What counts as AGI? Five competing definitions, the evidence, and why the choice matters

"AGI is close." "AGI is decades away." "AGI is already here." All three claims appear in the news in the same week, sometimes from people with similar credentials looking at the same models. Most of that disagreement is not about the evidence. It is about the word. Artificial general intelligence, or AGI, has no agreed definition, and the choice of definition quietly decides whether a given system counts.

This article lays out the five main definitions in serious use, makes the strongest case for each, and then applies them to the most recent evidence (as of October 2026). It keeps three things separate: what has been measured, what experts believe, and what is still speculation. By the end you should be able to hear any AGI claim and ask the one question that matters: AGI by which definition?

Why one word causes so much confusion

"AI" in everyday use means systems that infer how to produce useful outputs from data. The "general" in AGI was added to mark a contrast with narrow systems, like a chess engine or a spam filter, that are excellent at one thing and useless at everything else. The idea is a system that can handle a wide range of tasks the way a capable person can.

The trouble starts when you try to make "wide range" and "capable person" precise. Wide range of what: jobs, test questions, environments? Compared with which person: an average adult, a skilled professional, the best in the world? Measured how: by what the system can do, by how quickly it learns, or by what it actually does once deployed? Each answer produces a different definition, and each definition draws the finish line in a different place.

That would be an academic problem if nothing rode on it. It isn't. The October 2025 restructuring of the Microsoft–OpenAI partnership ties intellectual property rights and revenue sharing to the moment AGI is reached, and specifies that "once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel"². Billions of dollars now depend on a term without a fixed meaning.

Five definitions, each at its strongest

The fairest way to understand this debate is to take each camp seriously. Here is the best case for each.

1. The economic definition: AGI is about work. OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work"¹. The case for it: it measures what society actually cares about. If a system can do most paid work better than people, the economic and social consequences are enormous whatever is happening inside it. It is also concrete enough to write into a contract. The weakness: "most" and "economically valuable" need judgment calls, and the test depends on prices, regulation and adoption, not just capability.

2. The universal-intelligence definition: AGI is about goals across environments. Shane Legg (later a co-founder of DeepMind) and Marcus Hutter offered an informal working definition in 2007: "Intelligence measures an agent's ability to achieve goals in a wide range of environments"³. The case for it: it is general by construction and does not privilege human tasks or human ways of thinking. The weakness: their formal version is mathematically elegant but cannot be computed for real systems, so it guides thinking more than measurement.

3. The learning-efficiency definition: AGI is about how fast you learn something new. François Chollet argued in 2019 that skill alone is a poor measure, because skill can be bought with enough training data. He described intelligence instead as "skill-acquisition efficiency," measured relative to prior knowledge and experience⁴. The case for it: a system that has memorized the internet can look brilliant on familiar problems while failing at genuinely new ones. Measuring how efficiently it adapts to novelty tests the "general" part directly. The weakness: it can be hard to know how novel a test really is for a model trained on vast data, and efficiency is only one part of what people mean by intelligence.

4. The levels definition: AGI is a path, not a point. In a 2024 position paper at ICML, one of the main machine learning conferences, a Google DeepMind team led by Meredith Ringel Morris proposed rating systems on two axes: how general they are and how well they perform⁵. Their levels run from "Emerging" (equal to or somewhat better than an unskilled human) through "Competent" (at least the 50th percentile of skilled adults), "Expert" (90th), "Exceptional" (99th) and "Superhuman." They focus on capabilities rather than internal processes, on cognitive rather than physical tasks, and on "the path to AGI, not a single endpoint"⁵. As of September 2023 they classed frontier chatbots as "Emerging AGI." The case for it: it replaces a yes/no argument with a measurable ladder, and it lets different definitions sit on the same scale. The authors note, for example, that OpenAI's labor-replacement threshold "better matches 'Exceptional AGI'"⁵. The weakness: someone still has to pick the task list and the benchmarks.

5. The psychometric definition: AGI is a well-educated adult's cognitive profile. A group of 33 researchers led by Dan Hendrycks of the Center for AI Safety, including Montreal's Yoshua Bengio, defined AGI as "an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult"⁶. They borrow from the Cattell-Horn-Carroll (CHC) theory, the most widely validated model of human cognitive abilities used in psychology. The framework splits cognition into ten equally weighted areas, from general knowledge and math to working memory, long-term memory storage, visual and auditory processing, and speed. A score of 100% means AGI. The case for it: it rests on a century of research on human intelligence and produces a single, auditable number with a breakdown. The weakness: tests built for humans may not measure the same things in machines, and equal weighting of the ten areas is a choice, not a finding.

