Artificial General Intelligence (AGI)

Artificial General Intelligence is a capability-defined category: an AI system that can match or exceed human cognitive ability across the full breadth of tasks a person can perform, rather than excelling at one narrow domain. The defining word is general: AGI is contrasted with narrow AI, which is competent only within a bounded task or domain. The term has no single agreed definition, and the disagreement over how to define it carries practical consequences for contracts, corporate safety commitments, and governance.

AGI is distinct from several adjacent concepts. AGI Timelines tracks forecasts of when AGI arrives, whereas this page covers what it is. Transformative AI (TAI) is an impact-defined alternative that brackets the capability question. Superintelligence refers to capability beyond the human range (not yet a separate page). General-Purpose AI (GPAI) is a specific EU AI Act regulatory category.

Origins of the term

The aspiration is as old as the field: the 1956 Dartmouth workshop framed AI as building machines that could do anything a human mind can. For decades, practical AI was narrow. The phrase "artificial general intelligence" was popularized only around 2007, notably by the edited volume Artificial General Intelligence (Goertzel & Pennachin, Springer, 2007) and a community that wanted to re-mark the original ambition as something separate from applied or narrow AI (Source: Goertzel & Pennachin, eds., Artificial General Intelligence, Springer, 2007). The term re-entered mainstream discourse with the rise of large language models, whose breadth made the "general" question concrete rather than speculative.

The definitional contest

There is no agreed definition of AGI, and the disagreement is substantive rather than cosmetic. The main families of definition are:

• Economic / labor-replacement. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," framing AGI by what it can displace rather than by any cognitive benchmark (Source: openai.com).
• Capability matrix. Google DeepMind's "Levels of AGI" framework (Morris et al., 2023) rejects a single threshold and instead grades systems on two axes: performance (Emerging, Competent, Expert, Virtuoso, Superhuman) and generality (narrow vs. general). It defines AGI by capabilities rather than process or consciousness, and by potential rather than deployment (Source: arxiv.org).
• Human-level / cognitive-benchmark. Older "human-level machine intelligence" framings used in expert surveys, anchored to matching a typical human across cognitive tasks. Statements by Google DeepMind's leadership use a variant that requires all cognitive capabilities rather than most: Demis Hassabis's July 2026 framework essay defines AGI as "a system that exhibits all the cognitive capabilities the brain has" (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026)), and the DeepMind Institute's September 2026 founding statement, by Hassabis, Shane Legg and James Manyika, uses nearly the same wording, "a system that exhibits all the cognitive capabilities of the human brain" (Introducing the DeepMind Institute (Legg, Manyika and Hassabis, September 2026)). Google DeepMind's June 2026 report From AGI to ASI, on which Legg is a senior author, instead defines AGI as "median human-level on most 'cognitive' tasks" and grounds its definitions in the Legg-Hutter measure (From AGI to ASI (Google DeepMind, June 2026)).
• Process / understanding. Definitions that demand genuine reasoning, world-modeling, or understanding rather than behavioral parity, and so treat current LLMs as disqualified regardless of benchmark scores.
• Psychometric. "A Definition of AGI" (Hendrycks, Song, Szegedy et al., 2025) defines AGI as "an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult," and operationalises it against Cattell-Horn-Carroll theory as ten equally weighted cognitive components across acquired knowledge, central executive, perception and output (A Definition of AGI (Hendrycks, Song, Szegedy et al., 2025)). The equal weighting is the design choice that gives the definition teeth: a system cannot score highly by excelling in a few domains. Applying it, the authors report 27% for GPT-4 and 57% for GPT-5, with a "jagged" profile — proficiency in knowledge-intensive domains alongside deficits in foundational cognitive machinery, "particularly long-term memory storage." Its 33 co-authors span institutions and include researchers who disagree publicly about AI trajectories, among them Yoshua Bengio, Gary Marcus, Max Tegmark, Erik Brynjolfsson and Eric Schmidt. See Jagged Frontier.

The OpenAI–Microsoft commercial agreement reportedly ties "AGI" partly to a profit figure (on the order of $100B in cumulative profits), making AGI a contractual term whose triggering would reshape the partnership.

