"A Definition of AGI" is a paper by 33 authors led by Dan Hendrycks, Dawn Song and Christian Szegedy, posted to arXiv as 2510.18212 on October 21, 2025 and last revised December 3, 2025. It proposes a quantifiable operational definition of artificial general intelligence and a measurement framework derived from human psychometrics, and reports scores for GPT-4 and GPT-5 under that framework.
The paper is presented by its authors as addressing a definitional gap: "The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition." See Artificial General Intelligence (AGI), AGI Timelines.
The definition
The definition the paper proposes is a single sentence: "AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult." Two properties are doing the work — versatility, meaning breadth across cognitive domains, and proficiency, meaning depth within them — with the well-educated adult as the reference standard rather than expert or superhuman performance.
To make the definition measurable, the authors ground it in Cattell-Horn-Carroll theory, which they describe as "the most empirically validated model of human cognition," and adapt established human psychometric batteries to evaluate AI systems.
The ten components
General intelligence is divided into ten core cognitive components derived from CHC broad abilities and "weighted equally (10%) to prioritize breadth and cover major areas of cognition," grouped into four families:
| Group | Components |
|---|---|
| Acquired Knowledge | General Knowledge; Reading and Writing Ability; Mathematical Ability |
| Central Executive | On-the-Spot Reasoning; Working Memory; Long-Term Memory Storage; Long-Term Memory Retrieval |
| Perception | Visual Processing; Auditory Processing |
| Output | Speed |
The equal weighting is a design choice with a stated purpose — it prevents a system from scoring highly by excelling in a few domains — and it is what produces the paper's headline result. Each component is decomposed into sub-abilities with their own scores; General Knowledge, for instance, is scored across commonsense, science, social science, history and culture, with GPT-4 at 8% and GPT-5 at 9% on that component.
Results and the jaggedness claim
The paper reports composite AGI scores of 27% for GPT-4 and 57% for GPT-5. The authors read these as quantifying "both rapid progress and the substantial gap remaining before AGI."
The distributional finding is presented as the more informative one: applying the framework "reveals a highly 'jagged' cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage." Under equal weighting, a deficit in one component caps the composite regardless of performance elsewhere, so the jaggedness and the headline number are the same observation viewed two ways. See Jagged Frontier.
Authorship and provenance of the framework
The author list spans 30 institutional affiliations and includes researchers who disagree publicly about AI trajectories: Yoshua Bengio, Gary Marcus, Max Tegmark, Erik Brynjolfsson, Eric Schmidt, Jaan Tallinn, Dawn Song, Yarin Gal, David Duvenaud and Steve Omohundro among them. Hendrycks, Mazeika, Phan, Ingebretsen, Khoja and Zou are at the Center for AI Safety, under whose mark the project site is published. Kevin McGrew, of the Institute for Applied Psychometrics, is a CHC-theory researcher rather than an AI researcher, which is the link to the psychometric literature the framework adapts.
The document is a research paper with an author list, affiliations and results, not an open letter — the 33 names are co-authors rather than signatories.
Later citation
Marcus invoked the paper on September 6, 2026 in response to Nvidia chief executive Jensen Huang's declaration that the race to AGI was over, writing that the declaration came with "no evidence and no definitions" and that declaring victory without a definition "simply muddies the waters" (Source: garymarcus.substack.com). The citation was to an existing 2025 paper, not to a new statement.
Provenance
Captured from the arXiv abstract page, the canonical host, and from the authors' project site at agidefinition.ai, both HTTP 200, on September 7, 2026. The arXiv record gives submission on October 21, 2025, revisions on October 23, 2025 and December 3, 2025, subjects cs.AI and cs.LG, DOI 10.48550/arXiv.2510.18212, and a CC BY 4.0 licence. The BibTeX record gives year 2025.
The full PDF — roughly 20 MB with figures — was not downloaded. Held content covers the abstract, the complete author list with affiliations, the ten-component framework as the authors present it, one of ten per-component results tables, and the citation record. The psychometric battery adaptations, the methods, the nine other component tables and the bibliography are not held, and claims about them cannot be sourced from this capture.
Relationships
- depends-on: Artificial General Intelligence (AGI)
- supports: Jagged Frontier
- contradicts: AGI Timelines — supplies a measurement against which declarations that AGI has arrived can be tested, and reports a substantial remaining gap as of GPT-5
- related: ARC-AGI — a separate operationalisation of general intelligence, from abstraction and reasoning tasks rather than psychometric batteries
- related: Embodied AI vs AGI (China's Race Redirect)
- related: Dan Hendrycks
- related: Center for AI Safety (CAIS)