What Does "AGI" Even Mean?

In 1997, a graduate student at the University of Maryland named Mark Gubrud gave a paper at a nanotechnology conference. It was called “Nanotechnology and International Security”, and it was a warning. Gubrud worried that new technologies could make wars between rough equals more unstable, and one of those technologies was a kind of AI that didn’t exist yet.

He needed a name for it. The AI of 1997 was mostly narrow: programs that did one job, like configuring a computer order or playing chess. Gubrud meant something else. He described systems that could rival or surpass the human brain, that could acquire and reason with general knowledge, and that could be used in almost any part of industry or the military. He called it artificial general intelligence.

Almost nobody noticed. The paper had little impact, and the phrase went to sleep.

Around 2002, it was reinvented. Ben Goertzel and Cassio Pennachin were editing a book about building thinking machines and needed a title. “Real AI” felt too provocative. Goertzel asked friends for ideas, and Shane Legg, who later co-founded DeepMind, suggested “Artificial General Intelligence.” The book came out from Springer in 2005, and Goertzel spread the term through conferences and papers. Only later did someone point out that Gubrud had used it first.

Fast forward to August 2025. Sam Altman, CEO of OpenAI, a company whose stated mission is about AGI, was asked on CNBC whether the new GPT-5 brought us closer to it. His answer, as widely reported: AGI is “not a super useful term.” His reason was that different people mean different things by it.

That’s the right place to start this series. Over the next seven parts, I’ll walk through the big bets on how to build general intelligence: writing down rules, letting machines learn, rewarding them, scaling them up, teaching them to reason, and the wilder ideas in between. But you can’t judge a bet without knowing what winning means. So first: what do people actually mean by AGI?

The honest answer is that there are at least six serious definitions, and they don’t agree.

Definition 1: Can it pass for a human? (The Turing test, 1950)

The oldest test is older than the field of AI itself.

In 1950, the mathematician Alan Turing published a paper in the journal Mind called “Computing Machinery and Intelligence.” It opens with a famous line: “I propose to consider the question, ‘Can machines think?’”

Then he more or less refuses to answer it. A few pages later he writes that the original question is, in his view, “too meaningless to deserve discussion.” Words like “think” are too slippery. So he swaps the question for a game he could actually run.

He called it the imitation game. A human judge chats by text with two hidden players, one human and one machine. If the judge can’t reliably tell which is which, the machine has done something worth taking seriously.

Turing even made a prediction. He believed that in about fifty years, around 2000, computers would play the game well enough that an average judge would have no more than a 70% chance of picking correctly after five minutes of questions.

He was off by about 25 years. In 2025, Cameron Jones and Benjamin Bergen at UC San Diego ran a careful version of exactly this three-way game, with five-minute conversations. When told to adopt a particular persona, GPT-4.5 was judged to be the human 73% of the time, more often than the real humans it was paired with. Meta’s Llama 3.1 405B, with the same prompt, was picked 56% of the time. ELIZA, a 1960s chatbot used as a baseline, managed 23%.

So is GPT-4.5 AGI? Almost nobody thinks so. That’s the Turing test’s big limit: it tests imitation, not ability. A system can sound perfectly human and still be unable to book a flight, fix a bug or learn a new skill. The test was a brilliant way to dodge a philosophical argument in 1950. It’s not a good finish line in 2026.

Definition 2: Can it do the work? (OpenAI’s charter, 2018)

In April 2018, OpenAI published a charter. It defines AGI in one line:

“highly autonomous systems that outperform humans at most economically valuable work.”

This is a very different kind of definition. It doesn’t ask what’s going on inside the machine, or whether it seems human. It asks: can it do the jobs?

Look at the three parts:

• “Highly autonomous” means it works with little human supervision. A chatbot that answers when asked doesn’t really count.
• “Outperform humans” means better than people, not just comparable.
• “Most economically valuable work” means most of the work people get paid for, not one task.

