AGI is an argument about a definition, not a thing anyone has built and is keeping quiet about. The table below is the five definitions people actually argue from, who wrote each one, and what a system would have to do to satisfy it.
| Definition | Who and when | What it says | What would satisfy it | What it leaves out |
|---|---|---|---|---|
| The imitation game | Alan Turing, Mind, 1950 | If a judge in a text conversation cannot reliably tell the machine from the person, stop asking whether it thinks | A judged conversation, with the judge, the duration and the pass rate written down in advance | Says nothing about doing anything outside a chat window |
| Universal intelligence | Shane Legg and Marcus Hutter, Minds and Machines, 2007 | Intelligence is expected performance across all computable environments, weighted by simplicity | A formal score over a distribution of tasks | The measure is not computable, so it guides rather than decides |
| Economically valuable work | OpenAI Charter, April 2018 | Highly autonomous systems that outperform humans at most economically valuable work | A list of jobs, a definition of "most", and a measured comparison against people doing them | Never fixes the list or the threshold |
| Skill-acquisition efficiency | François Chollet, arXiv:1911.01547, 2019 | Intelligence is how efficiently a system turns experience into skill at tasks it has not seen | A test of novel tasks with prior knowledge and training exposure controlled, which is what ARC is | Deliberately refuses to score skill on its own |
| Performance × generality | Morris et al., Google DeepMind, arXiv:2311.02462, 2023 | Two axes: how well, against a human percentile, and over how wide a range of tasks | Reaching a named percentile of skilled adults across a broad, agreed task set | The task set itself is left to be written |
The five disagree about where the difficulty lies, which is why an argument about AGI is usually two people using two of these rows without saying which. Turing moves the question to behaviour and refuses to define thinking. Legg and Hutter give a formal measure and accept that it cannot be computed. The OpenAI charter defines the target by its consequences rather than its capabilities. Chollet argues that scoring skill measures the training data, not the system, and that efficiency at acquiring NEW skill is the only thing worth measuring. DeepMind's paper is the one that makes the question answerable, by splitting it into two axes that can each be measured and then insisting that both be reported. If a conversation about AGI is going nowhere, asking which of these five rows the other person is using usually ends it.
Why there is no single definition
Legg and Hutter collected more than seventy published definitions of intelligence before proposing their own, and the spread was not noise: psychologists, economists and computer scientists were each defining it for their own purpose. The same is true of the G.
"General" is used for at least three different claims, and they are not the same claim:
- General as in broad. It does many different things. A system that answers physics questions, writes code and reads a chest X-ray is general in this sense, and says nothing about how well it does any of them.
- General as in transferable. It does something it was never shown. This is Chollet's sense, and the reason ARC's tasks are built to be unlike anything in a training set.
- General as in autonomous. It decides what to do next without a person in the loop. DeepMind's paper treats this as a separate axis entirely, with six levels of its own, precisely because a highly capable system under tight human control and a moderately capable one running unsupervised are different situations with different risks.
A claim that a system "is AGI" is therefore not one claim. It is a claim about breadth, a claim about transfer, and a claim about autonomy, and it is usually made with evidence for one of the three.
This is the shortest honest route from a claim to a measurement, and it has exactly two questions in it. The first is which task set, because a task set chosen after the results are in can be made to support anything; a fixed, published set with a held-out portion cannot. The second is which human baseline, because "better than a human" is meaningless until the human is specified. DeepMind's 2023 paper does that by percentile of skilled adults, and the three endpoints here are its own three named rungs. A claim that survives both questions is a claim worth checking. A claim that fails the first one is not wrong so much as untestable, which is worse.
