Loading
Artificial general intelligence (AGI) refers to a hypothetical AI system able to understand, learn, and apply knowledge across the same broad range of tasks a human can, rather than being confined to the narrow domain it was built for. No government, standards body, or AI lab has agreed on a single technical test for it: OpenAI's charter defines AGI operationally as "highly autonomous systems that outperform humans at most economically valuable work," while Google DeepMind researchers have proposed a graded "Levels of AGI" framework that scores systems on performance and generality rather than declaring one pass/fail line. That gap is not just academic, it once sat inside a Microsoft-OpenAI contract clause that let an AGI declaration void billions of dollars in IP and revenue-sharing terms, and it was argued in front of a jury in the 2026 Musk v. Altman trial. For governance teams, the absence of an agreed definition means AGI functions less as a fixed regulatory line and more as a moving target that every framework, contract, or safety commitment using the term has to define for itself.
Run the free AI Health CheckArtificial General Intelligence (AGI), a hypothetical AI system capable of understanding, learning, and applying knowledge across the full range of tasks at which humans are capable, rather than being narrow to a specific domain.
AGI has no agreed technical definition. Different frontier labs use different operational definitions, OpenAI references "highly autonomous systems that outperform humans at most economically valuable work"; DeepMind has published a tiered framework (Levels 0-5). Most current AI systems, including frontier large language models, are narrow AI, extraordinarily capable in some domains while limited in others. The governance question of how to prepare for advanced AI is increasingly being addressed even where the technical question remains open.
Source: DeepMind AGI Levels framework; OpenAI mission statement
OpenAI's charter definition
OpenAI's charter states its mission is to ensure AGI, "by which we mean highly autonomous systems that outperform humans at most economically valuable work", benefits all of humanity. It's an economic, outcomes-based test rather than a technical one, and it deliberately leaves out abilities (like artistic or emotional intelligence) whose economic value is hard to price.
DeepMind's Levels of AGI
A November 2023 position paper by Google DeepMind researchers (Morris et al.) proposes scoring systems on two independent axes, performance (how good) and generality (how broad), across six levels from "No AI" to "Superhuman," rather than treating AGI as a single on/off switch.
Chollet's skill-acquisition test
In "On the Measure of Intelligence" (2019), François Chollet defines intelligence as skill-acquisition efficiency on novel tasks and built the ARC benchmark (later rebranded ARC-AGI) to test it, arguing that strong performance on tasks resembling training data isn't evidence of general intelligence.
The classic Turing Test
Alan Turing's 1950 "imitation game" judges a machine by whether its conversation is indistinguishable from a human's. It predates the term AGI by decades and is now widely regarded as necessary but not sufficient, since narrow chatbots can pass short conversational tests without general capability.
The DeepMind paper separates six performance levels, each anchored to a human-percentile benchmark: Level 0 (No AI), Level 1 Emerging (equal to or somewhat better than an unskilled human), Level 2 Competent (at least the 50th percentile of skilled adults), Level 3 Expert (at least the 90th percentile), Level 4 Exceptional (at least the 99th percentile), and Level 5 Superhuman (outperforms 100% of humans). Crucially, each level applies separately to "narrow" systems (a single task) and "general" systems (a wide range of cognitive tasks), producing a two-dimensional matrix rather than one line of progress.
As of the paper's September 2023 analysis, the authors classified frontier large language models of that time (ChatGPT, Bard, Llama 2, Gemini) as "Level 1 General AI", nicknamed "Emerging AGI", because they showed competent performance on select tasks without reliably clearing the 50th-percentile bar across most cognitive domains. That sits in contrast to narrow systems that had already reached Superhuman performance on single tasks with zero generality, such as AlphaFold, AlphaGo, and Stockfish. The paper argued that no system had yet reached "Competent AGI" (Level 2 General), let alone higher levels.
The same paper proposes a companion "Levels of Autonomy" scale, Level 0 No AI, 1 AI as Tool, 2 AI as Consultant, 3 AI as Collaborator, 4 AI as Expert, 5 AI as Agent, to separate what a system can do from how independently it is deployed. It links escalating autonomy to escalating risk categories: de-skilling and industry disruption at the lower levels, over-trust and targeted manipulation in the middle, and mass labor displacement, misalignment, and concentration of power at the top end.
