AGI: Definition, measurement, and research

AGI, or artificial general intelligence, refers to a system that can learn and solve a wide variety of problems, including unfamiliar ones, by transferring what it has learned across situations. Its definition and the threshold for demonstrating it remain disputed: strong performance on an exam, in a conversation, or on a benchmark does not establish this general capability.

What “general” means

Specialized intelligence handles a defined family of tasks. Assessing general intelligence also requires examining how a system acquires skills when the rules, data, or goals change. The question therefore goes beyond how many features an interface offers.

Consider a simple example: a program can organize files by their extensions. You then ask it to infer an unfamiliar filing system from examples and adapt that system when an exception appears. AGI research is concerned with how it learns this new organization, what information it needs, and whether it can reuse what it has understood.

Generality does not necessarily mean being able to do everything immediately. It can mean an ability to learn across a broad range of situations with limited resources. An evaluation must therefore distinguish skills a system already has from its ability to acquire new ones.

The paper Levels of AGI for Operationalizing Progress on the Path to AGI, first published in November 2023 and revised in September 2025, proposes separating depth of performance from breadth of capability. It also treats the autonomy granted to a system as a separate dimension. This proposal provides a research vocabulary; it is not a universal certification standard.

Our artificial general intelligence hub brings together pages covering the different dimensions of this question. Before accepting a claim about AGI, identify the definition behind it.

Concepts that need to stay distinct

AGI, language models, agents, and superintelligence describe different things. Treating them as interchangeable can lead you to attribute capabilities to a system that its behavior has not demonstrated.

A large language model is a category of model. An agent is a system that selects and executes actions in an environment. AGI describes an ambition for general capabilities; superintelligence adds the idea of performance far beyond human abilities.

Concept	What it describes	What to examine	What it does not establish
Language model	A family of models that process language, among other inputs	Response quality, transfer, errors	A general ability to learn
AI agent	A system organized to carry out sequences of actions	Action selection, verification, recovery from errors	General intelligence
AGI	A general ability to learn and solve problems	Breadth, adaptation, learning efficiency	Consciousness or intentions of its own
Superintelligence	Capabilities far beyond those of humans	Relevant domains and comparison conditions	Beneficial behavior

These distinctions have a practical consequence: adding memory and tools to a model can improve the system without demonstrating general intelligence. Conversely, a system may solve a difficult problem while still requiring human approval before every action.

Consciousness is another separate question. The tests discussed here observe behavior and results; they provide no access to any possible subjective experience.

To understand the action loop without confusing it with a measure of intelligence, see AI agent or chatbot: What is the difference?. An execution architecture and a level of capability need separate assessments.

Measuring intelligence rather than accumulated skills

In his essay On the Measure of Intelligence, published in November 2019, François Chollet argues for distinguishing skill from intelligence. Performance can depend heavily on prior knowledge and training experience. It therefore does not, by itself, reveal a system’s ability to learn.

He proposes understanding intelligence in terms of the efficiency of skill acquisition. This approach accounts for the scope of the problems, their generalization difficulty, initial knowledge, and available experience. The aim is to better isolate the ability to handle unfamiliar situations.

To interpret a result from this perspective, examine:

the information already available to the system;
the additional examples it needs to learn;
the changes introduced into the task;
the resources it consumes to succeed;
its ability to transfer learning to a different problem.

Consider a hypothetical document classification task. A system correctly recognizes familiar forms, then encounters an unfamiliar layout. Its initial success remains useful, but it does not answer the next question: can it discover the new rules from a few examples and apply them correctly?

This distinction helps shape an evaluation. Keep the underlying problem, change its presentation, and observe the errors and corrections required. The test should avoid becoming a matter of recognizing a familiar format.

Our page on generalization and skill acquisition explores this approach in more detail. It explains why the amount of knowledge a system draws on and its ability to adapt deserve different measurements.

What makes an evaluation useful

An AGI evaluation needs to state what it is trying to establish. Success across varied tasks remains too vague without specifying the domains, expected performance, and permitted assistance.