There is also a sixth position: stop using AGI as a goal at all. A 16-author position paper at ICML 2025 in Vancouver argues that the term creates an "illusion of consensus," supercharges "bad science" through poorly defined concepts, and hides political and ethical choices behind technical-sounding language¹⁴. The case for it: if experts mean different things, the term mostly generates confusion and hype. The weakness: the concept still drives investment, contracts and policy, so it has to be defined somewhere, even if imperfectly.

What the evidence says when you apply them

Definitions only become useful when you measure against them. Two recent efforts do this directly, and they tell a story of rapid but uneven progress.

On the psychometric scale, models are about halfway, and lopsided. Hendrycks and colleagues scored GPT-4 (2023) at 27% and GPT-5 (2025) at 57%⁶. The more revealing result is the shape. GPT-5 scored full marks in reading and writing and in math, but 0% on long-term memory storage: the ability to learn something new during use and keep it. The authors describe a "highly 'jagged' cognitive profile," with "critical deficits in foundational cognitive machinery"⁶. Caveat: this is one group's test battery, scored on two models, and the paper has not yet been peer reviewed.

On the learning-efficiency scale, the picture moved fast in 2026. The ARC Prize Foundation, which runs benchmarks based on Chollet's ideas, launched ARC-AGI-3 in March 2026. It drops models into unfamiliar, game-like environments with no instructions. They must explore, work out the goal and adapt. At launch, humans could solve 100% of the environments while frontier AI systems scored below 1%⁷. In May the foundation reported GPT-5.5 at 0.43% and Claude Opus 4.7 at 0.18% on its held-out test set⁸.

Then things changed. In July, OpenAI researchers reported that turning on two settings in the test harness (the software that connects a model to the games), retaining the model's reasoning between moves and compressing long conversations, raised GPT-5.6 Sol's score on the public set from 13.3% to 38.3% with no change to the model itself¹⁰. In September, the foundation verified OpenAI's GPT-6 Astra at 62.7% on the held-out set with its standard harness, and 99.9% with a provider-supplied harness⁹. The scoring compares a system's efficiency to human players, so a near-100% score means matching human efficiency, not just finishing the games.

Even so, the foundation wrote: "we are not claiming that it is AGI"⁹. That sentence captures the debate in miniature. A benchmark explicitly designed to measure general intelligence went from under 1% to 62.7% on its standard setup, and near-human efficiency on another, in about six months, and the people who built it still do not consider the finish line crossed. Two things are worth noticing. First, the harness made a large difference, which means "what the model can do" increasingly depends on the system wrapped around it. Second, every benchmark measures a slice. Passing one slice tells you about that slice.

Why the definition changes the answer

The clearest evidence that definitions drive the debate comes from surveys of researchers.

In late 2023, Katja Grace and colleagues surveyed 2,778 researchers who had published at top AI venues¹¹. They asked about two milestones. "High-level machine intelligence" was defined as "when unaided machines can accomplish every task better and more cheaply than human workers." "Full automation of labor" was when "all occupations are fully automatable." These sound almost identical. Yet the median forecast gave a 50% chance of the first by 2047 and of the second by 2116, a gap of 69 years created largely by wording¹¹. Caveat: the response rate was about 15%, and the survey was taken before the 2024–2026 wave of reasoning models.

The research community is also divided on how AGI would be reached. In a survey of 475 respondents for the AAAI's 2025 Presidential Panel on the Future of AI Research, mostly academics and mostly from North America¹², 76% rated it "unlikely" or "very unlikely" that scaling up current AI approaches would produce AGI, according to press reporting of the results¹³. That is a statement of belief, not measurement, and it predates most of the 2026 benchmark gains above. As the survey gap above shows, much of the disagreement about timing is really disagreement about which finish line people have in mind.

Separating evidence, opinion and speculation

To keep this debate honest, it helps to sort the claims.

Evidence (measured): today's leading models score highly on many cognitive tests (GPT-5 earned full marks for reading, writing and math on one battery), remain weak at learning and retaining new information over time, and have improved very quickly on benchmarks of adapting to novel tasks⁶⁹. Results depend heavily on the scaffolding around the model⁹¹⁰.

Opinion (expert judgment): whether scaling current methods is enough¹³, whether the jagged weaknesses are small fixes or deep gaps, and whether "AGI" is a useful research goal at all¹⁴.