Regulators sidestep the term
The EU AI Act does not regulate "AGI" as a category. It instead classifies a general-purpose AI model as carrying systemic risk once cumulative training compute exceeds 10^25 floating-point operations (with a lower 10^23 FLOPs bar for GPAI status generally), triggering EU AI Office notification duties, a measurable compute proxy rather than a judgment about general intelligence.
A contract clause once tied billions to the word
Microsoft's original OpenAI partnership let OpenAI's board unilaterally declare AGI, an event that would have ended Microsoft's IP and revenue-sharing rights. An October 2025 amendment moved that call to an independent expert panel; by April 2026 the two companies removed the AGI trigger altogether, fixing the IP license to 2032 and revenue share to 2030 regardless of any future AGI claim.
It became a jury question
Musk's original 2024 complaint invoked GPT-4 having reached an "AGI threshold" to argue OpenAI breached its founding nonprofit agreement, but the narrower claims that went to trial in 2026, breach of charitable trust and unjust enrichment over the nonprofit-to-PBC conversion, did not require the jury to rule on whether AGI existed. The jury dismissed those claims on statute-of-limitations grounds without the court ever adjudicating the definitional question, illustrating how the term can carry legal weight in a dispute's background even when it's not the question actually decided.
No standards body has a compliance test for it
Neither NIST's AI Risk Management Framework nor ISO/IEC 42001 defines a compliance test for "AGI"; both frame obligations around risk level, intended use, and system lifecycle instead, meaning organizations that reference AGI in their own policies are effectively adopting a lab's or vendor's definition by default.
Has AGI already been achieved?
There is no consensus. Google DeepMind researchers classified 2023-era frontier chatbots (ChatGPT, Bard, Gemini, Llama 2) as only "Emerging AGI", the second-lowest rung of their framework, because they weren't reliably competent across most cognitive task categories yet. Some industry figures have since claimed AGI-level capability has arrived (for example, Nvidia CEO Jensen Huang, as reported by Fortune in 2026), but such claims rest on different, often undisclosed definitions and remain contested rather than settled.
What's the difference between AGI and ASI (artificial superintelligence)?
AGI describes a system that matches broad human-level capability across most tasks. Artificial superintelligence (ASI) describes a system that substantially exceeds human capability across virtually all domains. DeepMind's Levels of AGI framework captures this as its top performance rung, "Superhuman," once combined with high generality, a higher and more speculative bar than AGI itself.
How does OpenAI define AGI?
OpenAI's charter states that its mission is to ensure AGI, "by which we mean highly autonomous systems that outperform humans at most economically valuable work", benefits all of humanity. It's an economic, outcomes-based test rather than a technical benchmark, and it does not specify how such performance would be measured or verified.
What is DeepMind's "Levels of AGI" framework?
A November 2023 Google DeepMind position paper (Morris et al.) that scores AI systems on two independent axes, performance (six levels from "Emerging" to "Superhuman", anchored to human percentiles) and generality (narrow vs broad task coverage), instead of treating AGI as a single pass/fail threshold. A companion "Levels of Autonomy" scale (0 No AI through 5 Agent) separates what a system can do from how independently it is deployed.
Does any law or regulator define AGI?
No current statute or regulator formally defines "artificial general intelligence." The EU AI Act regulates general-purpose AI models using training-compute thresholds (10^23 and 10^25 floating-point operations) instead of an AGI test, and frameworks such as NIST's AI Risk Management Framework and ISO/IEC 42001 organize obligations around risk level and intended use rather than a capability-based AGI definition.
Why did a lawsuit turn on the definition of AGI?
Musk's original 2024 complaint argued that OpenAI, having released GPT-4, had already "reached the threshold of AGI," making its shift to a closed, for-profit model a breach of the founding nonprofit agreement. By the time the case reached a jury in 2026, however, the claims actually being decided had narrowed to breach of charitable trust and unjust enrichment tied to OpenAI's nonprofit-to-public-benefit-corporation conversion, the jury was not asked to rule on whether AGI had been achieved. Jurors dismissed Musk's claims on statute-of-limitations grounds, leaving both the breach question and the underlying AGI-definition question unaddressed by the court.
Last reviewed July 2026
This page is general information about What Is Artificial General Intelligence (AGI)?, not legal, regulatory, or professional advice, and does not capture every nuance or exception. Requirements change and can be fact-specific. Always verify against primary sources and your own qualified legal counsel before relying on it.