Artificial intelligence

Artificial intelligence (AI) is the field of computer science that builds machines to perform tasks which, when people do them, are said to require intelligence: recognising a face, translating a sentence, steering a vehicle, or answering a question in prose. The phrase names both a research programme and the products that programme has produced, which is one reason arguments about AI often talk past one another.

Its modern centre of gravity is machine learning, in which behaviour is fitted to examples rather than written as explicit rules. That approach has produced striking competence inside narrow domains and no system that is generally intelligent in the way a person is; claims to the contrary remain claims rather than results.

This entry covers the definition and scope of the field, its history from symbolic reasoning to deep learning, its technical foundations, what deployed systems do, the economics of the industry that builds them, the governance being written around them, and the criticisms that divide their makers and their critics.

AI is defined by its goal rather than by a single method. The label covers a research discipline founded in the 1950s, a set of techniques that includes search, formal logic, probabilistic inference, and statistical learning, and a commercial category whose marketing uses the term loosely. John McCarthy, who coined the phrase in the proposal for the 1956 Dartmouth workshop, framed the field as the science and engineering of making intelligent machines 1. Marvin Minsky's formulation — machines doing things that would require intelligence if done by people — remains a workable operational starting point.

Two distinctions run through most definitions. The first separates systems that reason over explicit symbols from systems that learn patterns from data. The second separates narrow systems, which perform a specific task or a family of related tasks, from general ones, which could transfer competence across unrelated domains. Every system deployed commercially today is narrow in the second sense, however wide its range of tasks.

The definition is contested because intelligence is contested. Alan Turing's proposal to judge a machine by conversation deliberately sidestepped the question of inner understanding, and the philosophical literature since has argued about whether behaviour is the right evidence at all 2 3. Reference works therefore tend to define AI by its methods and its history rather than by a threshold that a system either passes or fails 3.

The field is usually dated to an eight-week summer workshop at Dartmouth College in 1956, convened on the basis of a 1955 proposal by McCarthy, Minsky, Nathaniel Rochester, and Claude Shannon 1. The proposal was explicit about the ambition: to find how to make machines use language, form abstractions and concepts, and improve themselves. Early work concentrated on symbolic reasoning, search, and theorem proving, and produced programmes that could play chess, prove simple theorems, and solve algebra word problems.

Optimism outran results. A 1973 report by James Lighthill in the United Kingdom concluded that the field had failed to deliver on its promises, and funding contracted in what became known as the first AI winter. A second boom arrived in the 1980s with expert systems, which encoded specialist knowledge as rules and were sold as decision support for insurance, medicine, and manufacturing; it ended in the late 1980s when those systems proved expensive to maintain and brittle outside their narrow domains.

The recovery came from statistics rather than logic. Through the 1990s and 2000s, machine learning, probabilistic graphical models, and large annotated datasets displaced hand-written rules, and the field's centre moved from laboratories to internet companies with data to spare. A convolutional network's win in the 2012 ImageNet competition marked the arrival of deep learning 4; the transformer architecture of 2017 made very large language models practical; and the public release of ChatGPT in 2022 turned a research programme into a mass-market product category 5.

The historiography is not settled. Some accounts treat the field as a sequence of research paradigms, others as a sequence of funding cycles, and others as a single long argument about whether intelligence can be engineered from the top down or must be learned from the bottom up 3.

Symbolic AI represents knowledge as rules, logic, or graphs and reasons over them. Its strengths are transparency and compositionality: a rule base can be inspected, a proof can be checked, and a planning system can explain the sequence it chose. It remains the basis of scheduling, verification, and business-rule engines. Its weakness is fragility: a rule-based system handles only the situations its authors anticipated, and the number of rules needed to cover ordinary reality grows faster than the budget 3.

Connectionist AI, the tradition behind neural networks, learns statistical associations from large datasets. It tolerates noisy and ambiguous input and is far more effective for perception and language, at the cost of opacity: the internal representation of a large network cannot be read as a set of rules 4.

Most deployed systems combine the two, usually without advertising it. A recommendation pipeline may use a learned model for ranking, a rules layer for eligibility, and a search index for retrieval; a planning assistant may call a language model to interpret a request and a symbolic solver to schedule the result. The symbolic-versus-connectionist split describes research traditions more accurately than it describes working software.

Three inputs determine what a system can do: data, computation, and the objective it is trained against. Modern systems are fitted to datasets measured in billions of examples, trained on specialised accelerators for weeks, and evaluated against benchmarks that stand in for the tasks their builders care about 4.

Scale became a method in its own right. Empirical scaling work found that loss falls predictably as model size, data, and computation grow together, which turned architecture choices into capacity planning 5. The AI Index reports that training computation for notable models doubled roughly every five months through the mid-2020s, and that datasets doubled about every eight months 5. Later work showed that many models had been trained on too little data for their size, and that smaller models fed more text can match larger ones at lower serving cost.

The objective is the part that is easiest to get wrong. A model optimises the number it is given, not the outcome its builders intended, and the gap between the two is where many documented failures originate: a ranking system that maximises clicks, a classifier that learns a hospital's equipment rather than a disease, a chatbot rewarded for confident answers rather than correct ones 6.