It’s practical and, in principle, measurable. The problems are just as obvious. “Most” is vague. “Economically valuable work” includes plumbing and nursing, which need a body, so does a brilliant AI without a robot body fail? And the economy changes as AI arrives: when a task gets automated, its value often drops, and new work appears.

Other people have proposed their own work-based tests:

• Mustafa Suleyman (a DeepMind co-founder, now running AI at Microsoft) proposed a “Modern Turing Test” in 2023: give an AI $100,000 and ask it to turn that into $1 million on a retail website in a few months, with minimal human help. He called the target “artificial capable intelligence,” and argued that what an AI can do matters more than what it can say.
• OpenAI’s internal ladder, reported by Bloomberg in July 2024, had five levels: Chatbots, Reasoners, Agents, Innovators, and finally Organizations, an AI that can do the work of a whole company. OpenAI didn’t publicly confirm the framework at the time.

These point at real-world impact, but they define intelligence by its price tag.

Definition 3: Levels, not a line (DeepMind, 2023)

In November 2023, a team at Google DeepMind, including Shane Legg (the same person who suggested “AGI” for Goertzel’s book), published a paper called “Levels of AGI for Operationalizing Progress on the Path to AGI.” Meredith Ringel Morris was the first author. It was presented at ICML 2024, a major machine learning conference.

Their main move was to stop treating AGI as a single finish line and turn it into a grid. They looked at many older definitions and pulled out six principles. Two matter most here:

• Measure what it can do, not how it does it. Whether the AI “really understands” or has consciousness is left out.
• Measure both depth and breadth. Depth is performance: how good it is compared with people. Breadth is generality: how many different kinds of tasks it can do.

The other principles say to focus on thinking tasks rather than physical ones, to test what a system could do rather than what it’s deployed to do, to use tests that resemble real tasks people care about, and to think about the path rather than one endpoint.

That gives you a table. Performance runs from Level 0 (no AI) up to Level 5 (superhuman). Generality has two columns: narrow (one clear task) and general (a wide range of tasks, including learning new ones).

Two things jump out from the grid.

First, narrow AI went superhuman long ago. Deep Blue beat the world chess champion in 1997. AlphaFold predicts protein structures better than any human could. Superhuman at one thing is old news.

Second, by this scheme, the chatbots of 2023 were already “AGI”, just at the lowest level. The paper called ChatGPT, Bard, Llama 2 and Gemini “Emerging AGI”: general, but only about as good as an unskilled human on most tasks, and “Competent” on some. Competent AGI, matching the median skilled adult across a wide range of thinking tasks, was marked “not yet achieved.”

The paper also added a second ladder, for autonomy: how much the AI acts on its own. It runs from Level 0 (no AI) through AI as a Tool, a Consultant, a Collaborator and an Expert, up to Level 5, AI as an Agent, working fully on its own. The authors’ point is that capability and autonomy are separate choices. A very capable model can still be used as a simple tool, and how much freedom we give it is a design and safety decision, not something that just comes with intelligence.

A small detail shows how fluid this all is: in a 2025 revision, the authors renamed Level 4 from “Virtuoso” to “Exceptional.”

Definition 4: How fast can it learn? (Chollet and ARC, 2019)

François Chollet, the creator of the Keras deep learning library, published a long paper in November 2019 called “On the Measure of Intelligence.” It attacked a habit the whole field had.

Most AI tests measure skill: can the system play Go, translate French, pass the bar exam? Chollet argued that skill alone tells you little about intelligence, because skill can be bought. Give a system enough built-in knowledge or enough training data and it can become very skilled at a task without being able to handle anything new.

So he defined intelligence as skill-acquisition efficiency: how quickly and cheaply a system picks up a new skill, given what it started with and how much experience it got. A child who learns a new board game from three examples is showing more intelligence, in this sense, than a program that needed millions of games.

