment that pays regular income to you. Examples: Guaranteed Investment Certificates, Canada Savings Bonds… + read full definition , and cash. They did not have to follow the suggestion they were given, but were free to invest the money any way they chose.

OSC researchers then measured how closely participants followed the investment suggestion. People who received the investment suggestion from a human using an AI tool (‘blended’) adhered to the investment suggestion most closely, although this difference was not significant.

There was also no significant difference in adherence to investment suggestions provided by a human or an AI tool. This indicates Canadian investors are receptive to taking advice from an AI system. Canadians may have an explicit or implicit view that the benefits of either human or AI investment advice can be maximized by combining the two.

This research underlines the need to ensure that AI systems investors use for financial information or advice are based on unbiased, high-quality data, and prioritize the best interests of investors rather than the firms who develop them.

OSC researchers examined current investor-facing uses of AI in Canada and abroad through a literature review and environmental scan. Three common uses were identified:

• Decision support: Involves AI systems that provide recommendations or advice to guide investment decisions.
• Automation: Consists of AI systems that automate portfolio Portfolio All the different investments that an individual or organization holds. May include stocks, bonds and… + read full definition and/or fund (e.g., ETF) management.
• Scams and fraud: Includes AI systems that either facilitate or mitigate scams targeting retail investors, as well as scams capitalizing on the “buzz” of AI.

Several key benefits and risks were identified in the research.

The benefits of using AI in retail investing include:

• Reduced cost : AI systems can reduce the cost of personalized advice and portfolio management. This can create considerable value for retail investors.
• Access to advice : More sophisticated and properly regulated AI systems can provide increased access to financial advice for retail investors. This is particularly important for people who cannot access advice through traditional channels.
• Improved decision-making : AI tools can be developed to guide investor decision-making around key areas such as portfolio diversification Diversification A way of spreading investment risk by by choosing a mix of investments. The idea… + read full definition and risk management, as well as tools to assist investors in identifying financial scams.
• Enhanced performance : Existing research has shown that AI systems can make more accurate predictions of earnings Earnings For companies, it’s the money they make and share with their shareholders. For investors, it’s… + read full definition changes and generate more profitable trading strategies compared to human analysts.

The risks of using AI in retail investing include:

• Bias: AI models are generally subject to the biases and assumptions of the humans who develop them. For example, an AI system developed by people employed by a specific company may be biased towards promoting that company’s products, even when buying those products is not in the investor’s best interests.
• Herding : The concentration of AI tools among a few providers may induce herding behaviour, convergence of investment strategies, and chain reactions that exacerbate volatility Volatility The rate at which the price of a security increases or decreases for a given… + read full definition during market shocks.
• Data quality : If an AI model is built on poor data quality, then the outputs, whether advice, recommendations, or otherwise, will be of poor quality as well.
• Governance and ethics : The ‘black box’ nature of AI systems and limitations around data privacy and transparency create concerns around clear accountability in cases where AI systems produce adverse outcomes for investors.

Artificial Intelligence and Retail Investing: Use Cases and Experimental Research builds on the OSC’s existing research in the area of artificial intelligence. It also reinforces the benefit Benefit Money, goods, or services that you get from your workplace or from a government program… + read full definition of using behavioural science as a policy tool by regulators. As AI continues to advance in capabilities, more research is needed to help capital markets Capital markets Where people buy and sell investments. + read full definition stakeholders better understand the implications for retail investors.

The OSC partnered with the consulting firm Behavioural Insights Team (BIT) to conduct this research. The experiment included 7,771 Canadian residents aged 18 years and older. Current investors made up 60% of the sample, with 40% of participants being non-investors.

Sign up for Investor News

Join 18,000+ subscribers and stay informed with timely articles, the latest investor warnings and financial literacy resources like videos, calculators and quizzes.

• August 15, 2024
• June 25, 2024
• June 11, 2024

Peer Reviewed

GPT-fabricated scientific papers on Google Scholar: Key features, spread, and implications for preempting evidence manipulation

Article metrics.

CrossRef Citations

Altmetric Score

PDF Downloads

Academic journals, archives, and repositories are seeing an increasing number of questionable research papers clearly produced using generative AI. They are often created with widely available, general-purpose AI applications, most likely ChatGPT, and mimic scientific writing. Google Scholar easily locates and lists these questionable papers alongside reputable, quality-controlled research. Our analysis of a selection of questionable GPT-fabricated scientific papers found in Google Scholar shows that many are about applied, often controversial topics susceptible to disinformation: the environment, health, and computing. The resulting enhanced potential for malicious manipulation of society’s evidence base, particularly in politically divisive domains, is a growing concern.

Swedish School of Library and Information Science, University of Borås, Sweden

Department of Arts and Cultural Sciences, Lund University, Sweden

Division of Environmental Communication, Swedish University of Agricultural Sciences, Sweden

Research Questions

• Where are questionable publications produced with generative pre-trained transformers (GPTs) that can be found via Google Scholar published or deposited?
• What are the main characteristics of these publications in relation to predominant subject categories?
• How are these publications spread in the research infrastructure for scholarly communication?
• How is the role of the scholarly communication infrastructure challenged in maintaining public trust in science and evidence through inappropriate use of generative AI?

research note Summary

• A sample of scientific papers with signs of GPT-use found on Google Scholar was retrieved, downloaded, and analyzed using a combination of qualitative coding and descriptive statistics. All papers contained at least one of two common phrases returned by conversational agents that use large language models (LLM) like OpenAI’s ChatGPT. Google Search was then used to determine the extent to which copies of questionable, GPT-fabricated papers were available in various repositories, archives, citation databases, and social media platforms.
• Roughly two-thirds of the retrieved papers were found to have been produced, at least in part, through undisclosed, potentially deceptive use of GPT. The majority (57%) of these questionable papers dealt with policy-relevant subjects (i.e., environment, health, computing), susceptible to influence operations. Most were available in several copies on different domains (e.g., social media, archives, and repositories).
• Two main risks arise from the increasingly common use of GPT to (mass-)produce fake, scientific publications. First, the abundance of fabricated “studies” seeping into all areas of the research infrastructure threatens to overwhelm the scholarly communication system and jeopardize the integrity of the scientific record. A second risk lies in the increased possibility that convincingly scientific-looking content was in fact deceitfully created with AI tools and is also optimized to be retrieved by publicly available academic search engines, particularly Google Scholar. However small, this possibility and awareness of it risks undermining the basis for trust in scientific knowledge and poses serious societal risks.