Start by defining the scope. A collection of math exercises can provide evidence about certain mathematical abilities, but it does not directly measure learning to use unfamiliar software or interacting with a physical environment. A sound conclusion stays within the bounds of the tasks observed.

Next, define what counts as new. A previously unseen question may closely reproduce a familiar pattern. Conversely, a visually similar presentation may require a different rule. Generalization must therefore be assessed through the transformations required, rather than simply the absence of duplicate text.

The conditions for success also need to be explicit. Does the system receive corrections? Can it consult documentation, run programs, or try again? Is the evaluation scoring its first response or the best result from several attempts?

Comparisons with humans require the same care. Specialists, beginners, and participants trained on the test are different reference groups. Any differences in their instructions and available resources need to be documented.

Finally, an evaluation must retain the failures. A system that sometimes produces an excellent solution may still be unreliable if there is no way to identify its incorrect answers. In real use, the ability to detect an error and ask for help matters alongside average success.

These criteria do not amount to an official definition. They make a claim testable and allow its limits to be discussed on evidence rather than general impressions.

ARC-AGI: Testing adaptation to unfamiliar problems

ARC-AGI offers an example of this approach. The official ARC-AGI-1 overview describes grid problems in which participants infer a transformation from examples, then apply it to a new input. The test targets the acquisition of a rule from limited information.

The format avoids making every problem depend on specialized knowledge. It nevertheless reflects choices about objects, space, and the patterns participants are expected to recognize. Those choices matter when interpreting results.

As of October 2026, static tasks also need to be distinguished from interactive environments. The official ARC-AGI-3 overview describes environments where an agent explores, discovers how things work and what the objectives are, and adapts its actions without natural-language instructions. The evaluation thus examines learning that happens through interaction.

The methodological difference matters. In a static problem, all supplied observations can be examined before answering. In an interactive environment, the system must choose its experiments: an action also serves to obtain useful information.

Success on these tests would still not cover every possible aspect of AGI. Language proficiency, social skills, and professional reliability would require other evaluations.

The ARC-AGI reference page explains the formats and their limitations. The benchmark’s name identifies a research program; it should not be read as a status automatically awarded to its strongest participants.

Why a high score is not enough

A score summarizes results under a particular protocol. Understanding it requires examining what that aggregate includes and what it leaves out.

The first challenge is separating training from evaluation. If questions or solutions were used to prepare the system, the result may reflect familiarity with the test. Even without direct access to answers, repeated tuning on the same exercises can encourage adaptation to the benchmark.

Robustness requires a different kind of check. The GSM-Symbolic study, first published in October 2024, examines controlled variations of math problems. It reports performance changes when certain elements of the questions change. Those observations apply to the models and conditions studied; they do not automatically describe systems available in October 2026.

The principle remains useful for test design: changing numbers, removing a superficial clue, or rephrasing an instruction can reveal whether success is stable. Any decline then needs analysis, because the changes may differ in difficulty.

Success also needs to be separated from cost. A method may improve its result by increasing the number of attempts, candidate programs, or tool calls. That can be useful progress, but a comparison must account for those resources.

Finally, an average can conceal an entire family of failures. Examine results by task type, unresolved cases, and conditions in which the system gives incorrect answers with confidence.

The article How to read model benchmark rankings offers a complementary framework. A ranking orders results; it does not replace an examination of the protocol.

What the evidence establishes in October 2026

As of October 24, 2026, the sources reviewed show active research on adaptation, efficiency, and evaluation conditions. They do not provide a universal protocol for certifying the full range of capabilities generally associated with AGI.

The official ARC Prize leaderboard, consulted on that date, distinguishes reasoning configurations, competition systems, and the supporting setups around models. It also relates performance to resources consumed. These distinctions prevent every score change from being attributed to the model alone.

The execution setup may retain notes, pass along interaction history, or preserve a working state. When these conditions change, the system being evaluated changes too. A serious comparison therefore needs to describe the complete configuration and its budget.

This calls for precise conclusions: a system solved certain tasks, under certain rules, using certain resources. Moving from that observation to a claim of general intelligence requires further evidence about the breadth and transfer of its capabilities.

