Artificial general intelligence, or AGI, is a machine that can learn and carry out any intellectual task a human being can, across any domain, without being purpose-built for each one. The defining word is general. A system that plays chess at superhuman level is not AGI. A system that learns chess because it was asked to, then applies the same machinery to law, medicine or logistics, is closer to the idea.
What separates AGI from today's AI
Current systems are trained for a task distribution. They are extraordinary within it and brittle outside it. An AGI would carry its own ability to acquire new skills: it would encounter a problem it has never seen, work out what it needs to learn, learn it, and act. That capacity for transfer is the line most researchers draw.
What would count as AGI
A system would need, at minimum, the ability to reason across unfamiliar domains, retain and use long-term memory, plan over long horizons, learn continuously after deployment, and operate reliably enough that its output can be trusted without a human checking every step. Missing any one of those leaves a system that is impressive but narrow.
Why the definition matters
Two people arguing about AGI often disagree about whether it has arrived, not about the technology. One counts capability, another counts autonomy, a third counts understanding. Agreeing on the test first is the only way the timeline question becomes answerable.