Why AGI Had No Definition, and What Borrowing One Costs

Artificial general intelligence appears in company mission statements, in contracts, in national policy documents, and in arguments about whether the current trajectory is dangerous. It is doing a great deal of work.

And until recently there was no operational definition anyone agreed on. Ask ten researchers and you get answers ranging from a system that can do any task a human can, to a system that can do most economically valuable work, to a system that can improve itself, to a system that would pass as human in unrestricted conversation.

Those are not variations on one idea. They pick out different systems, arrive at different times, and imply different policies.

The practical consequence is that claims about AGI cannot be checked. If someone says a system is close to AGI, or that AGI is decades away, there is no measurement that would settle it, so the disagreement is unfalsifiable in both directions.

This is the gap a paper published in October 2025, with Dan Hendrycks, Dawn Song, Christian Szegedy, Yarin Gal, Erik Brynjolfsson and more than thirty other authors, sets out to close.

The field has thousands of benchmarks, so it is worth being clear about why none of them answered this.

Benchmarks measure task performance. A model scores well on a set of exam questions, coding problems, or reading comprehension items. That tells you about the tasks in the set, and generalising from it requires an assumption that the set represents something broader.

Three problems follow. Benchmarks saturate: once models score near the ceiling, the benchmark stops distinguishing anything and a new one is needed, so the yardstick keeps changing. They are contaminable: a benchmark published on the internet may be in the training data, and a high score may measure recall. And they are unrepresentative by construction, since a benchmark contains what was easy to collect and grade rather than a sample of what intelligence involves.

Most importantly, benchmarks have no theory of coverage. A model that scores well on twenty benchmarks might still have a large hole, because nobody built a benchmark for the thing in the hole.

That last problem is the one the paper attacks. It wants a decomposition of intelligence that is complete in some defensible sense, and then measurement against each part.

The paper's definition is a single sentence. Artificial general intelligence is a system matching the cognitive versatility and proficiency of a well-educated adult.

Both halves are doing work.

Versatility means breadth. Not excellence at one thing but competence across the range of cognitive tasks a person handles. This rules out a system that is superhuman at mathematics and unable to hold a conversation.

Proficiency means a level, and the level is deliberately modest. Not the best human in each domain, not a genius, but a well-educated adult. Something a great many people already are.

Choosing a modest target is a defensible move rather than a lowering of the bar. A definition pinned to the best human in every field would be satisfied by nothing for a long time and would tell you little along the way. A definition pinned to an ordinary competent person gives a threshold that is meaningful, reachable in principle, and measurable, because psychology has spent a century measuring exactly that population.

That last point is what makes the rest of the paper possible.

Having anchored to human cognition, the paper needs a decomposition of it, and rather than inventing one it takes the one psychology already converged on.

Cattell-Horn-Carroll theory is the product of roughly a century of a specific empirical exercise. Give large populations many different cognitive tests. Look at which scores correlate. Use factor analysis to find the smallest set of underlying abilities that explains the pattern of correlations. Repeat, refine, and merge competing models.

The result is a hierarchical structure of broad abilities that has been replicated across populations and test batteries, and the paper describes it as the most empirically validated model of human cognition available.

The attraction for this purpose is that the decomposition was not chosen by the authors. It emerged from data about how human abilities actually cluster, which is what gives the coverage claim its force. If those broad abilities exhaust what human cognitive test performance varies on, then measuring all of them is a defensible definition of measuring general intelligence.

That is the paper's central methodological bet, and it is also where the criticism lands.

The move from a human framework to a machine assessment involves an inference worth making explicit, because it is easy to miss.

The framework was derived from correlations among human test scores. It says that in human populations, performance on many tests can be explained by a smaller number of broad abilities, because those abilities vary together across people in a stable way.

Applying it to machines assumes that the same decomposition carves machine cognition at the same joints. That is not obviously true, and there is reason to doubt it. The factor structure exists because of how human brains vary, shaped by shared neural machinery, development and genetics. A transformer has none of those constraints, so there is no particular reason its abilities should cluster the same way.

The paper has a partial answer, and it is a good one: even if the factors do not carve machine cognition naturally, the domains still enumerate things humans can do, and a system claiming human-level general capability has to do them.

So the framework may be better read as a coverage checklist derived from human abilities than as a theory of machine cognition. That reading is weaker, and it survives the objection.

Two design decisions turn the framework from a description into something that produces a number.

The first is exhaustiveness. Every broad ability in the framework must be measured. You cannot score highly by being excellent at the ones you happen to be good at, because the missing ones are counted as zero rather than omitted. This is what makes it a versatility measure rather than a capability measure.

The second is equal weighting. Each of the ten domains contributes ten percent, and the authors are explicit that this is to prioritise breadth. A system superb at four domains and absent on six scores forty percent, no matter how superb.

That second decision is a value judgment rather than an empirical finding, and the paper does not pretend otherwise. There is no fact about the world establishing that auditory processing matters exactly as much as mathematical ability for general intelligence. One could argue the weights should reflect economic value, or frequency of use, or difficulty.

Equal weighting is the choice that best serves the definition, since versatility is the thing being measured. It is a choice, and reading the score requires remembering that it was made.

Before the results, it is worth asking what a definition like this is actually supposed to achieve, because judging it against the wrong purpose produces bad criticism.

It is not supposed to settle what intelligence is. That is a philosophical question the paper does not claim to close.

It is not supposed to predict when AGI arrives. It provides a measuring stick, and a measuring stick makes no forecast.

What it is for is making claims checkable. Once a definition and a procedure exist, someone saying a system is nearly AGI can be asked which domains it scores well on and which it does not. The conversation moves from assertion to evidence, and the number can be disputed on method rather than on intuition.

A secondary purpose is direction. A framework that scores every domain shows which are weak, and weak domains are research targets. The paper's headline empirical finding is exactly of this kind, and it is more interesting than the total.

That is the next lesson: the ten domains, the scores, and the one that comes out at essentially zero.

One habit worth carrying into the next lesson, since this paper is a preprint and preprints move.

This one was submitted in October 2025 and revised in December. The headline figures shifted slightly between versions, and the per-domain breakdown in the first version sums to a total a percentage point away from the abstract's figure.

That is entirely normal for a preprint and it is worth noticing, because a number cited without a version is a number that may not be in the paper you find.

The general habit is to cite the version, and to treat exact figures from any preprint as provisional while treating the structure as the durable part. Here the structure is the definition, the ten domains, and the equal weighting. Those have not moved. The precise percentage for a given model is the part most likely to change, and also the part most likely to be quoted in a headline.

More than thirty authors on a paper is a signal that its framing has broad buy-in. It is not a signal that every number in it is final.