What Is Artificial General Intelligence (AGI)?

What the term means

Artificial general intelligence, usually shortened to AGI, is a label for an AI system with broad competence across many kinds of intellectual work. The word “general” distinguishes the idea from a system built for a defined task, such as classifying images, translating text, recommending products, or controlling a particular industrial process. In ordinary usage, AGI is often imagined as a machine that can learn new tasks, apply knowledge in unfamiliar settings, and transfer skills between domains without requiring a separate system to be designed for each one. There is no universally accepted definition, however. Researchers, companies, policymakers, and commentators use the term with different thresholds in mind, from broad performance across existing tests to flexible learning and competence comparable to people.

AGI is not a single technical design or a product category with a settled specification. It names a goal or a proposed level of capability. A definition may emphasize the range of tasks a system can perform, how well it performs them, how independently it can learn, or whether it can adapt when the circumstances differ from its training examples. Those criteria need not point to the same answer. A system might perform exceptionally well on many tests but still need extensive human direction, struggle with unfamiliar situations, or make errors that are difficult to detect. Conversely, a system might be adaptable in a limited environment without matching people in the breadth of everyday reasoning. Claims about AGI therefore depend partly on what the speaker counts as general intelligence.

The term also has no single agreed relationship to human intelligence. Some definitions ask whether a machine can perform most economically valuable tasks at or above human levels. Others ask whether it can carry out a broad range of cognitive tasks, even if it is not humanlike in how it does so. A system need not have human feelings, consciousness, a body, or a human life history to meet every proposed definition of AGI. These are distinct questions. Capability refers to what a system can do; consciousness concerns subjective experience, for which there is no established test applicable to current AI systems. Combining them without explanation can make a technical discussion sound more settled than it is.

How it differs from specialized AI

Most deployed AI systems are designed around particular tasks or operating conditions. A spam filter classifies messages; a speech-recognition system converts audio into text; a recommendation system ranks items for a user. Each may be highly useful and technically sophisticated, but its success in one setting does not automatically mean it can handle unrelated tasks. Such systems are often called narrow AI or specialized AI. The label does not mean they are simple. It means that their capabilities are bounded by their design, training, inputs, tools, or intended use, rather than spanning an open-ended range of activities.

Modern general-purpose models complicate this distinction. A language model can generate text, summarize documents, answer questions, translate, and write some kinds of computer code through a shared interface. A model that accepts images or audio may work across still more input types. This breadth can make the system appear general, but a wide menu of functions is not by itself proof of robust general intelligence. The same underlying model may fail when instructions are ambiguous, when an answer depends on recent or unavailable information, or when a task requires reliable action over many steps. It may also produce plausible statements that are inaccurate. Evaluating breadth requires testing not only whether a model can attempt many tasks, but also how well it performs, how consistently it succeeds, and whether it can recognize its limitations.

A further distinction is between a system that can produce an answer and one that can complete a task dependably. For example, a model can suggest a plan, but executing that plan may require access to tools, permission to act, and the ability to notice and recover from mistakes. A system that performs well in a conversation may not be able to maintain progress through a long project, respond safely to changing conditions, or transfer what it learned to a genuinely new environment. Thus, descriptions such as “general-purpose,” “multimodal,” “autonomous,” and “AGI” refer to different properties. They should not be treated as interchangeable labels.

Why definitions and tests are contested

There is no official exam that certifies a system as AGI. Existing benchmarks measure specified abilities under chosen conditions: answering questions, solving problems, understanding images, writing code, or completing simulated tasks. They can help compare systems, but they cover only a sample of possible capabilities. A benchmark result does not automatically show that a system has broad competence outside the test, can perform reliably in everyday settings, or can learn new skills independently. Results may also depend on the data used to train a model, the precise prompts, the scoring procedure, and whether the system has access to external tools.

A useful evaluation therefore asks what was measured and how. Did the system encounter questions similar to examples in its training data? Was the test administered by the model’s developer or by an independent evaluator? Did a person select prompts or correct mistakes? Were results repeated across multiple tasks and conditions? Did the test measure a model’s first attempt, or a workflow involving search, code execution, and repeated revisions? These details affect what a reported score can establish. Independent replication, clear test descriptions, and comparisons with appropriate human baselines can make evidence more informative, though no single evaluation can settle every definition of general intelligence.

The analogy to human intelligence is also difficult to operationalize. People vary in expertise and performance, learn throughout life, and use language, tools, and social cooperation. A machine may exceed people on a narrow task while remaining weak at another. An average score across a collection of tests can conceal important gaps, and a high score on familiar tasks may not predict performance when stakes or conditions change. Some proposed definitions focus on a system’s ability to learn efficiently from limited experience; others focus on the number and variety of tasks it can perform. Because these measures capture different qualities, it is more precise to report them separately than to compress them into one label.