Present systems are strong at pattern recognition, classification, prediction, and generation. They transcribe speech, label images, flag anomalies in industrial sensor data, rank search results, translate text, generate video, predict protein structure, and produce fluent prose. On several benchmarks they exceed typical human performance, though a benchmark score measures performance on that benchmark rather than competence in the world 5.

Large language models extended this competence to language and, to a contested degree, to reasoning. They answer questions, summarise documents, write and debug code, and follow multi-step instructions. Their errors are characteristic: confident, fluent, and sometimes untraceable to any source. The gap between fluent output and reliable judgement is the central practical problem of the current generation 6.

Systems that act — agents — extend the pattern from answering to doing. They call tools, browse, run code, and complete multi-step tasks in software environments, with reliability that falls sharply as the number of steps and the cost of error rise. Demonstrations are frequent and dependable autonomy is not, and the distance between the two is where most deployment work now happens 6.

The limits are documented. Systems degrade when the data they encounter differs from their training distribution, and they fail in ways people find surprising because the failure is not signalled. They do not reliably know what they do not know. They reproduce and amplify biases present in their training data, and they require substantial computation to build and to run 6.

Generalisation remains unsolved in the strong sense. A model that has mastered one task family rarely transfers its skill to an unrelated one without further training, and the sample efficiency of human learning remains far beyond current methods. Long-horizon planning, exact arithmetic without tools, and the tracking of many simultaneous constraints are weak relative to the fluency of the surrounding text.

Whether these are engineering obstacles or evidence that something is missing is an open and contested question. One camp treats every failure as a data or objective problem that scale and better training will reduce. Another argues that predicting text is not the same as modelling the world, and that systems trained this way cannot acquire the causal understanding that reliable reasoning requires 3. Neither position has been settled by experiment.

AI has become a capital-intensive industry. Stanford's AI Index recorded 109.1 billion US dollars of private AI investment in the United States in 2024, nearly twelve times the figure for China, with 33.9 billion dollars of generative-AI investment worldwide 5. It also recorded that 78 per cent of organisations reported using AI in 2024, up from 55 per cent a year earlier — a measure of adoption that includes experiments as well as production systems.

Prices fell even faster than capabilities rose. The Index estimates that the inference cost of a system performing at the level of GPT-3.5 fell by more than 280-fold between November 2022 and October 2024, while hardware costs declined and energy efficiency improved 5. Cheap inference is why AI features spread into ordinary software, and why the binding constraint on deployment has shifted from model quality to integration, evaluation, and liability.

The industry is concentrated. Training frontier models requires accelerators, data centres, and power contracts that few organisations can assemble, and the largest laboratories sit inside companies with existing cloud and distribution businesses 5. Open-weight releases have narrowed but not closed the performance gap, shifting competition toward efficiency, distribution, and trust rather than raw scale alone.

Governance moved from commentary to law within a decade. The OECD's AI Principles, adopted in 2019 and revised since, set out expectations for human-centred and trustworthy AI: inclusive growth, human rights, transparency, robustness, and accountability 7. They are not binding, but they supplied much of the vocabulary that later instruments borrowed.

The US National Institute of Standards and Technology published a voluntary risk-management framework in January 2023, organised around governing, mapping, measuring, and managing risk 6. A generative-AI profile followed in July 2024, and the framework is being revised under the 2025 White House AI Action Plan, which shifted federal emphasis toward adoption and infrastructure rather than restriction 6.

The European Union's AI Act, adopted in 2024, is the first comprehensive binding regime. It prohibits a short list of practices, imposes documentation and conformity-assessment duties on high-risk systems, requires transparency for systems that interact with people or generate synthetic content, and adds obligations for general-purpose models above a compute threshold 8. Prohibitions applied from February 2025, general-purpose model duties from August 2025, and most high-risk obligations from 2026 and 2027.

Elsewhere the pattern is fragmentary: sectoral regulators applying existing law, national strategies without enforcement, and a growing number of state and provincial statutes. The result is a patchwork in which the same system can be lawful in one market and restricted in another 8.

Criticism of AI divides into several arguments that are often bundled together. The first concerns present harms: bias in decisions about credit, hiring, and policing; surveillance and biometric identification; the generation of misinformation at scale; and the labour displaced by automation 6. These are documented in incident databases and regulatory findings rather than hypothetical.

The second concerns political economy. Training data is collected largely without consent or payment, model development is concentrated among a handful of firms, and the returns accrue to whoever owns the computation and the distribution channel. Copyright litigation in several jurisdictions has tested whether training on published work is lawful, and the central questions remain unresolved 8.

The third is philosophical. If a system manipulates symbols without understanding them, does it think? The Chinese room argument holds that symbol manipulation alone cannot produce meaning, and the debate about whether large models understand anything has not moved far beyond that question 3. A related argument holds that fluent prediction is a shortcut that mimics explanation without providing it.

The fourth concerns risk from more capable systems: loss of human control, concentration of power, and the difficulty of supervising systems whose internal representations are not legible. Researchers disagree about how urgent that risk is relative to present harms, and the disagreement is partly empirical and partly about which harms deserve priority 7. Differing accounts also diverge on economics: some treat AI as a general-purpose technology comparable to electricity, others as a narrower automation of specific tasks whose productivity effects will be uneven and slow 5.