Artificial Intelligence

Artificial intelligence (AI) is the project of building computer systems that perform tasks commonly associated with intelligent beings¹. Britannica defines it as the ability of a digital computer or computer-controlled robot to perform such tasks¹. Regulators describe the same object more operationally: under the OECD definition, an AI system is a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions². The EU AI Act defines an AI system in the same terms, adding that it is designed to operate with varying levels of autonomy and may exhibit adaptiveness after deployment³. NIST refers to an AI system as an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions⁴.

The label covers a wide working range of deployed systems. The OECD formulation speaks of outputs such as predictions, content, recommendations or decisions², and such outputs now appear in services from machine translation to large language models⁵⁶. Medical applications are already counted: in 2023 the U.S. FDA approved 223 AI-enabled medical devices, up from just six in 2015⁷. None of this amounts to general intelligence — no AI yet matches full human flexibility over wider domains or in tasks requiring much everyday knowledge, though some perform specific tasks as well as humans¹.

The name itself is dated. The term was coined for a summer 1956 research project at Dartmouth College, proposed in 1955 by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon⁸. What followed — symbolic programs and expert systems, a funding winter, a learning turn, deep networks and large language models — is the substance of this article⁸. Governance came later: OECD principles in 2019, the NIST risk framework in 2023, and the EU AI Act, in force since 1 August 2024²⁴³.

Alan Turing proposed the field's first test of machine thinking before it had a name⁹. His paper “Computing Machinery and Intelligence” appeared in Mind in October 1950⁹. Turing opened with the question “Can machines think?” and judged it too meaningless to deserve discussion, replacing it with a game he called the imitation game⁹. The game is played with three people — a man, a woman and an interrogator — and the interrogator must determine which of the two unseen players is the man and which is the woman, knowing them only as X and Y⁹. So that tone of voice could not help, the answers were to be written, or better still typewritten⁹.

Turing made his expectations explicit. In about fifty years' time, he predicted, it would be possible to programme computers to play the game so well that an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes of questioning⁹. He hoped machines would eventually compete with men in all purely intellectual fields, and noted that many people thought a very abstract activity, like chess, the best place to start⁹. The original question, he added, was not one to argue but to outgrow: he expected the growth of machine capability to settle it in practice, with educated opinion eventually speaking of machines thinking without expecting contradiction⁹.

The paper set a pattern that would recur. Five years after Turing wrote, the Dartmouth proposal turned the simulation of intelligence into a research programme: every aspect of learning or any other feature of intelligence, it conjectured, can be described precisely enough for a machine to be made to simulate it⁸. How far that programme has run, and where it has stalled, is measured in the milestones that follow⁸.

The name arrived by proposal. On 31 August 1955, John McCarthy of Dartmouth College, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories put forward a plan for the following summer⁸. It proposed a 2-month, 10-person study of artificial intelligence to be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire — the first appearance of the term⁸. The study was to proceed on a single founding 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 it⁸.

The proposal was ambitious about method as well as subject. An attempt would be made, it said, to find how to make machines use language, form abstractions and concepts, solve problems then reserved for humans, and improve themselves⁸. The authors judged that a significant advance could be made on one or more of these problems if a carefully selected group of scientists worked on them together for a summer⁸. The term coined for that workshop has named the field ever since⁸.

Early artificial intelligence was largely symbol manipulation: the Dartmouth plan itself aimed at language, abstraction and concept formation⁸. Its most durable early product was the expert system — a computer program that uses artificial-intelligence methods to solve problems within a specialised domain that ordinarily requires human expertise¹⁰. The first expert system was developed in 1965 by Edward Feigenbaum and Joshua Lederberg of Stanford University and became known as Dendral; it was designed to analyse chemical compounds¹⁰. Such systems rest on two components: a knowledge base, an organised collection of facts about the domain, and an inference engine that interprets and evaluates those facts to provide an answer¹⁰.

Knowledge for the base was acquired from human experts through interviews and observations, usually represented as if-then production rules: if some condition is true, then a certain inference can be made or action taken¹⁰. Expert systems incorporated heuristic rules of thumb and, increasingly, the ability to learn from experience; they remain aids to, rather than replacements for, human experts¹⁰. The approach nevertheless met a structural limit: the combinatorial explosion, in which the number of paths through a network grows with the alternatives open at one time and the number of steps¹¹.

The reckoning came in 1973, when the UK Science Research Council published James Lighthill's survey, commissioned to give an unbiased view of the state of AI research primarily in the UK¹¹. The report supported work related to automation and to computer simulation of neurophysiological and psychological processes, but was highly critical of basic research in foundation areas such as robotics and language processing¹¹. Its central finding was damaging: in no part of the field, Lighthill judged, had the discoveries made so far produced the major impact that was then promised¹¹. The effect was a massive loss of confidence in AI by the UK academic establishment, including the funding body — a condition that persisted for almost a decade, the first AI winter¹¹. AI research continued, but the next attempt to mount a major activity in the area did not come until a September 1982 meeting on intelligent knowledge-based systems¹¹.

