A car takes an alternate route without a driver’s instruction. Is that a sign of general intelligence? Not necessarily. Choosing a route can be a specialized AI capability; artificial general intelligence (AGI) would imply much broader ability to learn, reason, and adapt across tasks. There is no agreed, testable definition of AGI. section 11.1
Table 11.1 Definitions of artificial general intelligence
| Source |
Definition of AGI |
| OpenAI Charter [3] |
“highly autonomous systems that outperform humans at most economically valuable work” |
| Sébastien Bubeck et al., in “Sparks of Artificial General Intelligence: Early experiments with GPT-4” [4] |
“systems that demonstrate broad capabilities of intelligence, including reasoning, planning, and the ability to learn from experience, and with these capabilities at or above human-level” |
| Hal Hodson, in The Economist [5] |
“a hypothetical computer program that can perform intellectual tasks as well as, or better than, a human” |
| Stuart J. Russell and Peter Norvig, in Artificial Intelligence: A Modern Approach, 4th US ed. [6] |
“a universal algorithm for learning and acting under any environment” |
Compare definitions centered on economically valuable work, human-level reasoning and learning, intellectual tasks, or acting across environments. Their differences matter when someone claims a system has reached AGI. section 11.1
Large language models have intensified the debate. Some abilities, such as answering questions across subjects, improved markedly at larger model sizes. A Microsoft research team described GPT-4’s breadth as possible early, incomplete AGI. Neither a surprising ability nor a claim of broad performance settles whether a system meets a definition that researchers have not agreed on. section 2.14 section 11.1
Predictions about when AGI might arrive remain uncertain, and questions about AGI are distinct from questions about machine consciousness or sentience. The debate also has present-day consequences: AGI hype can shape funding and corporate priorities while drawing attention away from accountability, energy use, and who controls AI infrastructure. The practical question is not only how capable future systems might become, but who benefits from the systems being built now. section 11.0 section 11.6 section 11.7
