"AGI" is used daily as if it named one thing. It doesn't. It is a contested label covering at least four distinct claims, and knowing which one a speaker means is the first step in evaluating sentences like "AGI is here" or "AGI is decades away." This piece maps the definitions, the history, the measurement problem, and the honest current state.
"AGI" is a contested label, not one thing, so a claim that it is here or far away means little until you know which meaning is in use.
- One word, four meanings. Human-level at nearly all thinking tasks, able to do most economically valuable work, able to pursue long goals reliably, or a ladder of levels rather than a finish line.
- There is no agreed test. The Turing test, benchmarks and job-based tests have each failed or eroded, which is why headlines contradict each other.
- Today's systems are broad but not reliable. The 2023 levels paper put the frontier models of its day at “Emerging AGI”; where today's systems sit on that scale is disputed.
- AGI is not consciousness, and not a date. Capability and sentience are separate questions, and a published timeline is an argument, not a measurement.
- Ask two questions. Which meaning, and what test would settle it? If the speaker cannot say, the claim is opinion.
1 · Four load-bearing definitions
The capability definition: a system matching or exceeding humans across virtually all cognitive tasks — the classic reading, tracing to Mark Gubrud's 1997 usage and popularized by Shane Legg and Ben Goertzel in the 2000s.M. Gubrud, Nanotechnology and International Security, 1997 — the first documented use of “artificial general intelligence”. Archived copy; link checked 16 Sep 2026. The economic definition: systems that can perform most economically valuable work — the framing used in OpenAI's charter.OpenAI, OpenAI Charter, 2018, the source of the economic definition. Link checked 16 Sep 2026. The autonomy definition: agents that can pursue long-horizon goals in the world with human-level reliability. The generality-gradient definition: AGI as a spectrum, not a threshold — formalized in DeepMind's 2023 "Levels of AGI" paperMorris, Sohl-Dickstein, Fiedel, Warkentin, Dafoe, Faust, Farabet & Legg, “Levels of AGI”, arXiv:2311.02462, 2023. Link checked 16 Sep 2026. (emerging → competent → expert → virtuoso → superhuman, crossed with narrow vs general).
Find out which meaning is in play. The same system can count as AGI under a loose definition and fall well short under the strict one.
2 · The measurement problem
Every proposed test has failed or eroded. The Turing test was arguably passed in spirit and dismissed as measuring deception, not intelligence. Benchmark suites saturate within months of release and leak into training data. Employment-based tests ("can it do most jobs?") depend on deployment and law as much as capability. The uncomfortable, factual position: there is no agreed instrument, which is why headlines can simultaneously announce AGI's arrival and its impossibility.
Treat an "AGI has arrived" or "AGI is impossible" headline as an opinion about a test. Without an agreed instrument, neither can be settled by measurement.
3 · Where things honestly stand (mid-2026)
Frontier reasoning models and agent systems are broadly capable across text, code, vision, and increasingly action — clearing many pre-2020 informal AGI bars — while still failing at long-horizon reliability, genuine novelty under distribution shift, and consistent physical-world competence. On the levels framing, the paper's own authors placed the frontier language models of 2023 at Level 1, “Emerging AGI”, one step below “Competent AGI”.Morris et al., “Levels of AGI”, arXiv:2311.02462 (v5), read at source 17 Sep 2026: “Overall, current frontier language models would therefore be considered a Level 1 General AI (“Emerging AGI”) until the performance level increases for a broader set of tasks (at which point the Level 2 General AI, “Competent AGI,” criteria would be met).” An earlier version of this page said “credible observers” place today’s systems around “competent-general”; no source was given, and none was found. This site has found no agreed placement for 2026 systems on that scale, and there is sharp disagreement about the slope. Claims stronger than this, in either direction, outrun the evidence.
Be wary of confident claims in either direction. The evidence supports broad capability with real gaps in long-horizon reliability and novelty; anything stronger outruns it.
4 · What AGI is not
Not consciousness — capability claims and sentience claims are independent. Not superintelligence — ASI is a further claim about exceeding all humans at everything. Not a product launch — no lab's announcement settles a definitional dispute. And not a date — every published timeline is an argument, not a measurement.
Keep the questions apart. Whether a system is capable, whether it is conscious and when "AGI" arrives are separate claims, and a lab announcement settles none of them.
5 · Quick FAQ (the questions people actually search)
Is ChatGPT/Claude AGI? By the strictest meaning, no; by looser ones, partly — see the four definitions.
Is AGI conscious? Separate question entirely; capability ≠ sentience.
When will AGI arrive? Every date you've read is an argument, not a measurement — experts range from "effectively now" to "decades."
Is AGI the same as ASI? No — ASI means beating all humans at everything, a further claim.
6 · Why the definition fight matters
Contracts, safety commitments, and even corporate governance provisions can hinge on the phrase. Regulation drafted against a word without an instrument becomes lobbying terrain. The practical takeaway for a reader of this site: whenever you meet the term, ask which of the four claims — and what test the speaker would accept as settling it. If they can't answer, you've learned what you needed.
The word carries real stakes. Contracts and governance can hinge on it, so ask which meaning a speaker uses and what test they would accept as settling it.
Sources
Link states re-checked 16 September 2026. The Gubrud archive, the OpenAI Charter and the arXiv paper all answered. The Science paper returns 403 to an automated request, so it is named here but was not opened on that date.
· Gubrud, M. (1997) — Nanotechnology and International Security (first documented use of "artificial general intelligence")
· OpenAI (2018) — OpenAI Charter (the economic definition)
· Morris, Sohl-Dickstein, Fiedel, Warkentin, Dafoe, Faust, Farabet & Legg (2023) — Levels of AGI, arXiv:2311.02462 (the levels framework; also documents Legg/Goertzel popularizing the term ~2001)
· Mitchell, M. (2024) — Debates on the nature of AGI, Science (the definitional dispute itself)
Part of the Stay Human record. Related: how to ask AI well