Artificial general intelligence

Artificial general intelligence, or AGI, is the name given to a hypothetical system that matches or exceeds human performance across the full range of cognitive tasks. The term is used as a research goal, a marketing category, a policy object, and a warning. It has no agreed definition and no accepted test 1.

The ambiguity is not incidental. A definition that emphasises economic output implies different milestones from one that emphasises generality, autonomy, or learning efficiency, and the choice determines whether a system counts as AGI and whether it is near. Laboratories have their own formulations, and the differences are often substantive rather than semantic.

This entry covers the definitions, the history of the idea, the evidence that language models changed the debate, the problem of measurement, the timeline dispute, and the political uses of the term.

The most common informal definition is human-level performance across most economically valuable work. It appears in corporate charters, policy papers, and journalism, and it is vague in every clause: which humans, which work, and how much of it 2.

A more formal tradition defines intelligence as the ability to achieve goals in a wide range of environments. Legg and Hutter formalised this by weighting performance across environments by their complexity, producing a measure that is precise but not computable in practice 3. The definition captures generality as a property of the distribution of tasks rather than a checklist of skills.

Morris and colleagues proposed a framework with three axes: performance, generality, and autonomy. Performance runs from emerging through competent, expert, and virtuoso to superhuman; generality runs from narrow to general; autonomy runs from assistant to fully independent. AGI, on this account, is a region of that space rather than a threshold, and the framework is explicitly a tool for comparison rather than a definition 1.

The distinctions matter because they separate questions that are often conflated. A system can be expert at one task and narrow; it can be general and unreliable; it can be capable and dependent on a human operator. Claims about AGI usually slide between these axes without saying which one is meant.

The ambition is as old as the field. Turing's 1950 paper asked whether machines could think and proposed an imitation game as a test. The Dartmouth workshop of 1956 named artificial intelligence and assumed that a machine able to manipulate symbols could, in principle, do anything a mind does.

Decades of work on search, logic, expert systems, and planning produced impressive narrow systems and repeated disappointment. The label general intelligence emerged partly in reaction to that pattern: researchers who wanted to distinguish the original ambition from task-specific engineering began to speak of artificial general intelligence, and the phrase was popularised in the 2000s by the community around the Artificial General Intelligence conference 1.

The deep learning era changed the terms of the argument. Systems trained on large datasets began to perform well across domains without hand-coded rules, and language models in particular displayed a breadth that narrow systems had not. By the early 2020s, the question had shifted from whether general-purpose learning was possible to how general the resulting systems really are 4.

Large language models are trained on a single objective, next-token prediction, over a corpus that spans many domains. The resulting systems can translate, summarise, write code, answer questions, and follow instructions without task-specific training. That breadth is the strongest empirical argument that general capability is continuous with ordinary machine learning rather than a separate breakthrough.

It is also the weakest. Breadth on text tasks is not the same as generality across environments, and a model that performs well on an exam may fail at the same task when it is embedded in a workflow. The claim that a model shows sparks of general intelligence was offered as an observation, not a measurement, and it was contested from the moment it appeared 4.

Two counter-arguments are prominent. The first is that language models learn from human-generated text, so their apparent generality is a reflection of the distribution they were trained on rather than an ability to acquire new skills. The second is that understanding requires more than producing appropriate outputs, and that behavioural competence is not evidence of the mechanisms that underlie comprehension 5.

Neither argument settles the question. They do establish that benchmark breadth is not the same as generality, and that the word intelligence is doing more work in these discussions than the evidence supports.

There is no accepted test for AGI. The Turing test was designed as a thought experiment about imitation, not as a benchmark, and it has been passed in narrow settings by systems that are not generally intelligent. Exam-style benchmarks measure knowledge and pattern recognition more than the ability to act in novel situations.

Chollet argues that intelligence should be measured as skill-acquisition efficiency: how quickly a system learns a new task from a small number of examples, rather than how much it already knows. The benchmark he designed for this purpose consists of novel visual puzzles intended to be resistant to memorisation, and it has become a reference point for measuring generalisation 6.

Other proposals focus on economic tasks, long-horizon autonomy, or the ability to conduct research. Each captures something real and each can be gamed. An economic definition depends on which jobs are counted and how productivity is measured; a long-horizon definition depends on the task suite; a research definition depends on the judgement of experts.

The measurement problem is not merely technical. A test encodes a theory of what intelligence is, and the choice of test determines which systems appear close to the goal. That is why the field has not converged on one, and why claims of AGI are usually claims about a particular evaluation rather than about the thing itself 1.

