What Is AGI? How Artificial General Intelligence Is Defined and Why Timelines Differ

What is AGI?

Artificial general intelligence (AGI) is a hypothetical kind of AI that can handle a wide range of intellectual tasks at or above human level, instead of excelling at one narrow task such as translation or chess. Every serious definition shares that idea of generality, but the details differ: one 2025 research paper opens by arguing that the lack of a concrete definition obscures the gap between today's specialised AI and human-level cognition ^[3].

Three widely cited versions show the range:

• Economic: OpenAI's Charter defines AGI as 'highly autonomous systems that outperform humans at most economically valuable work' ^[1].
• Levels: Google DeepMind researchers rate systems on two axes, depth of capability (performance) and breadth (generality), and describe AGI as a path with levels rather than a single finish line ^[2].
• Psychometric: a 2025 paper led by Dan Hendrycks defines AGI as matching the cognitive versatility and proficiency of a well-educated adult, and tests it across ten cognitive domains ^[3].

Some leaders avoid the word entirely. Anthropic chief executive Dario Amodei calls AGI an imprecise term with sci-fi baggage and hype, and prefers 'powerful AI', which he describes as smarter than a Nobel Prize winner across most relevant fields ^[4].

How is AGI different from today's AI and from superintelligence?

Today's chatbots are general-purpose, but they are uneven. The International AI Safety Report 2026 notes that leading systems achieved gold-medal performance on International Mathematical Olympiad questions while still failing at some seemingly simple tasks ^[13]. Superintelligence means going beyond human level; in the DeepMind levels, the top general tier is labelled Artificial Superintelligence (ASI), with Superhuman performance defined as outperforming 100% of humans ^[2].

What are the levels of AGI?

The 'Levels of AGI' paper by Meredith Ringel Morris, Shane Legg and colleagues at Google DeepMind, published in the proceedings of ICML 2024, proposes a framework for classifying AGI models and their precursors, intended as a common language for comparing systems and tracking progress ^[2]. It sets out six principles, including judging AGI by what a system can do rather than how it works, focusing on cognitive and metacognitive tasks rather than physical ones, and treating AGI as a path rather than an endpoint ^[2].

Its general-AI levels are ^[2]:

1. Emerging: equal to or somewhat better than an unskilled human. The paper lists ChatGPT, Bard, Llama 2 and Gemini as examples.
2. Competent: at least the 50th percentile of skilled adults.
3. Expert: at least the 90th percentile of skilled adults.
4. Exceptional: at least the 99th percentile of skilled adults (called 'Virtuoso' in earlier versions).
5. Superhuman: outperforms 100% of humans.

A 2025 Google DeepMind safety paper builds on this scale and focuses on Exceptional AGI, which it defines as a system matching or exceeding the 99th percentile of skilled adults on a wide range of non-physical tasks ^[10].

How do researchers measure progress towards AGI?

Because definitions differ, so do the yardsticks. Two recent efforts illustrate the approaches.

A cognitive 'AGI score'

The Hendrycks-led framework adapts human psychometric test batteries to AI and grounds them in the Cattell-Horn-Carroll theory of human cognitive abilities ^[3]. It reports AGI scores of 27% for GPT-4 and 57% for GPT-5 ^[3]. The authors describe a 'jagged' profile: strong performance in knowledge-intensive areas but critical deficits in foundational abilities, especially long-term memory storage ^[3]. The score depends on the domains and weighting the authors chose, so it is one research group's measure, not an official grade.

Puzzles that are easy for people

The ARC Prize benchmark, ARC-AGI, is described by its organisers as a measure of few-shot generalisation on novel tasks ^[14]. In the 2025 competition on the ARC-AGI-2 dataset, 1,455 teams submitted 15,154 entries, and the top score reached 24% on the private evaluation set ^[14]. The organisers have previewed ARC-AGI-3, which adds interactive tasks testing exploration, planning, memory and goal acquisition ^[14].

Benchmarks have a known weakness: the ARC Prize report argues that frontier models' reasoning is limited by their knowledge coverage, which creates new forms of benchmark contamination ^[14]. The International AI Safety Report adds that methods for estimating how and when new capabilities emerge remain unreliable ^[13].

When will AGI arrive, according to AI leaders?

Public forecasts from company leaders are statements of belief, not evidence, and they often use different definitions. As of 3 November 2026 they span a wide range:

• Sam Altman (OpenAI) wrote in January 2025: 'We are now confident we know how to build AGI as we have traditionally understood it' ^[5].
• Dario Amodei (Anthropic) wrote in October 2024 that powerful AI could come as early as 2026, though it could take much longer ^[4]. In January 2025 he told CNBC that AI better than almost all humans at almost all tasks could emerge within two or three years ^[6].
• Demis Hassabis (Google DeepMind) said in March 2025 that AGI, which he described as a system able to exhibit all the complicated capabilities humans can, would start to emerge over the next five to 10 years ^[6].
• Google DeepMind's safety team wrote in April 2025 that it is highly uncertain about timelines but finds it plausible that powerful AI systems will be developed by 2030 ^[10].
• Yann LeCun, who left Meta to found the start-up AMI, told MIT Technology Review in January 2026 that large language models cannot truly reason or plan because they lack a model of the world, and that human-level AI is not going to happen next year or two years from now and will not be built on LLMs ^[7].
• Robin Li (Baidu) said in 2024 that AGI is more than 10 years away, as reported by CNBC ^[6].