Implications

The use of ChatGPT to generate text for academic papers has raised concerns about research integrity. Discussion of this phenomenon is ongoing in editorials, commentaries, opinion pieces, and on social media (Bom, 2023; Stokel-Walker, 2024; Thorp, 2023). There are now several lists of papers suspected of GPT misuse, and new papers are constantly being added. 1 See for example Academ-AI, https://www.academ-ai.info/ , and Retraction Watch, https://retractionwatch.com/papers-and-peer-reviews-with-evidence-of-chatgpt-writing/ . While many legitimate uses of GPT for research and academic writing exist (Huang & Tan, 2023; Kitamura, 2023; Lund et al., 2023), its undeclared use—beyond proofreading—has potentially far-reaching implications for both science and society, but especially for their relationship. It, therefore, seems important to extend the discussion to one of the most accessible and well-known intermediaries between science, but also certain types of misinformation, and the public, namely Google Scholar, also in response to the legitimate concerns that the discussion of generative AI and misinformation needs to be more nuanced and empirically substantiated  (Simon et al., 2023).

Google Scholar, https://scholar.google.com , is an easy-to-use academic search engine. It is available for free, and its index is extensive (Gusenbauer & Haddaway, 2020). It is also often touted as a credible source for academic literature and even recommended in library guides, by media and information literacy initiatives, and fact checkers (Tripodi et al., 2023). However, Google Scholar lacks the transparency and adherence to standards that usually characterize citation databases. Instead, Google Scholar uses automated crawlers, like Google’s web search engine (Martín-Martín et al., 2021), and the inclusion criteria are based on primarily technical standards, allowing any individual author—with or without scientific affiliation—to upload papers to be indexed (Google Scholar Help, n.d.). It has been shown that Google Scholar is susceptible to manipulation through citation exploits (Antkare, 2020) and by providing access to fake scientific papers (Dadkhah et al., 2017). A large part of Google Scholar’s index consists of publications from established scientific journals or other forms of quality-controlled, scholarly literature. However, the index also contains a large amount of gray literature, including student papers, working papers, reports, preprint servers, and academic networking sites, as well as material from so-called “questionable” academic journals, including paper mills. The search interface does not offer the possibility to filter the results meaningfully by material type, publication status, or form of quality control, such as limiting the search to peer-reviewed material.

To understand the occurrence of ChatGPT (co-)authored work in Google Scholar’s index, we scraped it for publications, including one of two common ChatGPT responses (see Appendix A) that we encountered on social media and in media reports (DeGeurin, 2024). The results of our descriptive statistical analyses showed that around 62% did not declare the use of GPTs. Most of these GPT-fabricated papers were found in non-indexed journals and working papers, but some cases included research published in mainstream scientific journals and conference proceedings. 2 Indexed journals mean scholarly journals indexed by abstract and citation databases such as Scopus and Web of Science, where the indexation implies journals with high scientific quality. Non-indexed journals are journals that fall outside of this indexation. More than half (57%) of these GPT-fabricated papers concerned policy-relevant subject areas susceptible to influence operations. To avoid increasing the visibility of these publications, we abstained from referencing them in this research note. However, we have made the data available in the Harvard Dataverse repository.

The publications were related to three issue areas—health (14.5%), environment (19.5%) and computing (23%)—with key terms such “healthcare,” “COVID-19,” or “infection”for health-related papers, and “analysis,” “sustainable,” and “global” for environment-related papers. In several cases, the papers had titles that strung together general keywords and buzzwords, thus alluding to very broad and current research. These terms included “biology,” “telehealth,” “climate policy,” “diversity,” and “disrupting,” to name just a few.  While the study’s scope and design did not include a detailed analysis of which parts of the articles included fabricated text, our dataset did contain the surrounding sentences for each occurrence of the suspicious phrases that formed the basis for our search and subsequent selection. Based on that, we can say that the phrases occurred in most sections typically found in scientific publications, including the literature review, methods, conceptual and theoretical frameworks, background, motivation or societal relevance, and even discussion. This was confirmed during the joint coding, where we read and discussed all articles. It became clear that not just the text related to the telltale phrases was created by GPT, but that almost all articles in our sample of questionable articles likely contained traces of GPT-fabricated text everywhere.

Evidence hacking and backfiring effects

Generative pre-trained transformers (GPTs) can be used to produce texts that mimic scientific writing. These texts, when made available online—as we demonstrate—leak into the databases of academic search engines and other parts of the research infrastructure for scholarly communication. This development exacerbates problems that were already present with less sophisticated text generators (Antkare, 2020; Cabanac & Labbé, 2021). Yet, the public release of ChatGPT in 2022, together with the way Google Scholar works, has increased the likelihood of lay people (e.g., media, politicians, patients, students) coming across questionable (or even entirely GPT-fabricated) papers and other problematic research findings. Previous research has emphasized that the ability to determine the value and status of scientific publications for lay people is at stake when misleading articles are passed off as reputable (Haider & Åström, 2017) and that systematic literature reviews risk being compromised (Dadkhah et al., 2017). It has also been highlighted that Google Scholar, in particular, can be and has been exploited for manipulating the evidence base for politically charged issues and to fuel conspiracy narratives (Tripodi et al., 2023). Both concerns are likely to be magnified in the future, increasing the risk of what we suggest calling evidence hacking —the strategic and coordinated malicious manipulation of society’s evidence base.

The authority of quality-controlled research as evidence to support legislation, policy, politics, and other forms of decision-making is undermined by the presence of undeclared GPT-fabricated content in publications professing to be scientific. Due to the large number of archives, repositories, mirror sites, and shadow libraries to which they spread, there is a clear risk that GPT-fabricated, questionable papers will reach audiences even after a possible retraction. There are considerable technical difficulties involved in identifying and tracing computer-fabricated papers (Cabanac & Labbé, 2021; Dadkhah et al., 2023; Jones, 2024), not to mention preventing and curbing their spread and uptake.

However, as the rise of the so-called anti-vaxx movement during the COVID-19 pandemic and the ongoing obstruction and denial of climate change show, retracting erroneous publications often fuels conspiracies and increases the following of these movements rather than stopping them. To illustrate this mechanism, climate deniers frequently question established scientific consensus by pointing to other, supposedly scientific, studies that support their claims. Usually, these are poorly executed, not peer-reviewed, based on obsolete data, or even fraudulent (Dunlap & Brulle, 2020). A similar strategy is successful in the alternative epistemic world of the global anti-vaccination movement (Carrion, 2018) and the persistence of flawed and questionable publications in the scientific record already poses significant problems for health research, policy, and lawmakers, and thus for society as a whole (Littell et al., 2024). Considering that a person’s support for “doing your own research” is associated with increased mistrust in scientific institutions (Chinn & Hasell, 2023), it will be of utmost importance to anticipate and consider such backfiring effects already when designing a technical solution, when suggesting industry or legal regulation, and in the planning of educational measures.