What current systems do and do not demonstrate

Large machine-learning models are trained using data and computational methods that let them detect statistical patterns and generate predictions or actions. In a language model, the central training objective commonly involves predicting text tokens from context. This process can support surprisingly flexible behavior: a model may answer questions, follow instructions, transform text, or combine patterns learned across many examples. These abilities are real and useful, but their appearance does not alone establish that a model understands the world in the same way a person does. Nor does a fluent response guarantee that the information is correct, that the model has checked it, or that it will behave consistently on a later attempt.

Many systems can be connected to tools such as web search, code interpreters, databases, or software interfaces. Tool use can improve performance by letting a model retrieve information, carry out calculations, or take actions that its internal parameters alone cannot provide. Yet tool access introduces additional questions: what sources are consulted, what actions are permitted, whether the user has authorized them, and how errors are identified. A system that completes a task with a carefully designed software workflow should be evaluated as that system, not solely as an isolated model. The distinction matters when comparing demonstrations, since the interface, tools, instructions, and human oversight all contribute to the outcome.

Reliability is another independent dimension. A system can be capable in a broad range of settings but still produce occasional serious mistakes. A useful evaluation considers how frequently errors occur, whether they are predictable, whether the system communicates uncertainty appropriately, and whether a human can detect and correct them. Safety testing examines potential harms in particular contexts, such as misuse, privacy exposure, biased outcomes, or unsafe recommendations. Safety is not established just by showing impressive capabilities, and capability is not established by the absence of observed harms in a limited test. Both require explicit evidence suited to the intended use.

Why the label matters

Debates about AGI can influence investment, public policy, education, employment discussions, and decisions about how to govern AI. The label can communicate an ambitious research goal, but it can also blur distinctions between present-day capabilities and hypothetical future systems. A claim that a model is “close to AGI” is difficult to interpret unless the speaker explains the definition, evidence, and remaining limitations. A company may use the term to describe a long-term objective; a researcher may use it as a shorthand for a particular capability threshold; a journalist may use it in a broad popular sense. These uses do not necessarily refer to the same milestone.

Practical decisions generally benefit from assessing the system that exists rather than relying on a disputed label. For a particular use, relevant questions include what task the system is intended to perform, how it was evaluated, what it does when it lacks information, what kinds of people or circumstances were included in testing, and who is accountable for its outputs. Other considerations include privacy, security, accessibility, the possibility of automation bias, and the availability of human review. These questions apply whether a system is called narrow AI, general-purpose AI, or AGI.

The label also does not settle forecasts. Experts disagree about whether progress will continue gradually, whether particular technical advances are necessary, and whether broad competence can be achieved through current approaches. Forecasts depend on assumptions about research, computing, data, evaluation, and the meaning of the destination. A date prediction is not a measurement of current capability. When discussing a forecast, it helps to distinguish what someone expects from what has already been demonstrated, and to look for the uncertainty range and assumptions behind the expectation.

A practical way to assess an AGI claim

Start by asking for an operational definition: what specific abilities would count, and what failures would show that the threshold has not been reached? Then examine the evidence. A credible account should specify the tasks, test conditions, scoring criteria, comparison points, and role of human assistance. It should distinguish a live demonstration from a reproducible evaluation and a research result from a product that has been tested in real-world conditions. If the claim relies on a benchmark, check whether the benchmark measures the ability at issue and whether independent evaluators have examined the result.

Next, consider transfer and reliability. Can the system handle genuinely unfamiliar tasks, or is it performing variations on patterns in its training and test data? Does it succeed across repeated attempts and different settings? Can it recognize when it is unsure or when information is missing? Does performance depend on extensive prompting, specialist tools, or a person monitoring each step? These are not reasons to dismiss a system’s achievement; they are information needed to describe it accurately. Results can be impressive and limited at the same time.

Finally, keep capability claims separate from claims about consciousness, agency, safety, or social effects. A system might be capable without being conscious; it might be useful without being reliable enough for unsupervised use; it might be broadly competent while still requiring rules and oversight. AGI remains a contested term rather than a universally measurable status. Careful explanations make that uncertainty explicit and describe the underlying evidence in plain language.

References

Stanford Encyclopedia of Philosophy, “Artificial Intelligence”: https://plato.stanford.edu/entries/artificial-intelligence/

National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework

OECD, Recommendation of the Council on Artificial Intelligence: https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449

Does AGI exist today?

No. Current AI systems, including large language models, are narrow: they perform well on specific tasks but do not reliably generalise the way a person can.

What is the difference between AI and AGI?

AI is the broad field of systems that perform tasks associated with intelligence. AGI is a hypothetical subset that could learn and do essentially any intellectual task a human can.

When will AGI arrive?

No one knows. Lab leaders and researchers give estimates ranging from a few years to many decades, and there is no agreed test for when it would be achieved.