Speculation (forecasts): when any given threshold will be crossed, and what happens after¹¹. Forecasts of this kind have wide uncertainty and, as the survey gap shows, are highly sensitive to wording.

This article's view, labelled as opinion: the levels approach⁵ is the most useful default for general readers, because it turns "Is it AGI?" into "How general, how good, and at what?" The psychometric and learning-efficiency tests⁶⁷ are the most informative public measurements so far, and both currently point to the same pattern: broad, often expert-level skill with specific, important gaps, especially in memory and continual learning. On that reading, the answer to "Is AGI here?" is "parts of several definitions are being met, none fully and none uncontroversially." That view could change quickly. The ARC-AGI-3 results show how fast one measure can move.

What this means for you

Practical moves you can use this week:

1. When you hear an AGI claim, ask "by which definition?" Economic, levels, psychometric, learning-efficiency, or something vaguer? A claim with no definition is a mood, not a forecast.
2. Ask "measured on what, and with what harness?" A score without the test, the date and the setup is hard to interpret. Harness choices alone tripled one model's score¹⁰.
3. Plan around capabilities, not labels. For your work, the useful question is which of your tasks today's systems do reliably, not whether they deserve a title. Test on your own hardest cases.
4. Watch the weak spots. Long-term memory and learning on the job are where models scored lowest⁶. If your workflow depends on a system remembering and improving over weeks, check that specifically.
5. Treat timelines as ranges tied to definitions. If someone gives you one year without the definition and the uncertainty, discount it.

Sources

1. OpenAI. (2018). OpenAI Charter. openai.com/charter
2. OpenAI & Microsoft. (2025, October 28). The next chapter of the Microsoft–OpenAI partnership (also filed as Exhibit 99.2 to Microsoft Form 8-K, October 29, 2025). openai.com/index/next-chapter-of-microsoft-openai-partnership
3. Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. arXiv:0712.3329. arxiv.org/abs/0712.3329
4. Chollet, F. (2019). On the measure of intelligence. arXiv:1911.01547. arxiv.org/abs/1911.01547
5. Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). Position: Levels of AGI for operationalizing progress on the path to AGI. Proceedings of the 41st International Conference on Machine Learning (ICML 2024), arXiv:2311.02462 (v5, September 2025). arxiv.org/abs/2311.02462
6. Hendrycks, D., Song, D., Szegedy, C., Lee, H., Gal, Y., Brynjolfsson, E., et al., Bengio, Y. (2025). A definition of AGI. arXiv:2510.18212 (v3, December 3, 2025). arxiv.org/abs/2510.18212, project site agidefinition.safe.ai
7. ARC Prize Foundation. (2026). ARC-AGI-3: A new challenge for frontier agentic intelligence. arXiv:2603.24621 (v2, April 17, 2026). arxiv.org/abs/2603.24621
8. ARC Prize Foundation. (2026, May 1). Analyzing GPT-5.5 & Opus 4.7 with ARC-AGI-3. arcprize.org/blog/arc-agi-3-gpt-5-5-opus-4-7-analysis
9. ARC Prize Foundation. (2026, September 3). OpenAI's GPT-6 Astra on ARC-AGI-3. arcprize.org/blog/astra
10. Bigio, I., & Sanders, T. (2026, July 29). How enabling two settings tripled our scores on the ARC-AGI-3 benchmark. OpenAI. openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores
11. Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., Brauner, J., & Korzekwa, R. C. (2024/2025). Thousands of AI authors on the future of AI. Journal of Artificial Intelligence Research. doi.org/10.1613/jair.1.19087, preprint arXiv:2401.02843 arxiv.org/abs/2401.02843
12. AAAI. (2025, March). AAAI 2025 Presidential Panel on the Future of AI Research. Association for the Advancement of Artificial Intelligence. aaai.org/wp-content/uploads/2025/03/AAAI-2025-PresPanel-Report-FINAL.pdf
13. The Decoder. (2025, March). Most AI researchers are skeptical about language models achieving AGI. the-decoder.com/most-ai-researchers-are-skeptical-about-language-models-achieving-agi
14. Blili-Hamelin, B., Graziul, C., Hancox-Li, L., Hazan, H., El-Mhamdi, E.-M., Ghosh, A., et al., Dori-Hacohen, S. (2025). Position: Stop treating "AGI" as the north-star goal of AI research. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025), PMLR 267. arXiv:2502.03689. arxiv.org/abs/2502.03689