What Is Artificial General Intelligence, Actually?

Core argument: No agreed definition of artificial general intelligence exists, and that is not sloppy language — it is structure. The academic core of the term is breadth of capability. The working definitions are written by organizations with products to sell and contracts to trigger: OpenAI measures it in economic output, DeepMind in levels, Anthropic avoids the word, LeCun rejects it. Read it as a direction, not a milestone.

This is the fifth chapter of an ongoing series on world models. So far: why LLMs hallucinate, what separates a world model from an LLM, how the four technical routes actually work, and who is building them. The series hub holds the full map.

1. The short answer

Ask what artificial general intelligence means and you will get a confident answer from every corner of the internet. Wikipedia leads with one version: a hypothetical type of AI that matches or surpasses human capabilities across virtually all cognitive tasks. Education companies publish their own explainers, and their pages fill the search results. Each version is reasonable. None is authoritative: the field has never agreed on a definition, and the organizations with the most at stake have each written one that suits them.

One abbreviation note before going further. The initials AGI also stand for adjusted gross income, the figure at the top of a US tax return — the IRS uses the term that way, and search engines cannot reliably tell an intelligence question from a tax question. This series spells the phrase out on first use and keeps the two meanings apart.

2. A short history of the word

The term is younger than it sounds. Mark Gubrud used it in a 1997 paper on the military implications of nanotechnology, and researchers Shane Legg and Ben Goertzel revived it around 2002 to separate work aimed at broadly capable minds from the narrow, task-specific AI that paid the bills. A 2007 essay collection edited by Ben Goertzel gave the phrase its name in print, and it spread through the research community from there.

Generality was not a modern upgrade — it was the founding ambition. The 1955 proposal that launched AI as a field, written by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, aimed at machines that could carry out every aspect of learning, not one task at a time. Sixty years of narrow, task-specific systems made the word “general” do real work again.

Then the term escaped the seminar room. By 2023, Microsoft Research was putting “artificial general intelligence” in the title of a paper about GPT-4, and from there the phrase moved into product launches and earnings calls. A term that once marked a research direction became a flag that organizations plant wherever they want attention.

3. Think of it like growing up — and where that fails

“Adult” is a word everyone uses and no one owns. The law draws a line at a fixed age and treats it as if crossing it changes a person in one step. Psychologists describe maturity as a set of separate capacities on different schedules: financial independence arrives before emotional regulation for some people, and never fully for others. The word works because everyone gets the picture, and it fails whenever someone treats one drawn line as the whole truth.

Artificial general intelligence sits in the same spot. Each institution draws its own line on its own axis — economic output, cognitive breadth, autonomous agency — and then argues from its line. A system can be superhuman on one axis while missing others entirely: today’s models write code beyond most professionals, yet cannot hold a goal for a week or move a cup. As with adulthood, the honest account is that the line is drawn, not discovered.

The analogy fails in a way that matters, though. Every human passes through adulthood on roughly the same path, so the markers arrive on their own and the debate is only about which ones to count. Nothing forces an AI system to march through human capabilities in order. The line will not be crossed quietly and then become obvious; it will be declared, by whoever holds the pen. That is why the four definitions below look the way they do.

4. Four organizations, four definitions

OpenAI: an economic bar

OpenAI’s charter defines artificial general intelligence as “highly autonomous systems that outperform humans at most economically valuable work” that benefits all of humanity. Read it closely: the unit of measurement is work, not mind. It is a definition built for an organization whose mission is deploying systems across the economy — and it has direct contractual consequences, more on that below.

Google DeepMind: a ladder, not a line

A 2023 Google DeepMind paper, “Levels of AGI,” defines artificial general intelligence as non-biological systems that match or exceed the cognitive capabilities of humans across a wide range of tasks. Instead of a yes-or-no line, the authors sort systems along two axes: performance (emerging, competent, expert, virtuoso, superhuman) crossed with generality (narrow versus general). They explicitly exclude autonomy, consciousness, and embodiment from the definition. Under this scheme, contemporary chatbots land at “emerging AGI” — general in breadth, still below skilled-adult depth. The ladder’s real contribution is turning a shouting match into coordinates.

Anthropic: refuse the word, keep the target

Dario Amodei’s 2024 essay “Machines of Loving Grace” barely uses the term. He writes about “powerful AI” and specifies it concretely: systems smarter than a Nobel Prize winner across most relevant fields, working at ten to one hundred times human speed, deployable in millions of copies — “a country of geniuses in a datacenter.” Speaking at Davos in January 2025, he called AGI a “marketing term.” Note what did not happen: the target did not move, only the word.

Yann LeCun: reject “general” itself

Yann LeCun has argued for years that intelligence is not a single scale and that human intelligence is not “general” either — so a destination named “general intelligence” smuggles in a wrong picture of what is being built. His company, founded in 2025, is called Advanced Machine Intelligence (AMI Labs). The name is the argument: he expects machines that are superhuman in some domains and absent in others, not a universally general mind. (We covered the company and its $1.03 billion seed round in the previous chapter.)

5. Why the definitions bend

Definitions written by interested parties bend toward the writer. Two forces do most of the bending.