To measure it, he built the Abstraction and Reasoning Corpus (ARC). Each task is a small grid puzzle: a few example pairs of coloured grids, where the output is some transformation of the input, and then a new input. You have to work out the rule and apply it. Humans find most of them easy. The tasks only rely on very basic “core knowledge” like objects, counting and symmetry, so memorising the internet doesn’t help much.

ARC’s history is the best illustration I know of how slippery AGI tests are:

• ARC-AGI-1 (2019) resisted for years. In December 2024, OpenAI’s preview of its o3 reasoning model scored 75.7% on a held-out set at a modest compute budget and 87.5% with about 172 times more compute. Chollet wrote that this was a real breakthrough, but that ARC is “not an acid test for AGI”, and that he didn’t think o3 was AGI, because it still failed some very easy tasks.
• ARC-AGI-2 (2025) was harder. In the ARC Prize 2025 competition, which limits how much computing a solution can use, the winning team scored 24% on the private set. Outside the competition, the best verified system ARC reported by December 2025, built on Gemini 3 Pro by a company called Poetiq, reached 54% at about $30 per task.
• ARC-AGI-3 (launched 25 March 2026) changed the format to interactive games with no instructions: the agent has to explore, work out what the goal is, and plan. At launch, humans could solve every environment, and every frontier AI scored below 1%. On 3 September 2026, ARC Prize reported that OpenAI’s new GPT-6 Astra scored 62.7% with ARC’s standard test setup (99.9% with an OpenAI-specific setup that keeps the model’s hidden reasoning between steps, at around $19,000 for the run). In the same post, ARC Prize wrote: “we are not claiming that it is AGI.”

Notice the pattern: build a test that measures generality, watch AI get close, explain why that’s not quite AGI, and build a harder test. Chollet’s camp would say that’s the point: the test is working as designed, flagging the next gap. Critics would say it’s a goalpost on wheels. Both are partly right, and I’ll come back to it.

Definition 5: Goals in any world (Legg and Hutter, 2007)

The most mathematical definition came from Shane Legg again, with his PhD supervisor Marcus Hutter. Their 2007 paper “Universal Intelligence: A Definition of Machine Intelligence” (in the journal Minds and Machines) collected dozens of definitions of intelligence from psychologists and AI researchers and boiled them down to one sentence:

“Intelligence measures an agent’s ability to achieve goals in a wide range of environments.”

Then they turned that into a formula. Imagine every possible environment a computer could simulate. Put your agent in each one and see how much reward it collects. Add up the scores, but give simple environments more weight than complicated ones (simplicity is measured by the length of the shortest program that describes the environment, a quantity called Kolmogorov complexity).

It’s elegant. It ignores bodies, human-likeness and jobs, connects naturally to reinforcement learning (learning from rewards, Part 4 of this series), and gives a clear meaning to “general”: good across many worlds, not one.

The catch is that you can’t compute it. There are infinitely many environments, and Kolmogorov complexity itself is not computable; the authors say it can only be approximated. So it’s a beautiful definition of the destination and almost useless as a speedometer.

Definition 6: A human cognitive profile (2025)

One more, because it’s recent and concrete. In October 2025, a group of 33 authors led by Dan Hendrycks of the Center for AI Safety published “A Definition of AGI.” They defined AGI as matching the cognitive versatility and proficiency of a well-educated adult, and borrowed the structure from the Cattell-Horn-Carroll theory, a widely used model of human abilities in psychology. They split intelligence into ten areas, such as reasoning, memory, reading and writing, maths and perception, and scored models on each.

Their headline numbers: GPT-4 scored 27%, GPT-5 scored 57%. The profile was very uneven: strong on knowledge, weak on things like storing new long-term memories. It’s one group’s yardstick, not an industry standard, but it shows a trend: people are moving from one magic test to a profile across many abilities.

Why the definition is worth billions

If AGI were just a word for academics, none of this would matter much. It matters because, for a while, it was written into one of the biggest deals in tech.

