A machine able to achieve goals across a wide range of environments, not just the one it was trained on — the defining line between “narrow AI” and something categorically different.
“AGI” gets used in both marketing copy and formal research as though it names one clear, agreed-upon threshold. It doesn’t. Every current AI system, including the large language models behind the present hype cycle, is narrow: trained and evaluated on specific distributions of tasks, without the general, transferable problem-solving capacity the term is meant to describe.
The most cited attempt to formalize the term comes from Shane Legg and Marcus Hutter, who defined machine intelligence as an agent’s ability to achieve goals across a wide range of environments — deliberately environment-agnostic, so it doesn’t smuggle in human-specific assumptions about what “general” means. Formalizing the definition this precisely also exposes how far current systems are from meeting it: benchmark performance on a fixed task set is not the same as the open-ended, cross-domain competence the definition requires.
In this POV’s framework, AGI hype is a symptom of Peak AI rather than a technical inevitability — see also Intelligence for why treating any predictive system as a general intelligence, achieved or impending, risks the same category error regardless of scale.
Source: Legg, S. & Hutter, M. (2007), “Universal Intelligence: A Definition of Machine Intelligence,” Minds and Machines, 17(4), 391–444.