Recommendations

Solutions should be based on simultaneous considerations of technical, educational, and regulatory approaches, as well as incentives, including social ones, across the entire research infrastructure. Paying attention to how these approaches and incentives relate to each other can help identify points and mechanisms for disruption. Recognizing fraudulent academic papers must happen alongside understanding how they reach their audiences and what reasons there might be for some of these papers successfully “sticking around.” A possible way to mitigate some of the risks associated with GPT-fabricated scholarly texts finding their way into academic search engine results would be to provide filtering options for facets such as indexed journals, gray literature, peer-review, and similar on the interface of publicly available academic search engines. Furthermore, evaluation tools for indexed journals 3 Such as LiU Journal CheckUp, https://ep.liu.se/JournalCheckup/default.aspx?lang=eng . could be integrated into the graphical user interfaces and the crawlers of these academic search engines. To enable accountability, it is important that the index (database) of such a search engine is populated according to criteria that are transparent, open to scrutiny, and appropriate to the workings of  science and other forms of academic research. Moreover, considering that Google Scholar has no real competitor, there is a strong case for establishing a freely accessible, non-specialized academic search engine that is not run for commercial reasons but for reasons of public interest. Such measures, together with educational initiatives aimed particularly at policymakers, science communicators, journalists, and other media workers, will be crucial to reducing the possibilities for and effects of malicious manipulation or evidence hacking. It is important not to present this as a technical problem that exists only because of AI text generators but to relate it to the wider concerns in which it is embedded. These range from a largely dysfunctional scholarly publishing system (Haider & Åström, 2017) and academia’s “publish or perish” paradigm to Google’s near-monopoly and ideological battles over the control of information and ultimately knowledge. Any intervention is likely to have systemic effects; these effects need to be considered and assessed in advance and, ideally, followed up on.

Our study focused on a selection of papers that were easily recognizable as fraudulent. We used this relatively small sample as a magnifying glass to examine, delineate, and understand a problem that goes beyond the scope of the sample itself, which however points towards larger concerns that require further investigation. The work of ongoing whistleblowing initiatives 4 Such as Academ-AI, https://www.academ-ai.info/ , and Retraction Watch, https://retractionwatch.com/papers-and-peer-reviews-with-evidence-of-chatgpt-writing/ . , recent media reports of journal closures (Subbaraman, 2024), or GPT-related changes in word use and writing style (Cabanac et al., 2021; Stokel-Walker, 2024) suggest that we only see the tip of the iceberg. There are already more sophisticated cases (Dadkhah et al., 2023) as well as cases involving fabricated images (Gu et al., 2022). Our analysis shows that questionable and potentially manipulative GPT-fabricated papers permeate the research infrastructure and are likely to become a widespread phenomenon. Our findings underline that the risk of fake scientific papers being used to maliciously manipulate evidence (see Dadkhah et al., 2017) must be taken seriously. Manipulation may involve undeclared automatic summaries of texts, inclusion in literature reviews, explicit scientific claims, or the concealment of errors in studies so that they are difficult to detect in peer review. However, the mere possibility of these things happening is a significant risk in its own right that can be strategically exploited and will have ramifications for trust in and perception of science. Society’s methods of evaluating sources and the foundations of media and information literacy are under threat and public trust in science is at risk of further erosion, with far-reaching consequences for society in dealing with information disorders. To address this multifaceted problem, we first need to understand why it exists and proliferates.

Finding 1: 139 GPT-fabricated, questionable papers were found and listed as regular results on the Google Scholar results page. Non-indexed journals dominate.

Most questionable papers we found were in non-indexed journals or were working papers, but we did also find some in established journals, publications, conferences, and repositories. We found a total of 139 papers with a suspected deceptive use of ChatGPT or similar LLM applications (see Table 1). Out of these, 19 were in indexed journals, 89 were in non-indexed journals, 19 were student papers found in university databases, and 12 were working papers (mostly in preprint databases). Table 1 divides these papers into categories. Health and environment papers made up around 34% (47) of the sample. Of these, 66% were present in non-indexed journals.

Finding 2: GPT-fabricated, questionable papers are disseminated online, permeating the research infrastructure for scholarly communication, often in multiple copies. Applied topics with practical implications dominate.

The 20 papers concerning health-related issues are distributed across 20 unique domains, accounting for 46 URLs. The 27 papers dealing with environmental issues can be found across 26 unique domains, accounting for 56 URLs.  Most of the identified papers exist in multiple copies and have already spread to several archives, repositories, and social media. It would be difficult, or impossible, to remove them from the scientific record.

As apparent from Table 2, GPT-fabricated, questionable papers are seeping into most parts of the online research infrastructure for scholarly communication. Platforms on which identified papers have appeared include ResearchGate, ORCiD, Journal of Population Therapeutics and Clinical Pharmacology (JPTCP), Easychair, Frontiers, the Institute of Electrical and Electronics Engineer (IEEE), and X/Twitter. Thus, even if they are retracted from their original source, it will prove very difficult to track, remove, or even just mark them up on other platforms. Moreover, unless regulated, Google Scholar will enable their continued and most likely unlabeled discoverability.

A word rain visualization (Centre for Digital Humanities Uppsala, 2023), which combines word prominences through TF-IDF 5 Term frequency–inverse document frequency , a method for measuring the significance of a word in a document compared to its frequency across all documents in a collection. scores with semantic similarity of the full texts of our sample of GPT-generated articles that fall into the “Environment” and “Health” categories, reflects the two categories in question. However, as can be seen in Figure 1, it also reveals overlap and sub-areas. The y-axis shows word prominences through word positions and font sizes, while the x-axis indicates semantic similarity. In addition to a certain amount of overlap, this reveals sub-areas, which are best described as two distinct events within the word rain. The event on the left bundles terms related to the development and management of health and healthcare with “challenges,” “impact,” and “potential of artificial intelligence”emerging as semantically related terms. Terms related to research infrastructures, environmental, epistemic, and technological concepts are arranged further down in the same event (e.g., “system,” “climate,” “understanding,” “knowledge,” “learning,” “education,” “sustainable”). A second distinct event further to the right bundles terms associated with fish farming and aquatic medicinal plants, highlighting the presence of an aquaculture cluster.  Here, the prominence of groups of terms such as “used,” “model,” “-based,” and “traditional” suggests the presence of applied research on these topics. The two events making up the word rain visualization, are linked by a less dominant but overlapping cluster of terms related to “energy” and “water.”