Contracts. The Information reported in December 2024 that a 2023 OpenAI-Microsoft agreement pinned the AGI trigger to a system generating at least $100 billion in profits — a dollar figure, not a capability test, because the word controlled when Microsoft’s share of the relationship changed. When a word moves money, the party that can define it owns it. The clause was later amended away — its own lesson: the definition was worth negotiating.

Marketing. The “Sparks of Artificial General Intelligence” paper put the phrase in a title next to a product name, and the caveats in its text did not survive into the headlines. Once a term carries that much attention, a lab gains nothing by defining it strictly and loses nothing by using it loosely.

And the goalpost slides in plain sight. Sam Altman wrote in January 2025: “We are now confident we know how to build AGI as we have traditionally understood it.” The hedge carries the sentence — it concedes that the meaning has already changed once and may change again. Nobody announces a finish line that their own definition keeps moving.

6. The measurement gap

If the definitions disagree, could a test settle it? Not yet. The most serious attempt to look past fixed skills is François Chollet’s 2019 argument that intelligence should be measured as skill-acquisition efficiency — how well a system learns new skills from few examples — rather than as performance on tasks it may have been trained on. His ARC benchmarks were built around that idea, and the current version, ARC-AGI-2, remains an open challenge.

An open benchmark can tell you progress is real; it cannot tell you when to declare a milestone, because the milestone has no owner. Every public claim that a system “is AGI” or “is not AGI” is therefore an argument about definitions dressed up as a measurement. When it might happen — and which past predictions have aged well — is the next chapter’s subject.

7. My take

Three readings, in descending order of confidence.

First: artificial general intelligence currently names two different things — a research direction (breadth of capability) and a business milestone (a contract or marketing trigger). Most confusion comes from treating the second as the first. When a company says AGI is near, check whether the sentence describes a mind or a business plan.

Second: no test exists today that would settle the question, and benchmark-based claims are gameable by construction, because systems are tuned on the tests that name them. The DeepMind ladder is the most honest vocabulary available. “Emerging AGI” is a defensible coordinate; the bare acronym used as a verdict is not.

Third, medium confidence: the labs that matter will abandon the term before they ever declare it met. Anthropic already switched vocabulary, LeCun named his company against it, and the CEO who called it marketing still predicts capable systems on a two-to-three-year clock. Watch the words labs reach for — powerful AI, advanced machine intelligence, frontier AI — rather than a declaration. If nobody plans to say the word when the thing arrives, the word was never the thing.

8. Questions people actually ask

Are current models already artificial general intelligence? By the DeepMind ladder they are “emerging”: broad in coverage, below skilled-adult depth on most axes, and no serious actor claims the charter-style bars have been met — Altman’s stated claim was about knowing how to build such systems, not having built one. The binary question is the wrong shape; the workable version is: which level, on which axis, by whose definition?

Why does this series spell out the full term instead of the abbreviation? Because the abbreviation AGI collides with adjusted gross income on US tax returns. Search engines, and readers, cannot always tell an intelligence question from a tax question. Full term first, abbreviation only after, in every chapter of this series.

What would count as proof that it has arrived? There is no instrument, so there is no proof in the formal sense. The nearest honest shape: a system that matches or exceeds skilled adults across a wide range of cognitive tasks and — if you accept Chollet’s framing — learns new tasks with human-like efficiency rather than after absorbing most of the written internet. Who predicts when, and what happened to those predictions, is the next chapter’s subject.

Sources

Shane Legg and Marcus Hutter, “Universal Intelligence: A Definition of Machine Intelligence”, Minds and Machines, 2007.
Google DeepMind (Morris et al.), “Levels of AGI for Operationalizing Progress on the Path to AGI”, 2023.
OpenAI, Charter — definition of artificial general intelligence quoted from the archived copy; the live page blocks automated requests.
Sam Altman, “Reflections”, January 6, 2025.
Business Insider, “Microsoft and OpenAI Have Put a Price Tag on Achieving AGI”, December 26, 2024, reporting The Information.
Dario Amodei, “Machines of Loving Grace”, October 2024.
Business Insider, “Anthropic CEO Says AGI Is a ‘Marketing Term’”, January 2025.
TechCrunch, “Yann LeCun confirms his new ‘world model’ startup, reportedly seeks $5B valuation”, December 19, 2025.
TechCrunch, “Yann LeCun’s AMI Labs raises $1.03 billion to build world models”, March 9, 2026.
François Chollet, “On the Measure of Intelligence”, 2019; the ARC Prize site documents the current benchmarks.
Sébastien Bubeck et al., “Sparks of Artificial General Intelligence: Early experiments with GPT-4”, Microsoft Research, 2023.
Wikipedia, “Artificial general intelligence” — used for the term history (Gubrud 1997; Legg and Goertzel around 2002) and as the baseline common definition; a tertiary source, cited only where a primary one does not exist.
IRS, “Definition of adjusted gross income” — the abbreviation collision.
Martin Ford (ed.), Architects of Intelligence, Packt, 2018 — LeCun’s stated objections to the term AGI; book-level citation, no article-level URL exists.
John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence”, August 31, 1955, reprinted in AI Magazine.

Next in this series: when will artificial general intelligence happen — the predictions on record, who made them, and which ones have already slipped. That chapter is built as a living timeline, because the only thing more revealing than a prediction is its expiration date.