Ask five AI researchers to define AGI and you’ll get five answers. One will describe a machine that can do any job a person can do. Another will quote a benchmark score. A third will say the question is badly posed and change the subject to something about embodiment. None of them will be obviously wrong.
The short definition is easy enough. Artificial general intelligence describes a system that can learn and perform any intellectual task a human can — not one skill, not a family of related skills, but anything a person could pick up with enough time and practice. A chess engine plays chess better than any human alive and cannot make a cup of tea. AGI is supposed to do both, and also the thing nobody thought to specify.
Everything interesting lives in the gap between that sentence and the real world.
The load-bearing word is “general”
Strip the definition down and you get three words doing very different amounts of work. Artificial is uncontroversial — built by humans, running on silicon. Intelligence has been argued over since at least 1921, when a symposium of psychologists failed to agree on a definition and published the disagreement instead.
General is where the real weight sits. Today’s systems are narrow, and narrow AI is not a criticism. AlphaFold predicted the structure of roughly 200 million proteins, a task that would have taken experimental biologists decades. It has no idea what a protein is for, cannot explain its own output in plain language, and would be useless on a question about tax law. Put a fence around a task and modern AI keeps winning. AGI is the proposal to remove the fence.
If you want the version of this with fewer footnotes, there’s a solid plain-English explanation of what AGI is and where the hype comes from that covers the basics without the academic scaffolding.
Where the phrase came from
“Artificial general intelligence” is younger than most people assume. It appears in a 1997 paper on military nanotechnology by Mark Gubrud, then gained real traction with a 2007 edited collection by Ben Goertzel and Cassio Pennachin that used it as a title. Before that, researchers mostly said “strong AI” or “human-level AI” and meant roughly the same vague thing.
One detail from the early days is worth knowing. Shane Legg and Goertzel ran a survey that collected dozens of formal definitions of intelligence from psychology, biology and computer science, then tried to merge them into one. What they produced was careful, useful, and still not something you could test a machine against on a Tuesday afternoon. The ambiguity in the AGI definition isn’t a recent failure of marketing. It was inherited.
Four definitions people actually use
When two people disagree about whether AGI is near, they are usually working from different definitions and haven’t noticed. Here are the four you’ll run into.
1. Human-level across the board
The classic version: a system that matches a typical adult human on the full range of cognitive tasks, from writing a lease to diagnosing a rash to noticing that a colleague is upset. It’s intuitive and almost impossible to operationalise, because “full range” is an enormous, unbounded set, and because the baseline human varies wildly.
2. The economic version
OpenAI’s charter, written in 2018, defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” That’s a deliberately different test. It doesn’t care whether the system understands anything. It cares whether it can replace the labour. Google DeepMind’s leadership has used similar language about systems that can do most cognitive tasks a human can.
3. Generality as transfer
François Chollet’s framing, which underpins the ARC-AGI benchmark, defines intelligence as skill-acquisition efficiency rather than skill itself. A general system should pick up a genuinely new problem from a handful of examples, the way a person learns a card game from one round of watching. Under this definition, a model trained on most of the internet and then fine-tuned on your specific task is not demonstrating generality. It’s demonstrating the internet.
4. Autonomy and agency
The loosest and most consequential one: a system that can pursue a goal over hours or weeks, course-correct when things go wrong, and decide what to do next without being prompted at each step. Some researchers fold this into AGI. Others argue it belongs to a separate discussion about agents.
Why the goalposts keep moving
Every few years, a task that was supposed to require general intelligence gets solved and quietly reclassified as narrow. Beating a grandmaster at Go was a milestone right up until 2016. Passing a bar exam was unthinkable until it wasn’t. There’s an old joke that AGI is whatever machines can’t do yet, and the joke survives because it keeps being accurate.
Part of this is shifting benchmarks, which is normal science. Part of it is commercial. Companies raise money against a definition, and nobody wants to fund a project that has already arrived. The disagreement isn’t only technical — it’s a structural problem in how artificial general intelligence gets defined and sold, and it shows up in every roadmap published with a three-to-five year horizon.
The vocabulary has drifted too. “AGI” and “AGI meaning” now get used interchangeably with frontier models, which are neither. There’s a useful breakdown of what the term AGI actually means and why consensus keeps failing if you want the semantic history in more depth.
What AGI is not
A lot of confusion clears up by subtraction.
- Not a large language model with tools bolted on. Search, code execution and image generation make a model more useful. They don’t make it general, because a human still chooses which tool to reach for.
- Not superintelligence. AGI is the human-level benchmark. Artificial superintelligence is the step after, where the system exceeds us everywhere. Mixing them up makes timelines sound far more dramatic than the underlying claims.
- Not consciousness. Nothing in the standard definition requires subjective experience. A system could be indistinguishable from a person in output and have nothing going on inside, and most researchers consider that possibility entirely live.
- Not a robot. Bodies help with physical chores. The definition is about cognition, not locomotion.
- Not a benchmark score. Chatbots that stay convincingly in character for 40 turns are genuinely impressive, as anyone who has tested these hands-on Character.AI prompt examples will tell you. It’s still a narrow skill.
The version with legal consequences
In 2019, OpenAI licensed its technology to Microsoft with a clause that ran only up to the point OpenAI declared AGI had been achieved. After that, the licence would lapse. For years it was the most expensive philosophical footnote in the industry, and it turned a definitional question into a corporate governance trigger.
That’s why the wording matters more than it looks. When a contract says “AGI,” someone has to decide whether the threshold has been crossed, and whoever decides holds real power. Regulators have the same problem. The EU AI Act regulates general-purpose AI models, a deliberately narrower category that sidesteps the AGI question entirely, because writing a legal test for “general” turned out to be harder than writing one for “large.”
How you’d actually recognise it
Forget the philosophy for a moment and imagine you had to design an entrance exam. Three properties do most of the work.
Sample efficiency. A person learns tic-tac-toe in about ten games. AlphaZero played millions before it stopped losing. If a system needs a dataset the size of a small country to acquire a new skill, it is not learning like a human, whatever its test scores say.
Novelty. The ARC-AGI benchmark uses coloured-grid puzzles that are trivial for people and were, until recently, brutal for machines. Progress in late 2024 was real and also required enormous compute per puzzle. That gap between “solved” and “solved cheaply and reliably” is where most AGI claims quietly fall apart.
Long horizon. Book a trip, have the airline cancel the connecting flight, rebook it, rebook the hotel, file the expense report in the right format, and flag that something looks off. No single step is hard. Doing forty of them in sequence without a human steering is a different order of problem.
If you want to check these claims yourself rather than trust a press release, this guide to running research with You.com that survives scrutiny covers how to trace a number back to its source paper instead of a summary of a summary.
A working definition you can actually use
Here’s the version I’d hand to someone who needs to make a decision rather than write a paper.
AGI is a system you could hand an unfamiliar job to, with no task-specific training, and expect it to figure that job out at roughly the level of a competent adult — including noticing when it’s stuck and saying so.
That definition has four testable parts. Unfamiliar rules out memorised benchmarks. No task-specific training rules out fine-tuning. Competent adult sets a baseline that isn’t perfection. Noticing failure is the one people forget, and it’s the one that matters most, because a system that is confidently wrong across every domain is worse than a narrow system that knows its limits.
No publicly available system meets all four today. Several come close on the first, most stumble on the second, and almost all fail the fourth when pushed outside their training distribution. That’s not a knock on the technology. It’s a description of where the definition still has teeth — and why it’s worth keeping a precise one around, even while everyone else argues about the timeline.