The alternative to symbols was nearly as old as the field itself. In 1943 Warren McCulloch and Walter Pitts showed that, because nervous activity is all-or-none, neural events and their relations can be treated by means of propositional logic¹². They found that the behaviour of every network of neurons could be described in logical terms, and that for any logical expression meeting certain conditions one could find a network behaving as it describes¹². Their paper became the first mathematical model of an artificial neuron and the seed of neural-network research¹².

Learning returned to the centre in 1986, when David Rumelhart, Geoffrey Hinton and Ronald Williams published “Learning representations by back-propagating errors” in Nature, volume 323, pages 533-536¹³. The procedure computes, for each weight of a network, a gradient indicating by what amount the error would increase or decrease if that weight were increased by a tiny amount¹³. The network is then adjusted according to that gradient, and the network's representations are learned from data rather than written by hand¹³.

The methods that resulted are now called deep learning: computational models composed of multiple processing layers that learn representations of data with multiple levels of abstraction¹³. In these models the layers of features are not designed by human engineers but discovered from data using a general-purpose learning procedure¹³. Their effect was felt well beyond vision: these methods dramatically improved the state of the art in speech recognition¹³.

ImageNet was built to give the field something to measure. Presented in 2009, it uses the hierarchical structure of WordNet, giving the database a large-scale ontology of images¹⁴. The team had constructed 12 subtrees containing 3.2 million images by the time of the paper, with the first release covering 5,247 synsets and 3.2 million images in total¹⁴. The images were labelled by human workers on Amazon Mechanical Turk, with about 50 million images planned for the completed database¹⁴.

The decisive contest was ILSVRC-2012. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images of the ImageNet challenge¹⁵. The network had 60 million parameters and 650,000 neurons, and consisted of five convolutional layers¹⁵. On ILSVRC-2010 test data it reached top-1 and top-5 error rates of 37.5% and 17.0%, considerably better than the previous state of the art¹⁵.

The winning entry was a variant entered in the ILSVRC-2012 competition, which achieved a top-5 test error rate of 15.3% against 26.2% for the second-best entry¹⁵. Training took between five and six days on two GTX 580 3GB GPUs, and the authors expected results to improve simply by waiting for faster GPUs and bigger datasets¹⁵. The deep-learning literature records that the success came from the efficient use of GPUs¹³, and that progress on object recognition had been slow until the ImageNet competition of 2012¹³.

In May 1997, IBM's Deep Blue became the first computer system to defeat a reigning world chess champion in a match under standard tournament controls¹⁶. Its six-game rematch against Garry Kasparov in New York ended 3.5-2.5 in the machine's favour¹⁶. The system evaluated 200 million chess positions per second, a processing speed of 11.38 billion floating-point operations per second¹⁶. The first encounter had ended differently: in February 1996 Deep Blue won the first game against Kasparov in Philadelphia — the first victory by a computer against a reigning world champion under regular time controls — but lost the match 4-2¹⁶. The 1997 victory marked an inflection point in computing¹⁶.

Go was the harder problem — long viewed as the most challenging of the classic games for artificial intelligence, owing to its enormous search space and the difficulty of evaluating board positions and moves¹⁷. Google DeepMind's AlphaGo attacked it with value networks to evaluate positions and policy networks to select moves, trained by supervised learning from human expert games combined with reinforcement learning from self-play¹⁷. The program achieved a 99.8% winning rate against other Go programs, and defeated the human European champion by 5 games to 0 — the first time a computer program had beaten a human professional in the full-sized game¹⁷.

Then came Lee Sedol. The player — winner of 18 world titles — was beaten 4-1 by AlphaGo in Seoul in March 2016, in a match watched by more than 200 million people worldwide¹⁸. Twenty years apart, the two matches made the same point in public: on bounded domains, machines could beat the best human players¹⁶¹⁷.

The milestones below run from the McCulloch-Pitts neuron of 1943 to the entry into force of the EU AI Act in 2024, as dated by the sources cited in this article¹²³.

In June 2017, Ashish Vaswani and colleagues published “Attention Is All You Need”, proposing the Transformer — a network architecture based solely on attention mechanisms, dispensing with recurrence and convolutions entirely⁵. It was more parallelizable than earlier architectures and required significantly less time to train⁵. On English-to-French translation it set a new single-model state-of-the-art BLEU score of 41.8 after 3.5 days on eight GPUs, a small fraction of the training costs of the best previous models⁵.