The bar chart of the terms in the paper subset (see Figure 2) complements the word rain visualization by depicting the most prominent terms in the full texts along the y-axis. Here, word prominences across health and environment papers are arranged descendingly, where values outside parentheses are TF-IDF values (relative frequencies) and values inside parentheses are raw term frequencies (absolute frequencies).

Finding 3: Google Scholar presents results from quality-controlled and non-controlled citation databases on the same interface, providing unfiltered access to GPT-fabricated questionable papers.

Google Scholar’s central position in the publicly accessible scholarly communication infrastructure, as well as its lack of standards, transparency, and accountability in terms of inclusion criteria, has potentially serious implications for public trust in science. This is likely to exacerbate the already-known potential to exploit Google Scholar for evidence hacking (Tripodi et al., 2023) and will have implications for any attempts to retract or remove fraudulent papers from their original publication venues. Any solution must consider the entirety of the research infrastructure for scholarly communication and the interplay of different actors, interests, and incentives.

We searched and scraped Google Scholar using the Python library Scholarly (Cholewiak et al., 2023) for papers that included specific phrases known to be common responses from ChatGPT and similar applications with the same underlying model (GPT3.5 or GPT4): “as of my last knowledge update” and/or “I don’t have access to real-time data” (see Appendix A). This facilitated the identification of papers that likely used generative AI to produce text, resulting in 227 retrieved papers. The papers’ bibliographic information was automatically added to a spreadsheet and downloaded into Zotero. 6 An open-source reference manager, https://zotero.org .

We employed multiple coding (Barbour, 2001) to classify the papers based on their content. First, we jointly assessed whether the paper was suspected of fraudulent use of ChatGPT (or similar) based on how the text was integrated into the papers and whether the paper was presented as original research output or the AI tool’s role was acknowledged. Second, in analyzing the content of the papers, we continued the multiple coding by classifying the fraudulent papers into four categories identified during an initial round of analysis—health, environment, computing, and others—and then determining which subjects were most affected by this issue (see Table 1). Out of the 227 retrieved papers, 88 papers were written with legitimate and/or declared use of GPTs (i.e., false positives, which were excluded from further analysis), and 139 papers were written with undeclared and/or fraudulent use (i.e., true positives, which were included in further analysis). The multiple coding was conducted jointly by all authors of the present article, who collaboratively coded and cross-checked each other’s interpretation of the data simultaneously in a shared spreadsheet file. This was done to single out coding discrepancies and settle coding disagreements, which in turn ensured methodological thoroughness and analytical consensus (see Barbour, 2001). Redoing the category coding later based on our established coding schedule, we achieved an intercoder reliability (Cohen’s kappa) of 0.806 after eradicating obvious differences.

The ranking algorithm of Google Scholar prioritizes highly cited and older publications (Martín-Martín et al., 2016). Therefore, the position of the articles on the search engine results pages was not particularly informative, considering the relatively small number of results in combination with the recency of the publications. Only the query “as of my last knowledge update” had more than two search engine result pages. On those, questionable articles with undeclared use of GPTs were evenly distributed across all result pages (min: 4, max: 9, mode: 8), with the proportion of undeclared use being slightly higher on average on later search result pages.

To understand how the papers making fraudulent use of generative AI were disseminated online, we programmatically searched for the paper titles (with exact string matching) in Google Search from our local IP address (see Appendix B) using the googlesearch – python library(Vikramaditya, 2020). We manually verified each search result to filter out false positives—results that were not related to the paper—and then compiled the most prominent URLs by field. This enabled the identification of other platforms through which the papers had been spread. We did not, however, investigate whether copies had spread into SciHub or other shadow libraries, or if they were referenced in Wikipedia.

We used descriptive statistics to count the prevalence of the number of GPT-fabricated papers across topics and venues and top domains by subject. The pandas software library for the Python programming language (The pandas development team, 2024) was used for this part of the analysis. Based on the multiple coding, paper occurrences were counted in relation to their categories, divided into indexed journals, non-indexed journals, student papers, and working papers. The schemes, subdomains, and subdirectories of the URL strings were filtered out while top-level domains and second-level domains were kept, which led to normalizing domain names. This, in turn, allowed the counting of domain frequencies in the environment and health categories. To distinguish word prominences and meanings in the environment and health-related GPT-fabricated questionable papers, a semantically-aware word cloud visualization was produced through the use of a word rain (Centre for Digital Humanities Uppsala, 2023) for full-text versions of the papers. Font size and y-axis positions indicate word prominences through TF-IDF scores for the environment and health papers (also visualized in a separate bar chart with raw term frequencies in parentheses), and words are positioned along the x-axis to reflect semantic similarity (Skeppstedt et al., 2024), with an English Word2vec skip gram model space (Fares et al., 2017). An English stop word list was used, along with a manually produced list including terms such as “https,” “volume,” or “years.”

• Artificial Intelligence
• / Search engines

Cite this Essay

Haider, J., Söderström, K. R., Ekström, B., & Rödl, M. (2024). GPT-fabricated scientific papers on Google Scholar: Key features, spread, and implications for preempting evidence manipulation. Harvard Kennedy School (HKS) Misinformation Review . https://doi.org/10.37016/mr-2020-156

• / Appendix B

Bibliography

Antkare, I. (2020). Ike Antkare, his publications, and those of his disciples. In M. Biagioli & A. Lippman (Eds.), Gaming the metrics (pp. 177–200). The MIT Press. https://doi.org/10.7551/mitpress/11087.003.0018

Barbour, R. S. (2001). Checklists for improving rigour in qualitative research: A case of the tail wagging the dog? BMJ , 322 (7294), 1115–1117. https://doi.org/10.1136/bmj.322.7294.1115

Bom, H.-S. H. (2023). Exploring the opportunities and challenges of ChatGPT in academic writing: A roundtable discussion. Nuclear Medicine and Molecular Imaging , 57 (4), 165–167. https://doi.org/10.1007/s13139-023-00809-2

Cabanac, G., & Labbé, C. (2021). Prevalence of nonsensical algorithmically generated papers in the scientific literature. Journal of the Association for Information Science and Technology , 72 (12), 1461–1476. https://doi.org/10.1002/asi.24495

Cabanac, G., Labbé, C., & Magazinov, A. (2021). Tortured phrases: A dubious writing style emerging in science. Evidence of critical issues affecting established journals . arXiv. https://doi.org/10.48550/arXiv.2107.06751

Carrion, M. L. (2018). “You need to do your research”: Vaccines, contestable science, and maternal epistemology. Public Understanding of Science , 27 (3), 310–324. https://doi.org/10.1177/0963662517728024

