What is artificial general intelligence (AGI)?

Defining artificial general intelligence

Artificial general intelligence refers to a system able to understand, learn, and apply knowledge across a wide range of tasks at a level comparable to — or exceeding — that of a human being. The defining word is general. A person who can read a contract, plan a trip, learn a new game, comfort a friend, and troubleshoot a leaking tap draws on the same underlying intelligence for all of these, and can pick up unfamiliar tasks with modest instruction. AGI names the aspiration to build software with that same breadth and adaptability, rather than a tool tuned for a single job.

This stands in contrast to the AI that exists today, usually called narrow AI or weak AI. Narrow systems can be extraordinarily capable within their domain — matching or beating people at chess, predicting protein structures, or drafting text — yet their competence does not automatically extend beyond the task they were built or trained for. A model that writes fluent prose cannot, by itself, drive a car; a champion game engine cannot file a tax return. An AGI would, in principle, take on all of these and transfer what it learns from one to another.

The term itself was popularised in the early 2000s and echoes earlier use in the late 1990s. It overlaps with older phrases such as "strong AI" and "human-level AI," although writers apply these with subtly different emphases.

"AGI," "strong AI," and "human-level AI" are related but not interchangeable. Some authors treat AGI as any system with broad, transferable competence; others reserve it for parity with skilled adults across almost all cognitive work. The absence of a shared meaning is itself part of the debate.

Narrow AI, general AI, and superintelligence

It helps to picture three reference points rather than a single dividing line.

Narrow AI is specialised. It performs particular functions — recognising faces, ranking search results, forecasting demand, generating images — and is the only kind of artificial intelligence found in deployed systems today.

Artificial general intelligence is broad. A genuine AGI would move fluidly between unrelated problems, learn new skills without being rebuilt for each one, and reason about situations it was never explicitly trained on, at roughly human competence or better.

Superintelligence sits beyond AGI. It describes a hypothetical intelligence that greatly surpasses the best human performance in virtually every field, including scientific creativity, strategic planning, and social skill. It is often discussed as a possible consequence of AGI rather than a separate research target, on the reasoning that a system able to improve itself might advance quickly.

These are categories of convenience, not precisely bounded stages. Real systems can be strong in some respects and weak in others, which makes it hard to say exactly where narrow capability ends and generality begins.

Why there is no agreed definition or test

There is no consensus definition of AGI, and no single accepted test, largely because there is no consensus on what intelligence is in the first place. Psychologists, neuroscientists, and computer scientists describe intelligence in different and sometimes incompatible ways, and each framing implies a different bar for machines.

Several distinct approaches recur. Capability definitions focus on the range of tasks a system can perform. Cognitive definitions emphasise underlying faculties such as reasoning, memory, planning, and abstraction. Economic definitions ask whether a system can do the work that people are paid to do. Comparative definitions set the target as matching a typical or an expert human.

Each approach runs into trouble. Task lists are never complete, and a system can pass many while failing others in ways that seem obviously non-human. Benchmarks that once looked like milestones — beating a grandmaster at chess, holding a fluent conversation — were reached without producing anything most observers would call general intelligence. This pattern, in which a hard problem is reclassified as "not really AI" once solved, is sometimes called the "AI effect," and it keeps any fixed finish line unstable.

Proposed tests and benchmarks

Over the years, researchers have suggested concrete tests meant to signal that a machine has reached broad, human-like competence. None is universally accepted, and passing any one of them would not settle the question, but together they show how varied the proposed criteria are.

The spread is telling. Some criteria stress conversation, some manual dexterity, some economic usefulness, and some the ability to generalise to genuinely new problems. A system might satisfy one while plainly failing another, which is one reason no single benchmark has become the accepted definition of AGI.

Where things stand in 2026

As of 2026, AGI has not been achieved, and this remains the mainstream view among researchers. The most capable systems in wide use are large "foundation" models that display broad and often surprising competence: they can write and edit text, generate and analyse images, produce working code, pass demanding professional and academic exams, and assist with research across many fields.

Whether this amounts to being "on the path" to AGI is genuinely contested. These systems remain uneven. Their performance is often described as jagged — strong on tasks that resemble their training data, brittle on small variations, prone to confident errors (sometimes called "hallucinations"), and limited in reliable long-horizon planning and consistent reasoning. A model can appear to grasp a problem while failing a simple reformulation of it.

Interpretations of the same evidence diverge. Some researchers see rapid, broad progress as the early stages of general capability; others see impressive pattern-matching that lacks ingredients they consider essential to general intelligence, such as robust reasoning, grounded understanding, or dependable transfer to unfamiliar tasks. Both readings are held by serious people, and the disagreement is not resolved by current benchmarks, several of which have been effectively saturated without ending the debate.

The timeline debate: near and far

Predictions about when, or whether, AGI will be built vary enormously, and surveys of AI researchers have repeatedly shown wide disagreement rather than a settled expectation.

Those who expect AGI relatively soon tend to argue that scaling — larger models, more data, more computation — has produced steady, general gains, and that continued scaling alongside new techniques could close the remaining gaps within years. Some leaders of frontier laboratories have publicly voiced timelines of this kind.

Those who expect AGI much later, or who doubt the current route entirely, argue that today's methods have known limitations that more scale may not fix, and that qualitatively new ideas may be required. They point to gaps in reasoning, causal understanding, data efficiency, and reliability, and note that earlier waves of AI optimism were followed by disappointment. A further group regards the question as ill-posed while "AGI" lacks a testable definition.

No forecasting method in this area has a strong track record, and confident single-date predictions — in either direction — are not well supported. The honest summary is that expert opinion spans from a few years, to many decades, to "not on the current approach," with no consensus.

Why AGI matters

AGI attracts intense attention, funding, and policy interest because the stakes attached to it — were it achieved — are unusually large. The discussion tends to fall into two broad areas.

The economic stakes concern work. A system able to perform most cognitive tasks at human level would touch a very wide range of jobs, potentially raising productivity and accelerating research while also disrupting labour markets and raising questions about how the resulting gains are distributed. Even partial progress toward general capability is already reshaping some occupations, which is why the topic matters well before any threshold is reached.

The safety stakes concern control and consequences. Researchers who study these questions ask how a highly capable, broadly competent system could be kept reliably aligned with human intentions, how misuse could be prevented, and how the power such a system confers might be concentrated or governed. Views range from those who regard advanced AI as a serious source of large-scale risk to those who consider such concerns premature or overstated relative to nearer-term harms like bias, misinformation, and misuse of existing tools.

Because definitions shape policy, the absence of an agreed meaning for AGI has practical consequences. Laws, corporate commitments, and safety frameworks increasingly refer to "AGI" or to "highly capable" systems, so how the term is defined affects what gets measured, disclosed, and regulated. That is a large part of why the seemingly abstract question — what would even count as artificial general intelligence — continues to matter.

Frequently asked questions

More from highlevel.ai

Continue with How AI works, a look at what a large language model is, an overview of where AI is heading, and a high-level history of AI.