Here’s the story, in order, using only what was published or reliably reported:

1. 2019 onwards. Microsoft invested in OpenAI and got rights to use its technology. OpenAI’s website said its nonprofit board would decide when AGI had been reached, and that AGI would be carved out of Microsoft’s commercial and IP licences, which applied only to pre-AGI technology. In plain terms: the day OpenAI’s board declared AGI, Microsoft’s rights to the newest technology would stop.
2. December 2024. The Information reported that a 2023 agreement between the companies had added a financial test: AGI would only count when OpenAI had built systems able to generate around $100 billion in profits. Neither company published that contract, so treat it as reporting, but it was widely covered. Notice what happened: the most important definition of AGI in the industry had quietly become a profit number.
3. 28 October 2025. As OpenAI restructured into a public benefit corporation, the two companies signed a new agreement. Microsoft’s own post says that once OpenAI declares AGI, the declaration will be verified by an independent expert panel. Microsoft’s IP rights were extended through 2032, including models after AGI, with safety guardrails, and Microsoft became free to pursue AGI on its own. Microsoft’s stake was about 27%.
4. 27 April 2026. They rewrote it again. Microsoft now has a non-exclusive licence to OpenAI’s models and products through 2032, and OpenAI’s revenue-share payments to Microsoft continue through 2030 “independent of OpenAI’s technology progress.” Microsoft’s announcement doesn’t mention AGI at all. Press coverage, including CNBC and The Decoder, described the effect as the end of the AGI clause.

So in about seven years, the contractual meaning of AGI went from “whatever OpenAI’s board says”, to a reported profit figure, to “whatever an expert panel confirms”, to, in practice, a fixed date on a calendar. As far as I could find, no AGI declaration has been made under any version of the deal, as of October 2026.

I find that history clarifying. When billions depended on the word, the people writing contracts found it too fuzzy to rely on and replaced it with dates and dollars.

The moving goalposts debate

There’s an old joke in AI, usually credited to the computer scientist Larry Tesler and popularised by Douglas Hofstadter: “AI is whatever hasn’t been done yet.” Once a machine can do something, people stop calling it intelligence. Chess was the ultimate test of thought until Deep Blue won in 1997, then it was “just search.” Conversation was the Turing test until chatbots passed it, then it was “just imitation.”

Both sides of the AGI argument accuse the other of moving goalposts.

One side says AGI is basically here, and sceptics keep raising the bar. In March 2023, a Microsoft Research team led by Sébastien Bubeck titled a paper on GPT-4 “Sparks of Artificial General Intelligence”, arguing it “could reasonably be viewed as an early (yet still incomplete) version” of AGI. In October 2023, Blaise Agüera y Arcas of Google and Peter Norvig, co-author of the most famous AI textbook, wrote an essay for Noema titled “Artificial General Intelligence Is Already Here”, arguing that frontier chatbots are general in a way older systems weren’t, despite their flaws. DeepMind’s grid, in effect, agrees: Level 1 general AI.

The other side says calling today’s systems AGI is the real goalpost shift. Gary Marcus answered the Noema essay by calling it “an epic act of goal-post shifting”: lowering the bar until current systems clear it. Chollet’s argument is similar in spirit: if a system needs vast data to learn what a child learns from a few examples, and still fails easy novel puzzles, “general” is the wrong word.

My honest read is that both are describing the same fact. Today’s systems are extremely broad and very uneven. They can pass a bar exam and then fumble a simple puzzle a child would solve. If your definition stresses breadth of knowledge, they look close. If it stresses fast learning of new things, reliability, or acting alone for long stretches, they look further away. The goalposts move because we keep discovering which ability we’d silently assumed came bundled with the others.

So what definition will this series use?

I won’t pick one winner, because the bets in this series aimed at different targets: logicians’ reasoning, reward in any environment, benchmark and economic usefulness. Instead I’ll ask the same practical questions of every bet:

• How general did it get? One task, a family of tasks, or many?
• How good did it get? Compared with skilled humans, using real numbers.
• How well does it learn something new? Chollet’s question.
• How much can it do alone? DeepMind’s autonomy ladder.