GPT-3 followed in 2020: an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, tested in the few-shot setting⁶. For all tasks it was applied without any gradient updates or fine-tuning, with tasks specified purely through text interaction with the model⁶. It performed strongly on many natural-language datasets — translation, question answering and cloze tasks among them — as well as on tasks requiring on-the-fly reasoning or domain adaptation⁶. The authors also identified datasets where few-shot learning still struggled, and others with methodological issues tied to training on large web corpora⁶. Human evaluators had difficulty distinguishing its generated news articles from articles written by humans⁶.

ChatGPT arrived on 30 November 2022 as a free research preview: a model that interacts in a conversational way, trained with reinforcement learning from human feedback — the same methods as InstructGPT, with slight differences in the data-collection setup¹⁹. It could answer follow-up questions, admit mistakes, challenge incorrect premises and reject inappropriate requests¹⁹. GPT-4 followed in March 2023, a large-scale multimodal model that accepts image and text inputs and produces text outputs²⁰.

GPT-4 is a Transformer-based model pre-trained to predict the next token in a document²⁰. While less capable than humans in many real-world scenarios, it exhibited human-level performance on various professional and academic benchmarks, including a simulated bar exam passed with a score around the top 10% of test takers²⁰. Its developer recorded another kind of result: some aspects of GPT-4's performance could be predicted from models trained with no more than 1/1,000th of its compute²⁰.

The limits are as documented as the capabilities. No AI yet matches full human flexibility over wider domains or in tasks requiring much everyday knowledge, though some perform specific tasks as well as humans¹. NIST's Generative AI Profile names confabulation — confidently presented erroneous or false content, colloquially called hallucination — as a phenomenon of generative systems, and a natural result of their design: they generate outputs that approximate the statistical distribution of their training data⁴. The NIST framework itself is intended to be voluntary, rights-preserving, non-sector-specific and use-case agnostic⁴. Even GPT-4, for all its examination results, was reported as less capable than humans in many real-world scenarios²⁰.

Governance began with principles. The OECD Recommendation on Artificial Intelligence — the first intergovernmental standard on AI — was adopted by the OECD Council at Ministerial level on 22 May 2019². It was revised on 8 November 2023 so that its definition of an AI system would stay technically accurate and reflect developments including generative AI, and the principles were updated in May 2024². They promote the use of AI that is innovative and trustworthy and that respects human rights and democratic values — from inclusive growth and human rights to transparency, robustness and accountability²; 47 adherents have committed to them².

In the United States, the federal instrument was a framework rather than a statute. NIST released its AI Risk Management Framework on 26 January 2023, after a consensus-driven, open, transparent and collaborative development process⁴. It is intended for voluntary use, to improve the ability to incorporate trustworthiness considerations into the design, development, use and evaluation of AI products, services and systems⁴. The framework is intended to be voluntary, rights-preserving, non-sector-specific and use-case agnostic, with a core composed of four functions: GOVERN, MAP, MEASURE and MANAGE⁴.

Europe legislated. The AI Act — Regulation (EU) 2024/1689 — entered into force on 1 August 2024 as the first-ever legal framework on AI, addressing the risks of AI³. It became applicable on 2 August 2026 with some exceptions: prohibited AI practices and AI literacy obligations applied from 2 February 2025, and the governance rules and obligations for general-purpose AI models became applicable on 2 August 2025³. AI use cases that can pose serious risks to health, safety or fundamental rights are classified as high-risk under the regulation³.

Adoption is now the measurable story⁷. In 2024, U.S. private AI investment grew to $109.1 billion — nearly 12 times China's $9.3 billion and 24 times the U.K.'s $4.5 billion⁷. Business usage accelerated alongside it: 78% of organizations reported using AI in 2024, up from 55% the year before⁷. Between November 2022 and October 2024, the inference cost of a system performing at the level of GPT-3.5 dropped over 280-fold, driven by increasingly capable small models⁷.

Capability gains are dated, not hypothetical⁷. Benchmarks introduced in 2023 to test advanced AI systems saw scores rise within a year by 18.8 points on MMMU, 48.9 on GPQA and 67.3 on SWE-bench⁷. In 2023 the U.S. FDA approved 223 AI-enabled medical devices, up from just six in 2015⁷. Two Nobel Prizes have recognised work related to deep learning⁷.

The frictions are documented alongside the gains⁷. AI-related incidents are rising sharply, yet standardized responsible-AI evaluations remain rare among major industrial model developers⁷. Governments are responding: in 2024, U.S. federal agencies introduced 59 AI-related regulations — more than double the number in 2023 — issued by twice as many agencies⁷. The gap the sources describe is one of assurance: what these systems can do is increasingly measured, and how reliably they can be trusted is not⁴.