The critique

A cluster of researchers argues the term is ill-defined to the point of being misleading:

• Unfalsifiability and goalpost-shifting. Because there is no agreed test, "AGI" claims can be neither confirmed nor refuted, and the bar moves as capabilities advance.
• Category error. Alison Gopnik and the cultural-technology school argue that treating LLMs as proto-minds en route to AGI misdescribes what they are, holding that a transformative technology need not be an agent or a mind.
• Marketing function. Critics including Emily M. Bender, Timnit Gebru, and Gary Marcus argue "AGI" often functions as hype that inflates valuations and racing pressure. The TESCREAL critique situates AGI discourse within a broader ideological bundle.

AGI is best treated as a contested term rather than a settled milestone. Statements of the form "X predicts AGI by year Y" are distinct from any claim that AGI will arrive.

The absence of an agreed test becomes practically consequential when arrival is declared. Nvidia chief executive Jensen Huang declared the race to AGI over on September 6, 2026; Marcus replied the same day that the declaration came with "no evidence and no definitions," that declaring victory without a definition "simply muddies the waters," and pointed Huang to the psychometric framework above (Source: garymarcus.substack.com). Marcus recorded that of the ten items in his standing bet with Miles Brundage he expects Astra may have reached autoformalization and possibly reliable coding while doubting it has met the other eight, that François Chollet of ARC Prize wrote on September 3, 2026 that saturating ARC 3 is not a claim of AGI, and that Huang had made a similar declaration about an earlier model roughly six months before. Days earlier, OpenAI president Greg Brockman had closed the Astra launch briefing with "Welcome to the AGI era" while leaving users to judge whether the model meets the definition. See AGI Timelines, ARC-AGI.

The geopolitical framing

Where the definitional contest concerns what AGI is, a separate strand asks what possessing it would do. Stokes (CNAS, August 2026) stipulates a world where AGI is imminent or exists rather than judging whether it will arrive, and argues that building it "would not automatically convert to geopolitical dominance": the state holding it "would have to apply it to translate technological capability into wealth, power, and influence" in ways cost-effective against alternatives, and "sometimes that AGI capabilities-to-power conversion could prove unworkable." The report distinguishes AGI from superintelligence — a system that "greatly exceeds the cognitive performance of humans in virtually all domains of interest" — and from transformative AI, noting that commentators often conflate them. It also records that emergence is unlikely to be binary, arriving instead "through a series of high-end evolutions that collectively add up to AGI," possibly with AGI as "a top layer that coordinates a series of other specialized agents," which bears directly on whether any observable threshold exists for policy to attach to. On terminology it notes that Beijing uses a different term for what it pursues — "general artificial intelligence," or "general-purpose AI" (Superpowers and AGI: How U.S.-China Competition Could Be Shaped by the Emergence of Artificial General Intelligence (CNAS, August 2026)). See US-China AI Competition: Different Races, Different Metrics.

Relation to policy

AGI definitions trigger concrete consequences: contract clauses (the OpenAI–Microsoft agreement), corporate safety commitments and capability thresholds in Responsible Scaling Policies, national-security framings of an "AGI race," and the urgency case for existential-risk governance. Which definition prevails shapes when obligations attach and how much pressure the competitive race carries, which is why the term remains central to AI policy despite its lack of an agreed meaning.

Relationships

• related: AGI Timelines — forecasts of when AGI arrives (this page covers what it is).
• contrasts-with: Transformative AI (TAI) — impact-defined alternative that deliberately brackets the capability question.
• related: General-Purpose AI (GPAI) — the EU AI Act's adjacent regulatory category.
• contradicts: AI as Cultural Technology, Stochastic Parrots (Bender, Gebru, McMillan-Major, Mitchell, 2021) and the Octopus Test — argue the AGI framing of LLMs is a category error.
• related: AI Existential Risk, TESCREAL Framework (Gebru and Torres), Recursive Self-Improvement (RSI).
• related: Superpowers and AGI: How U.S.-China Competition Could Be Shaped by the Emergence of Artificial General Intelligence (CNAS, August 2026), US-China AI Competition: Different Races, Different Metrics — the geopolitical framing of the term.
• related: Introducing the DeepMind Institute (Legg, Manyika and Hassabis, September 2026), A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026), From AGI to ASI (Google DeepMind, June 2026) — Google DeepMind definitions of AGI.

Sources

• (Source: openai.com) — OpenAI Charter: the "economically valuable work" definition of AGI.
• (Source: arxiv.org) — Morris et al., "Levels of AGI: Operationalizing Progress on the Path to AGI" (Google DeepMind, 2023): the performance × generality matrix.
• (Source: Goertzel & Pennachin, eds., Artificial General Intelligence, Springer, 2007) — the volume that popularized the term as distinct from narrow/applied AI.