What Is AGI? Definitions Compared

The encyclopedia's longer treatment is Definitions and operational criteria for AGI; the overview is Artificial general intelligence. Terms such as narrow AI are in the glossary.

#Quick reference

The grades describe the status of each claim, not which definition is "right". See the evidence framework.

#Where does the term AGI come from?

According to Morris et al., the original use of "artificial general intelligence" was in a 1997 article on military technologies by Mark Gubrud, who described AI systems that rival or surpass the human brain in complexity and speed and can acquire, manipulate and reason with general knowledge. The same paper says Shane Legg and Ben Goertzel popularised the term among computer scientists in 2001, describing AGI as a machine able to do the cognitive tasks that people can typically do.

The label became a field name with the edited volume Artificial General Intelligence (Springer, Cognitive Technologies series, 2007), edited by Ben Goertzel and Cassio Pennachin. It was used to separate work on broad, flexible intelligence from the specialised systems that dominated applied AI. Part of the motive was history: earlier promises of general machine intelligence had been followed by funding collapses, covered in The AI winters and AI winters explained.

#What is Legg and Hutter's definition?

Shane Legg and Marcus Hutter surveyed informal definitions of intelligence from psychology and AI and reduced them to one working sentence: "Intelligence measures an agent's ability to achieve goals in a wide range of environments" (Universal Intelligence: A Definition of Machine Intelligence, Minds and Machines, 2007).

They then formalised it. An agent's universal intelligence is its expected reward summed over every computable environment, with simpler environments given more weight. Simplicity is measured by Kolmogorov complexity, the length of the shortest program that describes the environment. Hutter's theoretical agent AIXI scores highest by construction.

The strength is precision: this is a definition of intelligence in general, not of "human-level" performance. The weakness is practical. The paper itself notes that Kolmogorov complexity is not computable and can only be approximated, so the measure cannot be run as a test, and it sets no threshold at which a system becomes "general".

#What is OpenAI's definition of AGI?

OpenAI's charter, published in 2018, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". (This wording is quoted from Morris et al.; we did not check it against the charter page itself, which was not reachable on 26 October 2026.)

This is an outcome definition: it asks what the system can do in the economy, not how it works. Morris et al. credit it for focusing on capabilities and for offering a yardstick, economic value. They also note two problems. Some things associated with intelligence, such as artistic creativity or emotional intelligence, do not have a clear economic value, and measuring value implies deployment, whereas a capability definition only needs the potential to do a task. The wording also leaves open how "most", "outperform" and "economically valuable" are measured.

#What are Google DeepMind's Levels of AGI?

Meredith Ringel Morris and seven co-authors at Google DeepMind proposed replacing a single AGI threshold with a matrix: Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv, first posted 4 November 2023; published at ICML 2024). One axis is performance, the other generality (narrow versus general).

The paper states six principles: focus on capabilities, not processes; on generality and performance; on cognitive and metacognitive, not physical, tasks; on potential, not deployment; on ecological validity; and on the path to AGI rather than a single endpoint. It maps older definitions onto its levels: the OpenAI threshold "better matches" Exceptional AGI, while Competent AGI is "probably the best catch-all" for many existing definitions.

Placing ChatGPT and similar models at Level 1 is the authors' own assessment, not an independent measurement. The framework is a proposal and has not been adopted as a standard.

#Are there newer definitions?

Yes. In October 2025, Dan Hendrycks and 32 co-authors posted A Definition of AGI (arXiv preprint), which defines AGI as "matching the cognitive versatility and proficiency of a well-educated adult". It splits general intelligence into ten cognitive domains drawn from psychometric theory and reports overall scores of 27% for GPT-4 and 57% for GPT-5. Those numbers come from the authors' own scoring scheme and are Preliminary.

Other proposals define intelligence as efficiency at acquiring new skills, which underlies the ARC benchmarks. See Testing claims of general intelligence and The Turing test and AI benchmarks.

#Why does the definition matter?

Because the answer to "is it AGI yet?" changes with it. Under the Levels matrix, 2023 chatbots were already Emerging AGI. Under the OpenAI wording, the question turns on economic aggregation. Under Legg and Hutter, there is no threshold at all. Timeline forecasts inherit the same problem: surveys that ask about "high-level machine intelligence" or "full automation of labour" get very different dates (see Expert surveys on AGI timelines and AGI forecasting).

Above the top level sits superintelligence, covered in What is artificial superintelligence?.

#Frequently asked questions

#Is there an official definition of AGI?

No. As of 26 October 2026 no standards body or scientific consensus has fixed one; the main definitions coexist and measure different things.

#Who coined the term artificial general intelligence?

Morris et al. trace the first use to a 1997 article by Mark Gubrud and credit Shane Legg and Ben Goertzel with popularising it around 2001. Goertzel and Pennachin's 2007 edited volume then used it as the name of a research field.

#Does any current system count as AGI?

It depends on the definition. Morris et al. classed 2023 chatbots as Level 1 "Emerging AGI", but that is the authors' own judgement, and under stricter definitions no system qualifies.

#What is the difference between AGI and narrow AI?

Narrow AI performs well on a specific task or domain, such as chess or protein-structure prediction. AGI is defined by breadth: competence across a wide range of tasks, usually including learning new ones.

#What is the difference between AGI and ASI?

AGI is roughly human-level breadth; artificial superintelligence greatly exceeds the best humans across virtually all domains. In the Levels of AGI matrix, ASI is the Level 5 general cell, "outperforms 100% of humans".

#Sources

• Legg, S., Hutter, M.: Universal Intelligence: A Definition of Machine Intelligence (Minds and Machines 17(4), 2007; DOI)
• Legg, S., Hutter, M.: A Collection of Definitions of Intelligence
• Goertzel, B., Pennachin, C. (eds.): Artificial General Intelligence (Springer, 2007)
• Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., Legg, S.: Levels of AGI for Operationalizing Progress on the Path to AGI; ICML 2024 version
• Hendrycks, D., Song, D., Szegedy, C., et al.: A Definition of AGI

Status: draft, reviewed 26 October 2026 against the primary sources above. The OpenAI Charter and Gubrud (1997) wordings are taken from Morris et al. because the charter page could not be fetched directly; new AGI definitions and model assessments appear often.