What it means
Every AI system in production today is narrow: extremely capable within a domain and useless outside it. A model that writes excellent code cannot drive a car, and a system that plays Go at superhuman level knows nothing else. Artificial general intelligence describes a system that transfers competence across domains the way a person does — learning a new task from a description, applying knowledge from one field to another, handling situations nobody anticipated.
No such system exists, and there is no agreement on what would count as one. Definitions range from "outperforms most humans at most economically valuable work" to far stricter tests involving autonomy and self-directed learning. That definitional gap is why credible researchers give wildly different timelines while appearing to discuss the same thing.
Why it matters
AGI does a lot of rhetorical work in AI discourse, on both sides. It is invoked to justify enormous investment and to argue for urgent regulation, and it is invoked to dismiss present-day systems as mere pattern-matchers. Both moves lean on a term with no agreed meaning. For anyone making practical decisions, the more useful question is what today's narrow systems can reliably do — which is a lot, and is measurable.
In practice
Treat any confident AGI timeline — near or far — as a claim about the speaker's definition as much as about technology. When you see a prediction, the first question worth asking is what test the speaker thinks would settle it.