Centre for Digital Humanities Uppsala (2023). CDHUppsala/word-rain [Computer software]. https://github.com/CDHUppsala/word-rain

Chinn, S., & Hasell, A. (2023). Support for “doing your own research” is associated with COVID-19 misperceptions and scientific mistrust. Harvard Kennedy School (HSK) Misinformation Review, 4 (3). https://doi.org/10.37016/mr-2020-117

Cholewiak, S. A., Ipeirotis, P., Silva, V., & Kannawadi, A. (2023). SCHOLARLY: Simple access to Google Scholar authors and citation using Python (1.5.0) [Computer software]. https://doi.org/10.5281/zenodo.5764801

Dadkhah, M., Lagzian, M., & Borchardt, G. (2017). Questionable papers in citation databases as an issue for literature review. Journal of Cell Communication and Signaling , 11 (2), 181–185. https://doi.org/10.1007/s12079-016-0370-6

Dadkhah, M., Oermann, M. H., Hegedüs, M., Raman, R., & Dávid, L. D. (2023). Detection of fake papers in the era of artificial intelligence. Diagnosis , 10 (4), 390–397. https://doi.org/10.1515/dx-2023-0090

DeGeurin, M. (2024, March 19). AI-generated nonsense is leaking into scientific journals. Popular Science. https://www.popsci.com/technology/ai-generated-text-scientific-journals/

Dunlap, R. E., & Brulle, R. J. (2020). Sources and amplifiers of climate change denial. In D.C. Holmes & L. M. Richardson (Eds.), Research handbook on communicating climate change (pp. 49–61). Edward Elgar Publishing. https://doi.org/10.4337/9781789900408.00013

Fares, M., Kutuzov, A., Oepen, S., & Velldal, E. (2017). Word vectors, reuse, and replicability: Towards a community repository of large-text resources. In J. Tiedemann & N. Tahmasebi (Eds.), Proceedings of the 21st Nordic Conference on Computational Linguistics (pp. 271–276). Association for Computational Linguistics. https://aclanthology.org/W17-0237

Google Scholar Help. (n.d.). Inclusion guidelines for webmasters . https://scholar.google.com/intl/en/scholar/inclusion.html

Gu, J., Wang, X., Li, C., Zhao, J., Fu, W., Liang, G., & Qiu, J. (2022). AI-enabled image fraud in scientific publications. Patterns , 3 (7), 100511. https://doi.org/10.1016/j.patter.2022.100511

Gusenbauer, M., & Haddaway, N. R. (2020). Which academic search systems are suitable for systematic reviews or meta-analyses? Evaluating retrieval qualities of Google Scholar, PubMed, and 26 other resources. Research Synthesis Methods , 11 (2), 181–217.   https://doi.org/10.1002/jrsm.1378

Haider, J., & Åström, F. (2017). Dimensions of trust in scholarly communication: Problematizing peer review in the aftermath of John Bohannon’s “Sting” in science. Journal of the Association for Information Science and Technology , 68 (2), 450–467. https://doi.org/10.1002/asi.23669

Huang, J., & Tan, M. (2023). The role of ChatGPT in scientific communication: Writing better scientific review articles. American Journal of Cancer Research , 13 (4), 1148–1154. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10164801/

Jones, N. (2024). How journals are fighting back against a wave of questionable images. Nature , 626 (8000), 697–698. https://doi.org/10.1038/d41586-024-00372-6

Kitamura, F. C. (2023). ChatGPT is shaping the future of medical writing but still requires human judgment. Radiology , 307 (2), e230171. https://doi.org/10.1148/radiol.230171

Littell, J. H., Abel, K. M., Biggs, M. A., Blum, R. W., Foster, D. G., Haddad, L. B., Major, B., Munk-Olsen, T., Polis, C. B., Robinson, G. E., Rocca, C. H., Russo, N. F., Steinberg, J. R., Stewart, D. E., Stotland, N. L., Upadhyay, U. D., & Ditzhuijzen, J. van. (2024). Correcting the scientific record on abortion and mental health outcomes. BMJ , 384 , e076518. https://doi.org/10.1136/bmj-2023-076518

Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, S., & Wang, Z. (2023). ChatGPT and a new academic reality: Artificial Intelligence-written research papers and the ethics of the large language models in scholarly publishing. Journal of the Association for Information Science and Technology, 74 (5), 570–581. https://doi.org/10.1002/asi.24750

Martín-Martín, A., Orduna-Malea, E., Ayllón, J. M., & Delgado López-Cózar, E. (2016). Back to the past: On the shoulders of an academic search engine giant. Scientometrics , 107 , 1477–1487. https://doi.org/10.1007/s11192-016-1917-2

Martín-Martín, A., Thelwall, M., Orduna-Malea, E., & Delgado López-Cózar, E. (2021). Google Scholar, Microsoft Academic, Scopus, Dimensions, Web of Science, and OpenCitations’ COCI: A multidisciplinary comparison of coverage via citations. Scientometrics , 126 (1), 871–906. https://doi.org/10.1007/s11192-020-03690-4

Simon, F. M., Altay, S., & Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School (HKS) Misinformation Review, 4 (5). https://doi.org/10.37016/mr-2020-127

Skeppstedt, M., Ahltorp, M., Kucher, K., & Lindström, M. (2024). From word clouds to Word Rain: Revisiting the classic word cloud to visualize climate change texts. Information Visualization , 23 (3), 217–238. https://doi.org/10.1177/14738716241236188

Swedish Research Council. (2017). Good research practice. Vetenskapsrådet.

Stokel-Walker, C. (2024, May 1.). AI Chatbots Have Thoroughly Infiltrated Scientific Publishing . Scientific American. https://www.scientificamerican.com/article/chatbots-have-thoroughly-infiltrated-scientific-publishing/

Subbaraman, N. (2024, May 14). Flood of fake science forces multiple journal closures: Wiley to shutter 19 more journals, some tainted by fraud. The Wall Street Journal . https://www.wsj.com/science/academic-studies-research-paper-mills-journals-publishing-f5a3d4bc

The pandas development team. (2024). pandas-dev/pandas: Pandas (v2.2.2) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.10957263

Thorp, H. H. (2023). ChatGPT is fun, but not an author. Science , 379 (6630), 313–313. https://doi.org/10.1126/science.adg7879

Tripodi, F. B., Garcia, L. C., & Marwick, A. E. (2023). ‘Do your own research’: Affordance activation and disinformation spread. Information, Communication & Society , 27 (6), 1212–1228. https://doi.org/10.1080/1369118X.2023.2245869

Vikramaditya, N. (2020). Nv7-GitHub/googlesearch [Computer software]. https://github.com/Nv7-GitHub/googlesearch

This research has been supported by Mistra, the Swedish Foundation for Strategic Environmental Research, through the research program Mistra Environmental Communication (Haider, Ekström, Rödl) and the Marcus and Amalia Wallenberg Foundation [2020.0004] (Söderström).