The ladder that makes the question answerable
DeepMind's 2023 paper is the most useful thing in this area for a reader who wants to stop arguing and start measuring, because it turns one yes-or-no question into a grid. One axis is performance, as a percentile of skilled adults. The other is generality: narrow, meaning a clearly scoped task, or general, meaning a wide range of non-physical tasks including learning new ones.
| Level | Performance bar | Narrow, as the paper placed it in 2023 | General, as the paper placed it in 2023 |
|---|---|---|---|
| 0 — No AI | None | A calculator, a compiler | Human-in-the-loop work such as Mechanical Turk |
| 1 — Emerging | Equal to or somewhat better than an unskilled human | Rule-based systems such as SHRDLU | Frontier chat models of 2023 |
| 2 — Competent | At least the 50th percentile of skilled adults | Voice assistants, toxicity classifiers | Not achieved |
| 3 — Expert | At least the 90th percentile | Grammar checkers, image generators | Not achieved |
| 4 — Virtuoso | At least the 99th percentile | Deep Blue, AlphaGo | Not achieved |
| 5 — Superhuman | Better than every human | AlphaFold, AlphaZero, Stockfish | Not achieved |
The column that matters is the right-hand one, and in the paper's own 2023 assessment every cell in it below the first two was empty. That is the whole argument in one table: narrow systems have been at level 5 since AlphaGo, and the general column had not been credited above level 1. The placements are the authors' judgements at the time of writing, not measurements, and they are the part of the table that dates fastest — the LEVELS are the durable contribution, because they say exactly what evidence would move a system up a row. Anyone claiming a system has reached level 2 general is claiming it beats half of skilled adults across a broad task set, which is a measurable claim and should come with the task set attached.
What is measured instead, while the definition is argued about
Because no agreed general task set exists, progress is reported on benchmarks, which are narrow by construction. A benchmark score is evidence about one column of one row of the grid above, and it is routinely quoted as if it were evidence about the whole grid. The specific failure modes — saturation, contamination, mismatched human baselines, compute not being reported — are the subject of the sibling page, AGI benchmarks explained.
One measurement does try to cross the gap. METR's March 2025 report measured the length of task a system can complete rather than the difficulty of a question, using the time a human expert takes as the unit, and reported the length at which a system succeeds half the time. The headline finding was that this horizon had been roughly doubling every seven months. It is worth knowing for two reasons: it measures something a person can picture, and it is an extrapolation from a small number of points across changing task suites, which its own authors say plainly.
Costs, mistakes and failure modes
Moving the goalposts, in both directions. Every capability that arrives stops being called intelligence, which is old enough to have a name in the field. The opposite error is as common: treating any newly passed benchmark as the definition being satisfied. Both are avoided the same way, by naming the definition and the threshold before the result.
Arguing about the word instead of the claim. "Is it AGI" cannot be settled. "Does it beat the 50th percentile of skilled adults on this published task set" can be, and it is usually the question that was actually meant.
Quoting a percentile with no population. "Better than a human" is unfalsifiable. "At least the 90th percentile of skilled adults at this task" is a claim, and it needs the population written down.
Treating a capability claim as a deployment claim. DeepMind's paper separates capability from autonomy for this reason. A system that COULD do a job and a system that is DOING it unsupervised are different facts, and the risks attached to each are different.
Key Takeaways
- There is no single definition of AGI. The five in the table above are the ones people argue from; Legg and Hutter counted more than seventy definitions of intelligence alone before adding theirs in 2007.
- "General" carries three separate claims — broad, transferable, autonomous — and evidence for one is usually offered for all three.
- DeepMind's 2023 grid is the most testable framing available: performance as a percentile of skilled adults (50th, 90th, 99th), crossed with narrow or general.
- By the paper's own 2023 placement, narrow systems had reached level 5 and the general column was credited no higher than level 1.
- The Turing test is from 1950 and specifies no duration, no judge and no pass rate; every version that has been run added its own.
- METR's March 2025 report measured task LENGTH rather than question difficulty and found the 50-per-cent horizon roughly doubling every seven months — an extrapolation its authors label as one.
- A claim worth checking names the task set before the result and the human baseline as a percentile. A claim that does neither is not testable.