What do surveys of AI researchers say?

Surveys of researchers are a useful check on executives' predictions, though they also measure opinion rather than fact.

The largest is a survey of 2,778 researchers who had published in top-tier AI venues, published by Katja Grace and colleagues in 2024 ^[8]. Respondents gave a 10% chance that machines would outperform humans in every possible task by 2027 and a 50% chance by 2047, a date 13 years earlier than in the same team's 2022 survey ^[8]. For all human occupations becoming fully automatable, the 50% estimate was 2116 ^[8].

The Association for the Advancement of Artificial Intelligence (AAAI) surveyed 475 respondents, about two thirds from academia, for its 2025 Presidential Panel report ^[9]. In that survey, 76% said that scaling up current AI approaches is unlikely or very unlikely to yield AGI ^[9]. Most respondents (77%) prioritised designing systems with an acceptable risk-benefit profile over directly pursuing AGI, and 70% opposed halting AGI research until full safety mechanisms exist ^[9].

The two results need not conflict: researchers can expect very capable AI within decades while doubting that today's methods, simply made bigger, will get there ^[8][9].

Why do experts disagree about AGI?

Several separate disagreements are bundled into the AGI debate:

• Definitions. A test based on economically valuable work ^[1], a percentile of skilled adults ^[2] and a psychometric profile ^[3] can give different answers about the same system.
• Methods. Some leaders expect current approaches to keep improving quickly ^[5], while LeCun argues new architectures based on world models are needed ^[7] and most AAAI respondents doubt that scaling alone will work ^[9].
• Uneven abilities. Systems that excel at hard maths can still fail at simple tasks, which makes a single verdict difficult ^[13]. The Hendrycks framework finds the same jaggedness ^[3].
• Forecasting itself. The International AI Safety Report says there is substantial uncertainty about how fast future progress will be, and that outcomes by 2030 could range from modest improvements to rapid gains, according to OECD analyses ^[13].
• Incentives. AGI has contractual and commercial meaning for some companies, as the Microsoft-OpenAI agreements show ^[11][12], so public statements can carry business weight.

Why does the definition of AGI matter for business?

The clearest example is the partnership between Microsoft and OpenAI. Under the agreement announced on 28 October 2025, any declaration of AGI by OpenAI was to be verified by an independent expert panel ^[11]. Microsoft's rights to OpenAI models and products were extended through 2032, its research IP rights were to last until the panel verified AGI or through 2030, whichever came first, and Microsoft gained the right to pursue AGI independently ^[11].

The companies changed course again on 27 April 2026. CNBC reported that OpenAI's revenue-share payments to Microsoft are now subject to a total cap and continue through 2030 'independent of OpenAI's technology progress', that Microsoft no longer needs to determine its response if OpenAI declares AGI, and that Microsoft's IP license through 2032 is no longer exclusive ^[12]. The episode shows that 'AGI' has functioned as a legal trigger as well as a scientific goal, and that its contractual role can be renegotiated.

What are the benefits and risks of pursuing AGI?

Claimed benefits

• Amodei predicts that AI-enabled biology and medicine could compress the progress human biologists would have made over the next 50 to 100 years into 5 to 10 years, and describes powerful AI as a 'country of geniuses in a datacenter' ^[4].
• OpenAI's stated mission is for AGI to benefit all of humanity ^[1].

Risks and concerns

• Google DeepMind's safety paper groups AGI risks into misuse, misalignment, mistakes and structural risks, and focuses on harms consequential enough to significantly harm humanity ^[10].
• The International AI Safety Report notes that some researchers and company leaders consider loss of control a serious possibility while others consider such scenarios implausible ^[13].
• In the AAAI survey, 82% of respondents said AGI systems developed by private entities should be publicly owned ^[9].

What happens next?

Expect more attempts to pin the term down with tests rather than slogans: frameworks such as the DeepMind levels ^[2] and the Hendrycks AGI score ^[3], and new benchmarks such as ARC-AGI-3 ^[14]. Expect, too, that each new model release will revive claims and counter-claims, because definitions remain contested ^[2][3].

What is still unknown

• Whether scaling current methods will keep delivering gains, which most AAAI respondents doubt ^[9].
• Whether foundational gaps such as long-term memory can be closed quickly ^[3].
• Whether researchers' forecasts, which moved 13 years earlier between the Grace team's 2022 survey and its latest one ^[8], will keep moving.

A practical reading for non-specialists: treat any claim that 'AGI has arrived' or 'AGI is impossible' by first asking which definition is being used and what evidence supports it.