Competing Interests

The authors declare no competing interests.

The research described in this article was carried out under Swedish legislation. According to the relevant EU and Swedish legislation (2003:460) on the ethical review of research involving humans (“Ethical Review Act”), the research reported on here is not subject to authorization by the Swedish Ethical Review Authority (“etikprövningsmyndigheten”) (SRC, 2017).

This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided that the original author and source are properly credited.

Data Availability

All data needed to replicate this study are available at the Harvard Dataverse: https://doi.org/10.7910/DVN/WUVD8X

Acknowledgements

The authors wish to thank two anonymous reviewers for their valuable comments on the article manuscript as well as the editorial group of Harvard Kennedy School (HKS) Misinformation Review for their thoughtful feedback and input.

• Skip to Guides Search
• Skip to breadcrumb
• Skip to main content
• Skip to footer
• Skip to chat link
• Report accessibility issues and get help
• Go to Penn Libraries Home
• Go to Franklin catalog

Critical Writing Program Fall 2024 Critical Writing Seminar in PHIL: The Ethics of Artificial Intelligence: Researching the White Paper

• Getting started
• News and Opinion Sites
• Academic Sources
• Grey Literature
• Substantive News Sources
• What to Do When You Are Stuck
• Understanding a citation
• Examples of Quotation
• Examples of Paraphrase
• Chicago Manual of Style: Citing Images
• Researching the Op-Ed
• Researching Prospective Employers
• Resume Resources
• Cover Letter Resources

Research the White Paper

Researching the White Paper:

The process of researching and composing a white paper shares some similarities with the kind of research and writing one does for a high school or college research paper. What’s important for writers of white papers to grasp, however, is how much this genre differs from a research paper.  First, the author of a white paper already recognizes that there is a problem to be solved, a decision to be made, and the job of the author is to provide readers with substantive information to help them make some kind of decision--which may include a decision to do more research because major gaps remain. 

Thus, a white paper author would not “brainstorm” a topic. Instead, the white paper author would get busy figuring out how the problem is defined by those who are experiencing it as a problem. Typically that research begins in popular culture--social media, surveys, interviews, newspapers. Once the author has a handle on how the problem is being defined and experienced, its history and its impact, what people in the trenches believe might be the best or worst ways of addressing it, the author then will turn to academic scholarship as well as “grey” literature (more about that later).  Unlike a school research paper, the author does not set out to argue for or against a particular position, and then devote the majority of effort to finding sources to support the selected position.  Instead, the author sets out in good faith to do as much fact-finding as possible, and thus research is likely to present multiple, conflicting, and overlapping perspectives. When people research out of a genuine desire to understand and solve a problem, they listen to every source that may offer helpful information. They will thus have to do much more analysis, synthesis, and sorting of that information, which will often not fall neatly into a “pro” or “con” camp:  Solution A may, for example, solve one part of the problem but exacerbate another part of the problem. Solution C may sound like what everyone wants, but what if it’s built on a set of data that have been criticized by another reliable source?  And so it goes. 

For example, if you are trying to write a white paper on the opioid crisis, you may focus on the value of  providing free, sterilized needles--which do indeed reduce disease, and also provide an opportunity for the health care provider distributing them to offer addiction treatment to the user. However, the free needles are sometimes discarded on the ground, posing a danger to others; or they may be shared; or they may encourage more drug usage. All of those things can be true at once; a reader will want to know about all of these considerations in order to make an informed decision. That is the challenging job of the white paper author.     
 The research you do for your white paper will require that you identify a specific problem, seek popular culture sources to help define the problem, its history, its significance and impact for people affected by it.  You will then delve into academic and grey literature to learn about the way scholars and others with professional expertise answer these same questions. In this way, you will create creating a layered, complex portrait that provides readers with a substantive exploration useful for deliberating and decision-making. You will also likely need to find or create images, including tables, figures, illustrations or photographs, and you will document all of your sources. 

Liaison Librarian

Connect to a Librarian Live Chat or "Ask a Question"

• Librarians staff live chat from 9-5 Monday through Friday . You can also text to chat: 215-543-7674
• You can submit a question 24 hours a day and we aim to respond within 24 hours 
• You can click the "Schedule Appointment" button above in librarian's profile box (to the left), to schedule a consultation with her in person or by video conference.  
• You can also make an appointment with a  Librarian by subject specialization . 
• Connect by email with a subject librarian

Find more easy contacts at our Quick Start Guide

• Next: Getting started >>
• Last Updated: Sep 9, 2024 4:14 PM
• URL: https://guides.library.upenn.edu/c.php?g=1423266

An official website of the United States government

Here’s how you know

Official websites use .gov A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS A lock ( Lock A locked padlock ) or https:// means you’ve safely connected to the .gov website. Share sensitive information only on official, secure websites.

https://www.nist.gov/news-events/news/2024/08/us-ai-safety-institute-signs-agreements-regarding-ai-safety-research

U.S. AI Safety Institute Signs Agreements Regarding AI Safety Research, Testing and Evaluation With Anthropic and OpenAI

These first-of-their-kind agreements between the u.s. government and industry will help advance safe and trustworthy ai innovation for all..

GAITHERSBURG, Md. — Today, the U.S. Artificial Intelligence Safety Institute at the U.S. Department of Commerce’s National Institute of Standards and Technology (NIST) announced agreements that enable formal collaboration on AI safety research, testing and evaluation with both Anthropic and OpenAI.

Each company’s Memorandum of Understanding establishes the framework for the U.S. AI Safety Institute to receive access to major new models from each company prior to and following their public release. The agreements will enable collaborative research on how to evaluate capabilities and safety risks, as well as methods to mitigate those risks. 

“Safety is essential to fueling breakthrough technological innovation. With these agreements in place, we look forward to beginning our technical collaborations with Anthropic and OpenAI to advance the science of AI safety,” said Elizabeth Kelly, director of the U.S. AI Safety Institute. “These agreements are just the start, but they are an important milestone as we work to help responsibly steward the future of AI.”

Additionally, the U.S. AI Safety Institute plans to provide feedback to Anthropic and OpenAI on potential safety improvements to their models, in close collaboration with its partners at the U.K. AI Safety Institute. 

The U.S. AI Safety Institute builds on NIST’s more than 120-year legacy of advancing measurement science, technology, standards and related tools. Evaluations under these agreements will further NIST’s work on AI by facilitating deep collaboration and exploratory research on advanced AI systems across a range of risk areas.