The absence of universal certification does not mean progress cannot be established. Researchers can measure better performance on an unfamiliar game, faster adaptation, or fewer required interventions. These results are informative when they can be reproduced.

To follow the research, prioritize publications that provide protocols, execution records, and explicit limitations. A selected demonstration illustrates a possibility; a systematic evaluation makes it possible to examine how often it occurs and under what conditions.

Research directions: Learning, searching, and reuse

Several mechanisms can contribute to learning new tasks. Keeping them distinct helps clarify what an experiment actually demonstrates.

A system can learn during training, draw on examples supplied in its context, or search for a solution during execution. It can also test a hypothesis with a tool, observe a failure, and try again. These mechanisms use different information and resources.

One direction involves searching for programs that explain the observed examples. A candidate program can then be executed to check its predictions. The benefit is an explicit representation of the proposed rule.

The DreamCoder paper, submitted in June 2020, presents a system combining program search, neural guidance, and the learning of reusable abstractions. It illustrates how components acquired on some problems can help solve others. The result applies to the experimental domains studied; it does not establish an unlimited general capability.

The search problem remains demanding: different rules can explain the same examples. The system must select hypotheses, organize attempts, and keep search costs manageable. Being able to verify a program does not guarantee that the system will quickly find the right one.

Another question concerns retaining what has been learned. A system may succeed by using its immediate history without being able to reuse the discovery in another session. Testing that transfer requires a separate experiment.

These approaches need to be compared on new tasks with explicit budgets. A promising architecture remains a hypothesis to evaluate; its underlying principle alone cannot establish AGI.

How to assess an AGI announcement

Read an announcement as an experimental claim. Start by restating it in observable terms: which tasks does the system learn, with what assistance, and within what limits?

Then look for the evidence needed to check the conclusion:

The definition used. Does the publication target an ability to learn, coverage across tasks, or a level of performance?
The complete system. Which tools, memories, instructions, and human interventions contribute to the result?
The evaluation data. Are the tasks public, new, protected from access, or selected after the trials?
The budget. Are failed attempts and additional computation included?
Verification. Do the protocols, execution records, and results allow independent reproduction?

Next, distinguish what was observed from what was extrapolated. A system that learns an unfamiliar environment provides evidence about that environment. Claiming that it could learn any profession or scientific discipline adds a hypothesis whose supporting evidence needs scrutiny.

Timelines deserve the same treatment. A forecast depends on the chosen threshold and assumptions about future progress. An announced date replaces neither a protocol nor a result.

For professional use, you can apply this discipline without waiting for the AGI debate to be settled. Define representative tasks, retain difficult cases, measure corrections, and control sensitive actions. The guide to evaluating an AI agent develops this method.

The practical question then becomes precise: does this system perform your task with acceptable reliability and cost? The answer can be useful even while its status as general intelligence remains disputed.

Key takeaways

AGI describes a general ability to learn and solve problems; the threshold for demonstrating it depends on an explicit definition.
High skill does not directly measure how efficiently a system acquires new skills.
A result needs to be reported alongside its protocol, resource use, and limits of generalization.
Tests such as ARC-AGI reveal certain capabilities; success does not cover every possible criterion for general intelligence.
To assess an announcement, examine unfamiliar tasks, failures, assistance received, and whether the results can be reproduced.

Frequently asked questions

What does AGI stand for?

AGI stands for artificial general intelligence. It describes a research goal whose precise criteria vary across publications. The acronym alone does not tell you which capabilities a researcher or organization considers necessary.

Would AGI necessarily be conscious?

Capability evaluations cannot determine whether a system has subjective experience. A machine could meet problem-solving criteria without settling the question of consciousness. These issues require separate arguments.

What score does a system need to qualify as AGI?

No universal score grants that status. A result must be interpreted within the scope of its test, including the data, permitted resources, and scoring method. Even a perfect score leaves open questions about capabilities the test does not assess.

Can we predict when AGI will arrive?

Any forecast depends on the definition being used and assumptions about future progress. It should specify the expected capabilities, the conditions for validating them, and the uncertainty involved. A deadline announced by a lab is not experimental evidence.

What is AGI? Definition, measurement, and research