What Would Count as AGI? Definitions That Can Be Tested
5 min read · updated August 3, 2026
Almost every public disagreement about whether AGI is close is a disagreement about what the letters stand for, conducted by people who have not said. The definitions below are all held by serious people, and they do not pick out the same event.
Why the word breaks arguments
“Artificial general intelligence” entered wide use as a way of distinguishing systems with broad competence from the narrow, single-task systems that dominated AI for decades. That contrast was useful when the alternative was a chess engine. It is much less useful now, because the interesting question is no longer general-versus- narrow but how general, at what level, measured how — and the term carries none of that.
The practical consequence is a specific and very common failure. One person says AGI is close, meaning a system that beats most humans on most text-based professional tasks. Another says it is decades away, meaning a system that can learn a new physical skill from a few demonstrations and retain it. Both may be right about their own claim. Neither statement contradicts the other, and the exchange produces nothing, because the shared word disguises the fact that two different propositions are on the table.
So the useful move is not to pick the correct definition. There is no correct one; the term is not a natural kind. The useful move is to ask of any definition: what would we observe if it were satisfied, and what would we observe if it were not?
Seven definitions people actually use
| Definition | Description |
|---|---|
| imitation | Turing's 1950 proposal: a system counts if an interrogator in text conversation cannot reliably distinguish it from a human. Turing offered it as a replacement for an unanswerable question, not as a capability bar. |
| economic | A system that can perform most economically valuable work at or above human level. OpenAI's charter uses a definition of this shape. It converts the question into an observable about labour markets rather than about cognition. |
| task battery | A fixed set of tasks or exams, with a threshold. The most operational family, and the one whose meaning depends most on the choice of tasks — which is what makes benchmark selection an argument in itself. |
| formal | Legg and Hutter's universal intelligence: expected performance across a weighted distribution of all computable environments. Fully precise, and not computable, so it functions as a definition rather than as a test. |
| efficiency | Chollet's framing in On the Measure of Intelligence: intelligence is skill-acquisition efficiency relative to priors and experience, not skill itself. A system that solves a task after unlimited exposure to similar tasks has demonstrated something, but not this. |
| graded | Morris and colleagues at Google DeepMind proposed replacing the binary with a matrix of performance level against generality, so a system can be narrow-superhuman and general-mediocre at once. This describes the actual situation better than any threshold does. |
| architectural | A system counts only if it has certain mechanisms — a world model, continual learning, grounded reference. A claim about how, not just what. It rules out passing by any behavioural test alone. |
One family is deliberately absent: definitions that route through consciousness, subjective experience or genuine understanding in a phenomenal sense. They are philosophically serious and they are not on this list because no one has proposed a procedure that settles them, which is exactly the property the next section grades on.
Grading them on falsifiability
Three questions. Is there a procedure somebody could actually run? Is the pass condition fixed in advance, or negotiated after seeing the result? Could a third party run it, without access the developer controls?
- Task battery scores best on all three, and is the most vulnerable to a different problem: a fixed public battery becomes a training target. Contamination is a documented and measurable phenomenon rather than a theoretical worry — see benchmark contamination for how it happens and how it is detected. A definition that a training run can satisfy directly is not measuring what it claims.
- Economic definitions are testable but slow. They resolve on data that arrives years after the capability does, and they are confounded by everything else that moves labour markets. Their virtue is that they cannot be satisfied by a demo.
- Efficiency definitions are testable by construction, because the measurement is of learning on tasks the system has not seen, with the prior knowledge stated up front. That is the design of the ARC benchmarks. The difficulty is proving the novelty of the tasks against a training corpus nobody outside the lab can inspect.
- The imitation game is a procedure with a moving pass mark. Its outcome depends on the interrogator, the duration, the topic and what the interrogator knows about current systems. It was proposed as a philosophical device and it does not survive being used as an engineering acceptance test.
- Formal definitions do not yield a test at all, by design. Universal intelligence is uncomputable; its value is in making explicit that any real test is a sample from a distribution of environments, and therefore that the sample is where the argument actually is.
- Architectural definitions can be falsified only with access. If AGI requires a world model, then deciding whether a system has one is an interpretability question — see world models for what probing evidence can and cannot establish. From outside the weights, the claim is not checkable.
- The graded matrix is not a test and does not try to be. It is a reporting format, and its contribution is making the question “at what level, on how wide a range” unavoidable.
The moving-goalpost complaint, examined
A recurring accusation is that the definition of AGI is revised upward every time a system passes a previous bar — chess, then Go, then professional exams, then long-horizon coding — so the target can never be reached. The observation is accurate. The inference from it is contested, and both readings are defensible.
One reading: this is ordinary scientific practice. A test is a proxy for a construct, and when a system passes the test without having the construct, the correct response is a better test, not a concession. Chess was thought to require general reasoning until a search algorithm did it without any, which was information about chess.
The other reading: an unfalsifiable target dressed as a moving standard. If every pass is reinterpreted as evidence that the test was bad, no evidence can ever count in favour, and the position has quietly stopped being empirical.
What distinguishes the two in any specific case is whether the revised test was specified before the result or after it, and whether the revision was principled — pointing at a capability the old test demonstrably failed to require — or merely harder. That is a question you can answer about a particular claim, which is why it is worth asking rather than treating the whole complaint as settled either way.
What to do before the next argument
Ask the other person which of the seven they are using. Most of the time they have not chosen, and the act of choosing ends the disagreement or converts it into a much more specific one — about whether a benchmark is contaminated, or whether labour-market data would show the effect, or whether probing evidence establishes a world model. Those are questions with methods attached.
For the same reason, treat “AGI by year X” claims as incomplete until the definition is attached. The bare year carries no information, because under the economic definition and the architectural definition the same forecaster would give very different answers. Whether such forecasts have any track record at all is the subject of AGI timeline predictions, and the reasons they are hard to make well are in why forecasting AI progress is so hard.