Home/What is AGI?

What is AGI?

AGI — artificial general intelligence — is a system with broad human-level capability across domains. It's also a term with no agreed definition, a finish line that has been moved every time a machine approached it, and the stakes behind most AI arguments. Here's the honest version.

AGI — artificial general intelligence — means a system that can match or exceed humans across most cognitive work: not one task, but the open-ended range of things people do with their minds. That's the working definition, and nearly everyone in the field would sign it. The trouble starts one question later — which work, measured how, matched at what threshold — because on those, there is no agreement at all. AGI is the field's most argued-about term, and the arguments never resolve for a reason this article can actually explain.

The short version of that reason: AGI isn't a technical milestone waiting to be reached. It's a definitional argument wearing the costume of one. Whether we've built AGI depends on what the word means, the word has never had a settled meaning, and — this is the part with a track record — the meaning has moved every single time a machine got close to the old one.

A definition built on an undefined word

Start one level down. AGI is "general intelligence," so its definition inherits whatever "intelligence" means — and intelligence is a word psychology has spent a century failing to define for humans, let alone machines. There are dozens of formal proposals; a well-known survey by Legg and Hutter collected over seventy definitions and found no consensus, before offering their own, which also didn't become consensus.

This isn't pedantry. Every argument about whether some system "is really AGI" bottoms out in this inherited vagueness. One camp means does economically useful cognitive work at human level; another means learns new domains from scratch the way a person can; a third means understands, as opposed to pattern-matching — and that last one imports the entire unresolved Chinese Room debate about whether behaviour can ever establish understanding at all. These camps talk past each other because they are not disagreeing about the machine. They are disagreeing about the word, and as the AI pillar argues at length, most "what is AI really" arguments have exactly this shape: vocabulary disputes wearing lab coats. AGI is the purest case.

## Where the term came from — and what it quietly admits

Here's a detail that reframes the whole discussion: general intelligence wasn't a new ambition added to AI. It was the original one. The 1955 Dartmouth proposal that named the field conjectured that every aspect of intelligence could be precisely described and simulated — generality wasn't a stretch goal, it was the premise. For the first few decades, "AI" simply meant that project.

What actually happened is that the field retreated. Symbolic AI chased the general project and hit the wall described in the machine learning pillar; the AI winters punished anyone whose grant application said "thinking machine." The survivors narrowed: chess programs, spam filters, recommenders — systems that worked precisely because they attempted one thing. By the 1990s "AI" had come to mean narrow AI in practice, without anyone announcing the change.

"AGI" was coined in the early 2000s — popularised by Ben Goertzel, Shane Legg and colleagues around a 2007 book of that name — as a deliberate act of rebranding the original ambition. The new word existed to say: we mean the old thing, the real thing, not the narrow systems that borrowed the name. Which means the term carries a confession in its etymology. We only needed "AGI" because "AI" had already failed to mean it. The field's most hyped word is a monument to its longest retreat — worth remembering when the same word is used to suggest the destination was always just around the corner.

The finish line has a history of moving

If AGI were a fixed target, you could ask how close we are. It isn't, and the record shows it.

The testIts status as a finish lineWhat happenedVerdict afterwards
The Turing testFor decades, the criterion for machine intelligenceModern chatbots pass informal versions routinelyReclassified as “a test of deception, not intelligence” — which, to be fair, Turing said in 1950 and everyone forgot
Chess, then Go“When a machine beats the world champion, something profound has happened”1997 and 2016“Just search”; “narrow”
Winograd schemasDesigned in 2011 as the common-sense test machines couldn't gameEffectively solved by large language modelsRetired; declared contaminated and too easy
Professional exams“A machine that passes the bar exam...”Passed, along with medical licensing questionsThe questions were in the training data” — sometimes true, which is precisely the problem with exams as finish lines
ARC-style novel reasoningThe current candidate: puzzles designed to resist memorisationScores climbing; goal revised upward as they doPending — but the pattern above suggests the verdict in advance

This is the AI effect — the pattern where tasks stop counting as intelligence once machines do them — operating on its own ultimate finish line. Each demotion had a defensible local reason: the Turing test does reward deception; exam questions do leak into training data; champion-level Go is narrow. But step back and the aggregate is hard to unsee: the definition of "general intelligence" has been, in practice, whatever machines can't do yet. A finish line that retreats when approached is not a finish line. It's a horizon.

And the movement runs in both directions, which is the detail partisans on each side omit. Sceptics move the line away as capabilities land ("that's not real reasoning"). Labs and boosters move it closer as incentives demand ("our next model may be AGI"), because the term anchors valuations, mission statements, and — in at least one widely reported case — contract clauses about when a partnership's terms change. A word that decides money flows does not get to stay precise.

chess1997 ✓Turing testpassed, demotedprofessional examspassed, disputed“AGI”current positionhistorically, it moves
Retired finish lines for machine intelligence. Each was “the test” until a machine passed it, at which point it was reclassified as narrow, gameable, or contaminated. The current line has no structural reason to behave differently.

