Understand AGI › Syllabus › Module 1
Five published definitions, side by side, and what each one would actually require before you could say it had been satisfied.
There are as many informal definitions of AGI as there are people using the term. Comparing those gets you nowhere. What is worth comparing is the small number of definitions that somebody wrote down, published and can be held to - because a written definition has a testable shape, and an opinion does not.
Five are set out below. For each one the useful question is the same: if you wanted to argue that this definition had been satisfied, what would you have to show? That question separates the definitions much more sharply than their wording does.
The best-known published definition comes from OpenAI’s charter, which describes AGI as highly autonomous systems that outperform humans at most economically valuable work.
What it would require
Evidence about work, not about cognition. You would need to identify the range of economically valuable tasks, show the system outperforming humans across most of them, and show it doing so with high autonomy - which rules out a system that only performs when a person is steering it.
What is attractive about it
It is indifferent to how the system works. There is no argument to have about whether the system is really thinking, really understands, or is conscious. The test is external and in principle measurable.
What is awkward about it
“Most economically valuable work” is not a fixed target. Work changes in response to automation, and the denominator moves. It is also a definition an organisation can approach by narrowing what counts as economically valuable, which is a reason to watch how the phrase is used over time rather than to treat it as settled.
Researchers at Google DeepMind published a framework in 2023 that refuses the single-threshold framing entirely. Instead of asking whether a system is AGI, it places systems on two axes at once: how well they perform, and how broadly.
Performance runs through emerging, competent, expert, virtuoso and superhuman. Generality runs from narrow to general. A chess engine is superhuman and narrow. A capable modern assistant is somewhere around emerging-to-competent and fairly general. The framework also insists on separating what a system can do from what has actually been deployed, which are routinely conflated.
What it would require
Nothing single. You would show where a system sits on both axes, with evidence for each, and the interesting claim becomes “competent AGI” or “expert AGI” rather than AGI full stop.
Why it matters here
If this framing is right, then the question “when will AGI arrive?” has no answer, because there is no single event to date. That is not evasion. It is a substantive claim about the shape of the thing, and it is the claim this site takes most seriously.
The academic tradition tends towards something much more abstract. The best-known formulation, from work published in 2007 by Shane Legg and Marcus Hutter, defines intelligence as an agent’s ability to achieve goals across a wide range of environments.
What it would require
Performance measured over a distribution of environments, weighted so that simpler environments count for more. It is a definition designed to be mathematically precise rather than practically checkable, and that is the trade it makes.
Why it is still useful
It puts the emphasis exactly where the informal arguments put it without noticing: on range. A system that is extraordinary in one environment and helpless in the next scores badly here, however impressive the first environment was. That intuition is doing work in almost every AGI argument you will read.
A fourth position, argued at length by François Chollet in 2019, holds that measuring what a system can do is the wrong test altogether, because a system can be given enormous amounts of relevant experience in advance. On this view intelligence is not skill. It is the efficiency with which new skill is acquired, given what the system started with.
What it would require
Tasks the system cannot have seen, solved from very few examples. This is why the associated benchmark is built from puzzles designed to be novel and to resist being solved by having read the internet.
Why this one bites
It is the definition under which the last few years of progress look least conclusive. A system trained on a very large fraction of written human output has, by construction, been given a great deal of experience. Demonstrating skill-acquisition efficiency separately from that is genuinely hard, and the attempts to do it are among the most informative results in the field.
The imitation game, proposed by Alan Turing in 1950, asks whether a machine can be distinguished from a person in open conversation. For decades it was the reference point.
It is included here because of what happened to it rather than because anyone still proposes it. Conversational fluency turned out to be separable from the other things people assumed came with it - reliability, reasoning, knowing when you do not know. A test that was supposed to be a proxy for general intelligence turned out to be a test of conversational fluency, which is a narrower thing.
That pattern - a test that stood in for something larger until systems got good at the test specifically - is the subject of Module 2, and it has repeated more than once since.
The five definitions disagree about the kind of evidence that settles the question.
Economic
Settled by what work actually gets done, and at what cost.
Levels
Settled by where a system sits on two axes, not by a single line.
Formal
Settled by performance across a distribution of environments, not one of them.
Learning
Settled by how little experience was needed to pick something new up.
Turing
Settled by whether you can tell, in conversation, which one you are talking to.
A system can satisfy one of these and plainly fail another, which means AGI arguments between two people using different definitions cannot be resolved by either of them getting better evidence. They are not disagreeing about the world. They are disagreeing about the question.
When you next read a timeline, the useful move is to work out which of these five the speaker is working from. Very often they have not said, and occasionally it becomes clear that they are using more than one at different points in the same argument.
Last reviewed
Enrol early
We will email you when each module publishes. You will also get AGI, Made Clear, keeping you up to date on AGI and ASI developments in a field that moves faster than anyone can follow.
Double opt-in. One-click unsubscribe. We never share your address.
AGI Made Clear covers what is happening, the evidence behind it and the claims that are not supported. This site teaches the subject. That one tracks it.
Visit AGI Made Clear →