A Definition of AGI

Paper

Dan Hendrycks¹, Dawn Song², Christian Szegedy³, Honglak Lee⁴, Yarin Gal⁵

Sharon Li⁶, Andy Zou^1,7,8, Lionel Levine⁹, Bo Han¹⁰, Jie Fu¹¹, Ziwei Liu¹², Jinwoo Shin¹³

Kimin Lee¹³, Mantas Mazeika¹, Long Phan¹, George Ingebretsen¹, Adam Khoja¹

Cihang Xie¹⁴, Olawale Salaudeen¹⁵, Matthias Hein¹⁶, Kevin Zhao¹⁷, Alex Pan²

David Duvenaud^18,19, Bo Li²⁰, Steve Omohundro²¹, Gabriel Alfour²², Max Tegmark¹⁵

Kevin McGrew²³, Gary Marcus²⁴, Jaan Tallinn²⁵, Eric Schmidt¹⁵, Yoshua Bengio^26,27

Dan Hendrycks¹, Dawn Song², Christian Szegedy³

Honglak Lee⁴, Yarin Gal⁵, Sharon Li⁶

Andy Zou^1,7,8, Lionel Levine⁹, Bo Han¹⁰

Jie Fu¹¹, Ziwei Liu¹², Jinwoo Shin¹³

Kimin Lee¹³, Mantas Mazeika¹, Long Phan¹

George Ingebretsen¹, Adam Khoja¹, Cihang Xie¹⁴

Olawale Salaudeen¹⁵, Matthias Hein¹⁶, Kevin Zhao¹⁷

Alex Pan², David Duvenaud^18,19, Bo Li²⁰

Steve Omohundro²¹, Gabriel Alfour²², Max Tegmark¹⁵

Kevin McGrew²³, Gary Marcus²⁴, Jaan Tallinn²⁵

Eric Schmidt¹⁵, Yoshua Bengio^26,27

¹Center for AI Safety

²University of California, Berkeley

³Morph Labs

⁴University of Michigan

⁵University of Oxford

⁶University of Wisconsin–Madison

⁷Gray Swan AI

⁸Carnegie Mellon University

⁹Cornell University

¹⁰Hong Kong Baptist University

¹¹HKUST

¹²Nanyang Technological University

¹³KAIST

¹⁴University of California, Santa Cruz

¹⁵Massachusetts Institute of Technology

¹⁶University of Tübingen

¹⁷University of Washington

¹⁸University of Toronto

¹⁹Vector Institute

²⁰University of Chicago

²¹Beneficial AI Research

²²Conjecture

²³Institute for Applied Psychometrics

²⁴New York University

²⁵CSER

²⁶Université de Montréal

²⁷LawZero

Introduction

The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today’s specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition.

The framework dissects general intelligence into ten core cognitive domains—including reasoning, memory, and perception—and adapts established human psychometric batteries to evaluate AI systems. Application of this 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.

The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 58%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

GPT-4 and GPT-5 capabilities radar chart

The capabilities of GPT-4 and GPT-5.

Definition

"AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult."

The framework comprises ten core cognitive components, derived from CHC broad abilities and weighted equally (10%) to prioritize breadth and cover major areas of cognition:

Acquired Knowledge

Perception

Central Executive

Output

Citation

@misc{hendrycks2025agidefinition,
      title={AGI Definition}, 
      author={Dan Hendrycks and Dawn Song and Christian Szegedy and Honglak Lee and Yarin Gal and Sharon Li and Andy Zou and Lionel Levine and Bo Han and Jie Fu and Ziwei Liu and Jinwoo Shin and Kimin Lee and Mantas Mazeika and Long Phan and George Ingebretsen and Adam Khoja and Cihang Xie and Olawale Salaudeen and Matthias Hein and Kevin Zhao and Alex Pan and David Duvenaud and Bo Li and Steve Omohundro and Gabriel Alfour and Max Tegmark and Kevin McGrew and Gary Marcus and Jaan Tallinn and Eric Schmidt and Yoshua Bengio},
      year={2025},
}