Evaluations conducted pursuant to these agreements will help advance the safe, secure and trustworthy development and use of AI by building on the Biden-Harris administration’s Executive Order on AI and the voluntary commitments made to the administration by leading AI model developers.

About the U.S. AI Safety Institute

The U.S. AI Safety Institute , located within the Department of Commerce at the National Institute of Standards and Technology (NIST), was established following the Biden-Harris administration’s 2023 Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence to advance the science of AI safety and address the risks posed by advanced AI systems. It is tasked with developing the testing, evaluations and guidelines that will help accelerate safe AI innovation here in the United States and around the world. 

• More Blog Popular
• Who's Who Legal
• Instruct Counsel
• My newsfeed
• Save & file
• View original
• Follow Please login to follow content.

add to folder:

• My saved (default)

Register now for your free, tailored, daily legal newsfeed service.

Find out more about Lexology or get in touch by visiting our About page.

Digital Replicas and the First Amendment: The Latest in Artificial Intelligence Legislation

Laws target unauthorized AI-generated replicas of an individual's voice or appearance

Image-generating technology is accelerating quickly, making it much more likely that you will be seeing "digital replicas" (sometimes referred to as "deepfakes") of celebrities and non-celebrities alike across film, television, documentaries, marketing, advertising, and election materials. Meanwhile, legislators are advocating for protections against the exploitation of name, image, and likeness while attempting to balance the First Amendment rights creatives enjoy.

In this post, we analyze significant recent developments in the regulation of digital replicas, particularly the U.S. Copyright Office Report on Copyright and Artificial Intelligence, Part 1: Digital Replicas , the federal NO FAKES Act , and other legislative and regulatory efforts to protect individuals against unauthorized generative AI re-creations, and how they implicate First Amendment principles or might impact the creation and distribution of content.

U.S. Copyright Office and First Amendment Considerations

On July 31, 2024, the U.S. Copyright Office published a report summarizing its proposal for a new federal law that would protect against unauthorized digital replicas, recognizing there are gaps in existing legal frameworks ( i.e. , a patchwork of state right-of-publicity law and regulation, the Copyright Act, the Federal Trade Commission Act, the Lanham Act, and the Communications Act).

After collecting thousands of comments and conducting its own study, the office outlined the contours of a proposed new federal law targeting unauthorized digital replicas (AI-generated or otherwise). Some of those features include:

• Protection for private individuals as well as celebrities;
• Liability for distribution only (not creation);
• Secondary liability for distributors or other types of intermediaries (and exclusion from Section 230), but with safe harbors that incentivize prompt removal; and
• Availability of injunctive relief, statutory damages, and attorneys' fees.

Although the office's proposal is not strictly limited to commercial uses, the office acknowledged that digital replicas have legitimate uses in the context of constitutionally protected speech, such as in news reporting, artistic works, parody, and political opinion.

There was significant disagreement among the thousands of comments the Copyright Office received on how to achieve the balance between protecting individuals from unauthorized deepfakes while respecting speech and distribution of content. Many commenters supported specific categorical exemptions from any federal law ( e.g. , for news reporting, various types of expressive works, sports broadcasting, as well as parody, comment, and criticism), which provide greater certainty. The Motion Picture Association's comments, for instance, included several examples of expressive uses deserving of categorical protection, such as documentaries that use digital replicas "to re-create scenes from history where no actual footage exists." Other commenters instead preferred a fact-specific balancing test, worried that categorical exclusions could be simultaneously over- and under-inclusive depending on the context.

Ultimately, the Copyright Office recommended that any federal law include a balancing framework rather than categorical exemptions, which would require courts to assess a full range of factors including:

• the purpose of the use, including whether it is commercial;
• whether the use is expressive or political in nature;
• the relevance of the digital replica to the purpose of the use;
• whether the use is intentionally deceptive;
• whether the replica was labeled;
• the extent of the harm caused; and
• the good faith of the user.

As with any balancing test, this provides greater flexibility to courts and other decisionmakers in striking the appropriate balance on a case-by-case basis, but also creates greater uncertainty, the likelihood of inconsistent decisions across different jurisdictions, increased litigation expense, and the potential for chilling speech as the result of creators wanting to steer clear of potential liability.

NO FAKES Act and Other Legislation

On July 31, 2024, the same day the U.S. Copyright Office released its report, a bipartisan group of federal lawmakers including Senators Coons, Blackburn, Klobuchar, and Tillis introduced the Nurture Originals, Foster Art, and Keep Entertainment Safe Act, commonly referred to as the NO FAKES Act . That bill, which is currently pending in the Senate, is designed to protect the voice and likeness of actors, singers, performers, and other individuals from the unauthorized use of AI-generated replicas. Among other things, the legislation attempts to address concerns regarding deepfake technology and seeks to create uniform protections across the United States. Upon release, entertainment industry stakeholders, including SAG-AFTRA, the Recording Industry Association of America, the Motion Picture Association, and The Walt Disney Company, among others, released statements endorsing the proposed legislation, as did technology companies OpenAI and IBM.

While the NO FAKES Act seeks to hold individuals or companies liable for damages for producing, hosting, or sharing digital replicas, including AI-generated replicas, of an individual performing in audiovisual works, images, or sound recordings without that person's consent or participation, the act attempts to balance First Amendment protections in the draft legislation, including exclusions that apply when the digital replica is:

• Produced or used in a bona fide news, public affairs, or sports broadcast if the replica is the subject of or materially relevant to the subject of the broadcast;
• Used in a documentary or in a historical or biographical manner, including some degree of fictionalization, unless the use intends to and does create the false impression the work is authentic and the person participated;
• Produced or used consistent with the public interest in bona fide commentary, criticism, scholarship, satire, or parody;
• Used in a fleeting or negligible manner;
• Produced or used in an advertisement or commercial announcement for any of the foregoing purposes.

Of course, even with explicit First Amendment exclusions being part of the draft bill, the legislation raises various questions about when these protections apply—for example, what is a "bona fide news" broadcast? What constitutes "some degree of fictionalization," and how "fleeting or negligible" does the use have to be to qualify? And do these First Amendment protections sufficiently protect creative and artistic (or even commercial) works? If this legislation is passed, courts will likely be wrestling with these and other questions.