The definitions actually on the table

Because the folk definition can't settle anything, several groups have tried to pin the term down properly. The attempts disagree in instructive ways.

Definition familyAGI means...What it impliesThe weakness
Economic (used in lab charters)Outperforming humans at most economically valuable cognitive workMeasurable in principle via labour statistics; arrives gradually, occupation by occupation“Most” and “valuable” are doing enormous work; a system could qualify while failing at things a child does easily
Capability levels (Morris et al., 2023)A matrix: how skilled × how general, from “emerging” to “superhuman”Replaces the binary with a dial — today's systems land at “emerging AGI,” competent at many tasks, expert at fewHonest, but dissolves the question rather than answering it; nobody argues about “Level 2”
Skill-acquisition (Chollet, 2019)Intelligence is efficiency at learning new skills, not possession of existing onesBenchmarks must use novel problems; memorised competence doesn't count, however broadNovel-problem tests keep getting partially solved and then revised — the horizon problem, again
Behavioural/folk“Can do anything a person can do, cognitively”Intuitive; what most public argument silently assumesInherits every unresolved dispute about intelligence and understanding; unfalsifiable in both directions

Notice what the serious attempts have in common: each replaces the yes/no question with something gradable — a percentage of occupations, a level on a matrix, an efficiency score. That is the tell. When a field's best minds respond to "is it AGI?" by changing the question, the original question was malformed. The honest technical position in 2026 is that current systems are strikingly general by any historical standard — one model writes code, translates, reasons through problems step by step, and handles domains it was never specifically trained for via in-context learningand clearly short of the folk meaning, failing unpredictably at tasks a careful human would not fail, and unable to be left alone with consequential work, which is why agent reliability rather than exam scores has become the frontier's real scoreboard.

Why the argument can't resolve

Three structural reasons, all visible in the material above.

It's definitional, not factual. Two people watching the same demo, agreeing on every observable fact, can disagree about AGI because they mean different things by it. No experiment settles a disagreement about a word. Dijkstra's old line — asking whether a machine can think is like asking whether a submarine can swim — applies with full force: the submarine's motion is a fact; swimming is a choice about vocabulary.

The measurements are contested exactly where it matters. Every proposed empirical criterion runs through benchmarks, and benchmark results at the frontier are precisely where measurement is least trustworthy: contamination inflates scores, and the field's most famous claim about sudden capability jumps — emergence — turned out to be substantially a plotting artefact, a rebuttal that won a best-paper award and changed far fewer minds than the original chart. A threshold you cannot measure cleanly cannot function as a finish line, whatever you name it.

The incentives point in every direction at once. Labs benefit from AGI being close (funding, talent, urgency); the same labs benefit from it being not here yet (contract terms, regulatory breathing room); sceptics stake reputations on the current approach being fundamentally limited; safety advocates need the term vivid enough to motivate alignment work without being so vivid it reads as marketing. Everyone in the argument holds a position on the definition that happens to serve them. That doesn't make anyone dishonest. It makes the word unfixable.

What would actually be evidence

Given all that, the useful move is to stop watching the word and start watching capabilities that are checkable regardless of what anyone calls them. Three that carry real information:

Long-horizon autonomy. Can a system carry a multi-day task — with sub-goals, recoveries from its own errors, and no human rescuing it — to completion? Today's agents fail this in a specific, measurable way: per-step reliability compounds, so a 90%-per-step system fails most ten-step tasks. Watch the task-length numbers, not the demos.

Generalization to the genuinely novel. Performance on problems that verifiably postdate training, in formats designed to resist memorisation. Not exam scores — exams are where contamination lives.

Knowing what it doesn't know. A system that reliably flags its own uncertainty instead of confidently fabricating would represent more progress toward anything worth calling general intelligence than another benchmark record. It is also, not coincidentally, the capability current systems most conspicuously lack.

None of these is "AGI." All of them are real, and progress on them is progress whatever the vocabulary does. The scaling-laws question — whether these capabilities keep improving smoothly with size and compute, or whether the curve bends — is the live empirical dispute underneath the definitional noise, and unlike the definitional noise, it will actually be settled by evidence.

What to watch

Whether the folk question quietly retires. The historical pattern for "is it intelligent?" arguments is not resolution but abandonment: chess stopped being debated when the answer stopped feeling important. If "is it AGI?" fades in favour of "what can it be trusted to do unattended?", that will be the argument ending the only way it ever could.

