What is Artificial General Intelligence?
Also called AGI.
Artificial general intelligence is a hypothetical form of artificial intelligence able to learn and perform the full range of intellectual tasks that people can perform, rather than excelling only in narrow domains. The term is contested, because researchers disagree about what counts as general capability and about how it would be measured. No agreed definition or accepted test exists.
Definitions vary widely. Some framings emphasize breadth, meaning competence across arbitrary tasks without task specific retraining. Others emphasize autonomy and the ability to set and pursue goals, or use economic framings that describe the ability to carry out most work that can be done remotely. Because these criteria do not coincide, two people can reasonably disagree about whether a given system qualifies.
The disagreement is partly philosophical. Whether a system that produces appropriate behavior therefore understands anything has been debated since the earliest days of the field, in arguments such as the Chinese room thought experiment. Behaviorist positions treat observable competence as sufficient, while other positions hold that internal states, grounding in the world, or consciousness matter. This entry takes no position.
Measurement is unsettled. Benchmarks are designed for specific capabilities and saturate quickly, and strong results can reflect contamination of the training data rather than general ability. Proposed alternatives include tests of abstraction and of skill acquisition efficiency, but none is accepted as decisive. Claims that a system has or has not achieved general intelligence usually rest on the speaker's chosen definition.
Estimates of when such systems might exist vary by orders of magnitude among qualified researchers, and this entry makes no forecast. Discussion of general intelligence is closely tied to work on alignment and governance, since capabilities that generalize broadly would also generalize the consequences of misuse or of objectives that do not match the intentions of the people deploying them.
Key points
- Hypothetical AI with broad, human level intellectual competence.
- Definition is contested and no accepted test exists.
- Behavioral competence versus understanding is a live dispute.
- Benchmarks saturate and can be contaminated by training data.
- Expert estimates of timing vary enormously.
In practice
Consider a system that writes competent code, passes professional examinations, and drafts legal summaries. One observer calls this general intelligence because the breadth of tasks is wide. Another says it is a collection of narrow abilities learned from text, since the system cannot reliably acquire a genuinely new skill from a few trials. Both are applying different, defensible criteria.