Say “AGI” in a room full of researchers and you’ll get at least four different definitions before the coffee goes cold. One person means a system that can do any job a person can. Another means something smarter than every human alive combined. A third insists the whole term is a marketing device. They’re all arguing about the same three letters.
The meaning of AGI is far less settled than the confident headlines suggest. What follows is what the term stands for, where it came from, how the major definitions differ, and how to judge an AGI claim when you run into one.
AGI stands for artificial general intelligence
AGI is short for artificial general intelligence. In plain English, it describes a system that can learn and carry out any intellectual task a human being can, instead of being very good at one narrow job.
That’s the short version. The long version is where things get slippery, because “any intellectual task” is doing enormous work in that sentence. Can the system pick up a genuinely new skill it was never trained on? Can it transfer what it learned in one domain to another? Does it need to match an average adult, or the best expert in the field? Every definition answers those questions differently.
Narrow AI vs AGI: the distinction that actually matters
Almost everything described as AI today is narrow. Translation models, recommendation engines, fraud detectors, image classifiers and coding assistants are all built to do one thing extremely well. Ask a chess engine to write a sonnet and it has nothing to say.
Biology has produced some spectacular narrow systems. AlphaFold, DeepMind’s protein-structure predictor, has mapped the shapes of more than 200 million proteins, a job that would have taken humans centuries. That is staggering, and it is still narrow, because the model doesn’t wander off and start designing clinical trials on its own. The same pattern shows up across the sciences, where machine learning does genuine discovery work in modelling the human cell without any claim to general intelligence.
The three ingredients most definitions share
- Generality. The system handles a wide range of tasks rather than one, including tasks nobody explicitly trained it for.
- Transfer. Skills and concepts learned in one domain carry over to unfamiliar ones without starting from scratch.
- Autonomy. It can set sub-goals, plan, and adapt when the environment changes, with minimal hand-holding.
Human parity is usually implied too, though where that bar sits varies wildly between definitions.
Where the term came from
The phrase “artificial general intelligence” appears in a 1997 military-technology paper by Mark Gubrud, but it was Shane Legg and Ben Goertzel who turned it into a research banner. Legg went on to co-found DeepMind. Goertzel launched the first AGI conference in 2008 and spent years pushing the idea when most of the field dismissed it as science fiction with a spreadsheet.
The definition that has shaped the most headlines, though, comes from OpenAI’s 2018 charter, which describes AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Notice what’s missing. No mention of consciousness, no mention of understanding, no requirement to be human-like at all. It’s an economic definition wearing a philosophical costume.
Four competing definitions, side by side
When two people disagree about whether AGI has arrived, they are usually working from different definitions. These four show up most often:
- Human-level across the board. The system performs at or above a skilled adult on essentially any cognitive task. Intuitive, popular, and almost impossible to test cleanly.
- The economic definition. It can do most jobs at least as well and as cheaply as a person. This is the version driving both investment and labour-market anxiety.
- Transformative or superhuman. Not just human-level, but better than all humans at everything that matters. Some writers reserve “superintelligence” for this and use AGI for the rung below.
- Skill-acquisition efficiency. François Chollet, creator of the ARC benchmark, argues intelligence isn’t the skills you already have, it’s how quickly you pick up new ones. A system that aces every test in its training set may simply have memorised the tests.
What AGI is not
A lot of the confusion comes from words getting welded together. These are not the same thing:
- AGI is not a chatbot. A fluent conversation partner can still be narrow. Fluency is one skill, not general intelligence. The 2023 Microsoft paper claiming GPT-4 showed “sparks of AGI” was roundly criticised for exactly this reason.
- AGI is not superintelligence. Human-level and vastly superhuman are different claims with different timelines and very different risk profiles.
- AGI is not consciousness. A system could pass every practical test and have no inner life whatsoever. Nobody knows how to test the inner part.
- AGI is not a finished thing sitting in a lab right now, whatever a press release implies. Benchmark scores are evidence, not proof.
Why the definition keeps sliding
There’s an old joke, usually credited to Larry Tesler, that AI is whatever hasn’t been done yet. The moment a machine can do something, we stop calling it intelligence and start calling it computation. Something similar happens to AGI. Chess, Go, protein folding and passing professional exams were all named as milestones at various points. Each one fell, and the goalposts moved.
Part of that is healthy scepticism. Part of it is self-interest. A definition that shifts with the weather can’t be falsified, and unfalsifiable claims are excellent for fundraising. It’s one reason AI seed startups keep pulling in valuations that would look strange in almost any other sector.
Why the meaning matters in practice
Treating AGI as a fuzzy philosophical puzzle is tempting, but the definition has real consequences.
Funding is the obvious one. If AGI means “systems that outperform humans at most economically valuable work,” then every labour-market projection, safety budget and national strategy is quietly built on that specific reading. Regulators drafting rules around “general-purpose AI” are borrowing the same idea.
Safety is the other. A system that operates across domains is far harder to constrain than one that only recommends videos, because the harm it can cause is not limited to its training distribution. Claims that today’s techniques will scale to that problem deserve more scrutiny than they usually get, as critics of narrowly framed AI safety proposals have argued.
The path people are actually betting on
Whatever AGI ends up meaning, the current route to it runs through scale: bigger models, more data, more compute, and post-training tricks that squeeze more reasoning out of the same weights. Labs are pushing this hard, and Meta’s recent frontier model release is a decent illustration of how fast the top tier moves.
Hardware is the constraint. Training runs burn through enormous amounts of GPU time, and keeping thousands of accelerators busy without wasting cycles has become an engineering discipline in its own right, with serious work going into scheduling GPU clusters efficiently. If AGI ever arrives, it will arrive on hardware that somebody had to keep cool and fed.
How to read an AGI claim without getting fooled
Next time the term turns up in a headline, four questions will sort most of it out.
- Which definition is being used, and is it stated or just assumed?
- What’s the evidence — a benchmark, a live demo, or a vibe?
- Has the system been tested on tasks it was never trained for?
- Who benefits if you believe the claim?
A working definition you can actually use
If you want one sentence to carry around, try this: AGI is a system that can acquire new skills and apply them across domains at roughly human level, without being rebuilt for each new job. It says nothing about consciousness, nothing about superintelligence, and nothing about whether we are close.
That version has the virtue of being testable. It also explains why the debate stays hot. Every few months a model does something that looks general, and then we find out it had seen something like it before, or that it collapses after a small change in phrasing. ARC-AGI, Chollet’s benchmark of novel visual puzzles, was designed to catch exactly that failure. OpenAI’s o3 crossed 75% on the original set in December 2024, an impressive jump, at a reported cost of roughly $20 in compute per puzzle. Cheap enough to count as real progress. Expensive enough that it is not yet a general mind.
So the honest answer to what AGI means is this: a target that keeps moving, defined mostly by whoever is aiming at it. The useful version is the narrow one — skills, transfer, autonomy — because you can test it, argue with it, and notice the moment somebody quietly swaps it for a shinier definition.