In recent months, other lawmakers have proposed their own bills to address digital replicas, all of which also grapple with balancing prohibitions with the First Amendment. For example, on August 9, 2024, Congressman Issa (CA-48) publicly released a discussion draft of the Preventing Abuse of Digital Replicas Act (PADRA) , which seeks to modify the Lanham (Trademark) Act to address digital replicas, including by providing a rebuttable presumption that such uses are likely to cause confusion, mistake, or deceive, with a carveout where the use is protected under the First Amendment. In Tennessee, the recent ELVIS Act extends the right of publicity protection for names and likenesses to include individuals' voices in light of the increased popularity and accessibility of AI-generated audio tracks, while at the same time memorializing First Amendment and fair use as an express exemption. Illinois also passed a similar amendment to the state's Right of Publicity Act to cover digital replicas, with express exclusions for news, sports broadcasts, or use in political or public interest purposes, documentary or biographical works, or for satire or parody—all so long as they do not create the false impression that the replica is authentic.

And in California, state lawmakers recently passed AB 1836 , one of California's proposed digital replica laws, which would modify the state's right of publicity statute to prohibit the use of digital replicas of a deceased person in expressive audiovisual works or sound recordings without prior consent. That legislation has similar First Amendment exclusions to those proposed in the NO FAKES Act discussed above. It is currently awaiting the governor's signature. And as of August 27, 2024, California lawmakers sent the governor AB 2602 , which would require movie studios, streamers, and other content creators to seek permission from performers to create digital replicas. These are but a few of the bills being debated, all of which attempt to balance the goals of the legislation while still protecting artists and content creators.

There is tension among the legislative proposals regarding how and to what degree they intend to incorporate and apply First Amendment protections for expressive works, with solutions ranging from exempting categories of works from the bill's scope to creating a balancing test of factors for courts to consider. Ultimately, it is still not clear how lawmakers—and eventually, courts—will resolve these pressing and evolving issues.

Filed under

• IT & Data Protection
• Davis Wright Tremaine LLP
• Artificial intelligence
• Generative AI

Organisations

• Federal Trade Commission (USA)
• Motion Picture Association

Popular articles from this firm

How does the demise of chevron deference affect employee benefit plans and erisa regulatory actions and litigation *, stay advised: 2024, issue 17 *, abcs of distribution to grocery *, california legislature sends bills regulating ai to the governor *, navigating the financial markets regulatory landscape: takeaways after the totsa settlement *.

If you would like to learn how Lexology can drive your content marketing strategy forward, please email [email protected] .

Professional development

An Introduction to Data Visualisation - How to Tell Stories with Data - Learn Live

Contract & Commercial Law Update 2024 for Contract Drafters & In-House Counsel

ChatGPT for HR Professionals - Unravelling the Potential

AI & Market Competition - Live at Your Desk - Learn Live

Advanced Copyright Law - Key Areas in Depth

Related practical resources PRO

• How-to guide How-to guide: How to develop a vulnerability disclosure program (VDP) for your organization to ensure cybersecurity (USA)
• How-to guide How-to guide: How to prepare for and respond to a governmental investigation or enforcement action for violation of US privacy laws (USA)
• How-to guide How-to guide: AI and smart contracts (USA)

Related research hubs

University of Washington Information School

i School's Caliskan wins award to battle bias in artificial intelligence

Imagine losing out on your dream job due to bias in AI tools used in the resume screening process or having your health care compromised for the same reason.

Those are the disturbing scenarios that Aylin Caliskan , an assistant professor in the University of Washington Information School, is dedicated to thwarting.

Caliskan was recently awarded a $603,342 National Science Foundation Faculty Early Career Development (CAREER) Award for her project titled, “The Impact of Associations and Biases in Generative AI on Society.” She is planning to develop computational methods to measure biases in generative artificial intelligence systems and their impact on humans and society. Caliskan says her goal is reducing bias in AI and human-AI collaboration.

“Hopefully, in the long term, we will be able to raise awareness and provide tools to reduce the harmful consequences of bias,” said Caliskan, who became a co-director of the UW Tech Policy Lab earlier this year. Her research in computer science and artificial intelligence will also provide empirical evidence for tech policy.

Caliskan noted that AI is used in a variety of places that many people don’t realize. Companies often use AI to screen job applications; some colleges use it to screen student applications; and health-care providers use AI in reviewing patient data. 

But because AI is trained on data produced by humans, it learns biases similar to those found in society. Women and people of different ethnicities are more frequently discriminated against in AI than white males, Caliskan said. She cited an example from current use of generative AI in health care, where African American patients may receive less effective or lower-cost medications when prescribed through AI than patients of European descent.

Caliskan’s work was among the first to develop methods to detect and quantify bias in AI. One of the difficulties she faces is that AI doesn’t work or “think” exactly like humans, despite being developed by them. However, AI is being used on a large scale and is helping to shape society. 

Another challenge for Caliskan is that not all AI is the same. Many companies have their own proprietary AI systems that they may or may not be willing to allow researchers like Caliskan to study. 

One of the keys to reducing bias in AI is understanding the mechanisms of bias and where the bias originated, she said. Some bias is cultural, societal or historical. Figuring out what is “fair” in a specific context and task isn’t trivial.

“There are many fairness notions,” Caliskan said. “We don’t have simple, straightforward answers to these complex open questions.”

Caliskan notes she grew up a multilingual immigrant, which fostered her interest in the subject of fairness. She speaks German, Turkish and Bulgarian as well as English.

“I was able to observe and live in different cultures in my childhood and observe different societies,” she said. “I have always been fascinated by culture and languages.”

Since coming to the UW, Caliskan has been invited to speak at AI-related events at Stanford University, Howard University, the Santa Fe Institute, and the International Joint Conferences on Artificial Intelligence. Her paper rigorously showing that AI reflects cultural stereotypes was published in Science magazine. 

In 2023, Caliskan was listed among the 100 Brilliant Women in AI Ethics by the Women in AI Ethics organization. She previously received an NSF award for her work on privacy and fairness in planning while using third-party sources. Caliskan is teaching a course on generative AI literacy this fall. 

Caliskan’s NSF grant will last for five years, but she doesn’t see her work on the subject ending then. 

“I see this research going on my entire life,” she said. “Since bias cannot be entirely eliminated, this is a lifelong problem.”

However, Caliskan believes that identifying, measuring and reducing bias can help align AI with societal values and raise awareness.

“I don’t think eliminating bias entirely in AI or people is possible,” she said. But “when we know we are biased, we adjust our behavior.”

Full Results

Customize your experience.

• DOI: 10.17762/ijritcc.v11i9.9328
• Corpus ID: 266992298

Generative Artificial Intelligence and GPT using Deep Learning: A Comprehensive Vision, Applications Trends and Challenges

• Et al. Ruchi Sharma
• Published in International Journal on… 5 November 2023
• Computer Science, Art

Related Papers

Showing 1 through 3 of 0 Related Papers