That’s enough to compare a 1950s logic program with a 2026 reasoning model without pretending either one is or isn’t “AGI.”

Scoreboard

Key ideas

• “AGI” was coined by Mark Gubrud in 1997 and reinvented around 2002 by Shane Legg for a book by Ben Goertzel and Cassio Pennachin.
• The Turing test measures imitation, not ability. A model passed a careful three-party version in 2025, and few people think that settled anything.
• Work-based definitions (OpenAI’s charter, Suleyman’s $1 million test) are practical but define intelligence by economic output.
• DeepMind’s Levels of AGI split the question into performance and generality, plus a separate autonomy ladder. In 2023 it called chatbots “Emerging AGI.”
• Chollet defines intelligence as how efficiently you learn new skills, and ARC keeps building harder tests as AI catches up.
• The definition has real money attached. The Microsoft-OpenAI AGI clause went from board decision, to reported profit test, to expert panel, to fixed dates.
• “Moving goalposts” cuts both ways. Today’s systems are broad but uneven, and which label you pick depends on which ability you think matters most.

Sources: Mark Gubrud, “Nanotechnology and International Security” (Fifth Foresight Conference on Molecular Nanotechnology, 1997; abstract at foresight.org); Ben Goertzel, “Who coined the term ‘AGI’?” (goertzel.org); Wikipedia, “Artificial general intelligence” (history of the term); A. M. Turing, “Computing Machinery and Intelligence” (Mind, 1950); Cameron R. Jones and Benjamin K. Bergen, “Large Language Models Pass the Turing Test” (arXiv 2503.23674, 2025); OpenAI Charter (2018); MIT Technology Review and coverage of Mustafa Suleyman’s “Modern Turing Test” (2023); Axios, “OpenAI nears ‘reasoning’-capable AI” (15 July 2024), reporting Bloomberg on OpenAI’s five levels; Morris et al., “Levels of AGI for Operationalizing Progress on the Path to AGI” (arXiv 2311.02462, ICML 2024, v5 September 2025); François Chollet, “On the Measure of Intelligence” (arXiv 1911.01547, 2019); ARC Prize, “OpenAI o3 Breakthrough High Score on ARC-AGI-Pub” (20 December 2024), “ARC Prize 2025 Results and Analysis”, “Announcing ARC-AGI-3” (March 2026) and “OpenAI’s GPT-6 Astra on ARC-AGI-3” (3 September 2026); Shane Legg and Marcus Hutter, “Universal Intelligence: A Definition of Machine Intelligence” (Minds and Machines, 2007); Hendrycks et al., “A Definition of AGI” (arXiv 2510.18212, 2025); Sherwood News and Gizmodo, reporting The Information on the Microsoft-OpenAI profit-based AGI definition (December 2024); VentureBeat, “OpenAI’s six-member board will decide ‘when we’ve attained AGI’” (2023); Microsoft, “The next chapter of the Microsoft-OpenAI partnership” (28 October 2025) and “The next phase of the Microsoft-OpenAI partnership” (27 April 2026); CNBC, “OpenAI shakes up partnership with Microsoft, capping revenue share payments” (27 April 2026); The Decoder, “OpenAI and Microsoft rewrite their deal: no more exclusivity, no more AGI clause” (27 April 2026); Fortune and Android Headlines on Sam Altman’s CNBC interview (August 2025); Bubeck et al., “Sparks of Artificial General Intelligence: Early experiments with GPT-4” (2023); Blaise Agüera y Arcas and Peter Norvig, “Artificial General Intelligence Is Already Here” (Noema, October 2023); Gary Marcus, “Reports of the birth of AGI are greatly exaggerated” (Substack, 2023).

Related: next in the series, Bet #1: Write Down the Rules (Symbolic AI); the finale, AGI: Where We Are Now (October 2026); and Benchmarks and Goodhart’s Law.