The AGI definition consensus
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
For years, the term "artificial general intelligence" was a moving target—a concept whose meaning shifted with each new model release, allowing companies to claim progress while critics dismissed the goal as ill-defined. In October 2025, a coalition led by Turing Award winner Yoshua Bengio did something unprecedented: they gave AGI a quantifiable, testable definition anchored in the most empirically validated model of human cognition. The result is not just a definition—it is a diagnostic tool that reveals how far we still have to go.
The Definition That Matters
The paper, A Definition of AGI, published in October 2025, was authored by an interdisciplinary team including Dan Hendrycks (Director of the Center for AI Safety), Yoshua Bengio (Turing Award laureate), Eric Schmidt, Gary Marcus, Max Tegmark, and numerous other leading researchers and entrepreneurs. Its core definition is deceptively simple:
AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult.
This definition emphasises two critical dimensions: versatility—the breadth of skills across multiple domains—and proficiency—the depth of competence in each. It anchors AGI not to superhuman performance, economic metrics, or fuzzy notions of "human-like" intelligence, but to the cognitive capabilities of an educated human. As the authors note, this is a definition of human-level AGI, not economy-level AI.
The Ten Cognitive Domains
To operationalise this definition, the research team grounded their methodology in the Cattell-Horn-Carroll (CHC) theory, the most empirically validated model of human intelligence. The framework decomposes general intelligence into ten core cognitive domains, each weighted equally to emphasise breadth:
General Knowledge (K) —factual understanding of the world, including commonsense, culture, science, and history. GPT-5 scored only 9%.
Reading and Writing (RW) —comprehension and expression in written language. GPT-5 scored 10%.
Mathematics (M) —knowledge and skills across arithmetic, algebra, geometry, probability, and calculus. GPT-5 scored 10%.
On-the-Spot Reasoning (R) —solving novel problems without relying solely on previously learned schemas, tested via deduction and induction. GPT-5 scored 7%.
Working Memory (WM) —maintaining and manipulating information in active attention across textual, auditory, and visual modalities. GPT-5 scored 4%.
Long-Term Memory Storage (MS) —the capability to continually learn new information (associative, meaningful, and verbatim). GPT-5 scored 0%—a catastrophic failure in one of the most foundational cognitive faculties.
Long-Term Memory Retrieval (MR) —the fluency and precision of accessing stored knowledge, including the critical ability to avoid hallucinations. GPT-5 scored 4%.
Visual Processing (V) —perceiving, analysing, reasoning about, and generating visual information. GPT-5 scored 4%.
Auditory Processing (A) —discriminating, recognising, and working creatively with auditory stimuli. GPT-5 scored 6%.
Speed (S) —performing simple cognitive tasks quickly, encompassing perceptual speed and processing fluency. GPT-5 scored 3%.
The Jagged Profile and the Gap
The results reveal a "jagged" cognitive profile . While models are proficient in knowledge-intensive domains—general knowledge, reading and writing, mathematics—they possess critical deficits in foundational cognitive machinery. The most significant deficit is Long-Term Memory Storage, where current models score near zero per cent. This results in a form of "amnesia," forcing the AI to re-learn context in every interaction.
The reliance on massive context windows (Working Memory) is a "capability contortion" used to compensate for this lack of persistent memory. The framework quantifies the gap to AGI concretely: GPT-4 scored 27%; GPT-5 reached 57%.
Why the Definition Matters
The paper's contribution extends beyond measurement. By anchoring AGI to a quantifiable framework, it challenges the "moving goalpost" problem that has plagued the field. The definition provides:
A common reference point for evaluating progress across different models and approaches
A diagnostic tool for identifying specific cognitive deficits that must be addressed
A governance instrument for determining when regulatory thresholds have been crossed
The authors explicitly distinguish their definition from economy-level definitions, such as OpenAI and Microsoft's reported agreement to define AGI as "AI that can create $100 billion in profits". This separation of cognitive from economic capability is crucial: it prevents commercial interests from redefining AGI to suit quarterly earnings.
The GFN Context
For Global Future Nexus, the Bengio framework is essential infrastructure for AGI governance. Without a shared, testable definition of AGI, regulation is impossible, accountability is unenforceable, and claims of progress are unverifiable. The framework's granular assessment of cognitive deficits—particularly the zero in Long-Term Memory Storage—reveals that current systems, for all their fluency, lack the cognitive machinery that defines genuine general intelligence.
This gap is not a failure—it is a diagnostic. It tells us where the science must go and where governance must focus. GFN's work on AGI identity, cross-species trust, and anticipatory governance must account for systems that excel in some domains while failing catastrophically in others. The Bengio framework provides the map. The question is whether we will use it.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