Whether any definition acquires teeth. A definition matters when something binds to it — a contract trigger, a regulatory threshold, an insurance category. Watch for AGI acquiring a legal definition somewhere, because the first one with money attached will become, de facto, the definition, ending seventy years of philosophy by paperwork.

The task-length curve. Of every number in the field, the length of task a system can complete autonomously and reliably is the one that most resembles what people actually mean by generality. It is currently short. Whether it doubles, plateaus, or bends is checkable, and none of the parties to the definitional argument can spin it.

Everything linked above is explained at five depths — read as far as you need and stop. Start with AGI itself, follow it to Intelligence and the Turing Test, or open the map and see how the argument connects to everything else. The companion pillars cover the two questions underneath this one: what AI is, and how machine learning works.

Frequently asked questions

What does AGI stand for?

Artificial general intelligence: a system with broad, human-level cognitive capability across domains, as opposed to today's mostly task-shaped systems. The "general" is the contested part — there is no agreed test for it, and every historical candidate test has been passed and then reclassified as not counting.

Is AGI the same as superintelligence?

No. AGI usually means roughly human-level generality; superintelligence means decisively beyond human level. The terms blur in public discussion because some argue the first would rapidly produce the second — but that's a claim about dynamics, not a definition, and it's disputed.

Has AGI been achieved?

By some definitions arguably yes, by most definitions no, and the honest answer is that the question is less factual than it sounds. Current systems are more general than anything before them and still fail unpredictably at things careful humans don't. Whether that combination "is AGI" depends entirely on which definition you adopt — which is the subject of most of this article.

How close is AGI?

Predictions range from a few years to never, and the spread itself is the finding: experts with the same evidence disagree by decades because they disagree about what would count. The checkable proxy worth watching is autonomous task length — how long a task a system can complete reliably without a human rescuing it. It's currently short, and its growth curve is a real number in a debate otherwise made of vocabulary.

Why do AI companies talk about AGI so much?

Because the word does work: it anchors missions, attracts talent and capital, and frames current products as steps toward something larger. Some of that is sincere conviction, some is strategy, and the two are not separable from outside. It's worth noticing that the same organisations benefit from AGI being imminent in some contexts and not-yet-here in others — a flexibility only an undefined term can provide.

Would AGI be dangerous?

The serious version of the concern doesn't require science fiction: a system pursuing objectives with human-level competence, at machine speed, would inherit the alignment problem — the gap between the objective you specified and the outcome you meant — at much higher stakes than today's systems. How large that risk is, and how it compares to nearer-term harms, is genuinely contested among researchers; the entry on alignment covers the debate rather than one side of it.

Sources & further reading

The primary literature behind the claims above, drawn from the concept entries this article links to — so a claim carries the same source here as it does there.

  • Sainz et al. (2023), NLP Evaluation in Trouble: On the Need to Measure LLM Data Contamination for each Benchmark — the problem stated plainly. Benchmark Contamination
  • Golchin & Surdeanu (2023), Time Travel in LLMs: Tracing Data Contamination in Large Language Models — detecting it from outside, without corpus access. Benchmark Contamination
  • Zhou et al. (2023), Don't Make Your LLM an Evaluation Benchmark Cheater — how contamination inflates scores, and what it does to comparisons. Benchmark Contamination
  • Amodei et al. (2016), Concrete Problems in AI Safety — still the clearest framing of the near-term technical issues. AI Alignment
  • Christiano et al. (2017), Deep Reinforcement Learning from Human Preferences — the technique behind RLHF. AI Alignment
  • Bai et al. (2022), Constitutional AI — one approach to supervision that doesn't scale with human labellers. AI Alignment
  • Morris et al. (2023), Levels of AGI: Operationalizing Progress on the Path to AGI — an attempt to make the term measurable, and a fair account of why it's hard. AGI (Artificial General Intelligence)
  • Bubeck et al. (2023), Sparks of Artificial General Intelligence — the most cited argument that something changed, and worth reading with its critics. AGI (Artificial General Intelligence)
  • Chollet (2019), On the Measure of Intelligence — the case that current benchmarks measure skill, not intelligence, and a proposed alternative. AGI (Artificial General Intelligence)
  • Legg & Hutter (2007), A Collection of Definitions of Intelligence — over seventy published definitions, which is itself the finding. Intelligence
  • Turing (1950), Computing Machinery and Intelligence — where the field chose to sidestep the definition and ask about behaviour instead. Still the most influential dodge in computer science. Intelligence
  • Searle (1980), Minds, Brains, and Programs — the Chinese Room; forty-five years unresolved. Turing Test

Go deeper, one concept at a time

See how all 202 concepts connect →