Forecasts vary by decades. In a 2023 survey of AI researchers, the aggregate forecast gave a 50 per cent chance of high-level machine intelligence by 2047, thirteen years earlier than the same survey's 2022 result, with substantial dispersion across respondents 7. Some respondents gave dates within a few years; others gave centuries or declined to answer.

The disagreement tracks different models of progress. One view extrapolates from scaling: capability has improved predictably with compute and data, the improvements have not yet plateaued, and the remaining gaps are engineering problems. Another view treats intelligence as requiring capacities that current methods do not address, such as causal reasoning, persistent memory, and embodied interaction, and expects progress to be slower and less continuous.

Institutional claims are not forecasts in the same sense. A laboratory that says AGI is near is also making a case for investment, talent, and regulatory forbearance. A laboratory that says it is far away is making a case for patience. Both can be sincere, and both are strategically useful, which is a reason to read public timelines alongside the evidence rather than instead of it.

The one thing the forecasts agree on is uncertainty. The distributions are wide, the questions are ambiguous, and the track record of prediction in this field is poor 8.

As an economic concept, AGI implies a system that can substitute for human labour across most tasks. That would be a different kind of technology from the ones that automate specific tasks, and the policy questions it raises — redistribution, taxation, labour market adjustment, competition — are correspondingly larger. Whether such a system is possible is the prior question, and it is unresolved.

As a policy object, AGI is awkward because it does not map onto existing legal categories. The European Union's AI Act regulates general-purpose AI models through transparency and systemic-risk obligations, but it does not legislate for AGI as a threshold. National security frameworks treat frontier capability as a strategic asset. International reports summarise the evidence without asserting a timeline 8.

The term also functions as a coordination device. A shared goal gives researchers a common reference, funders a narrative, and regulators a symbol. It can also distort: resources shift toward systems that score well on generality tests, and the safety debate becomes a debate about a hypothetical rather than about the systems already deployed 2.

A recurring proposal is to regulate capabilities rather than labels: evaluate what a system can do, require reporting above thresholds, and impose obligations that scale with demonstrated capability. That approach does not require agreement on what AGI means, which is its main advantage, and it requires evaluation that does not yet exist, which is its main difficulty.

The strongest critique is that general intelligence is not a single quantity. Human cognition is modular, uneven, and context-dependent; a system can be excellent at mathematics and helpless in a kitchen. Treating intelligence as a scalar invites comparisons that flatter whatever the test measures, and the history of IQ testing is a cautionary example of what happens when a scalar is treated as a natural kind 1.

A second critique is that the term is a moving target. Systems that pass one threshold are redefined as narrow, and the goalposts shift to whatever remains unachieved. This makes AGI unfalsifiable in practice: no result can count as success because the definition can always be tightened.

A third is that the concept centres a particular kind of cognition. Language and abstraction are emphasised; perception, motor control, social understanding, and emotional attunement are treated as peripheral or as prerequisites that will come later. Embodied cognition researchers argue that this ordering gets the problem backwards.

A fourth is political. Critics argue that AGI talk concentrates attention and capital on a few laboratories, frames policy around a hypothetical future, and distracts from measurable harms in deployed systems. Defenders reply that a long-term goal is necessary for research direction and that the same systems raise both kinds of concern.

On feasibility, the spectrum runs from inevitability to impossibility. Believers point to the absence of a known obstacle: nothing in physics forbids general learning, and the trend of the past decade is toward broader capability. Sceptics point to the absence of a known mechanism: current systems lack persistent memory, reliable planning, and causal models, and no one has demonstrated how to build them 6.

On timelines, the disagreement is even wider. Survey medians around mid-century coexist with claims of a few years and claims of a century. The gap reflects different assumptions about scaling, data, algorithms, and the difficulty of the remaining problems, and neither side has a decisive argument 7.

On whether language models are the path, opinion divides between those who see them as the first general technology and those who see them as sophisticated interpolation over human text. The former note that tool use, multimodal input, and reinforcement learning extend the base model into action; the latter note that these extensions are engineered around the model rather than learned by it 5.

On risk, the disagreement is about both probability and priority. Some researchers treat AGI as the most important safety problem because a system more capable than its overseers cannot be controlled by supervision alone. Others treat that scenario as speculative and argue that present harms deserve the attention. Both positions are represented in the public reports, and neither has been settled by evidence 8.