Artificial General Intelligence, usually abbreviated to AGI, refers to an AI system capable of performing, at human level or above, any cognitive task that a human can perform. The emphasis is on generality: today's AI systems are narrow, meaning they do one kind of thing well, whether that is generating text, recognising images, or playing chess. An AGI would not be limited to a domain. Asked to write a legal brief, compose music, diagnose a medical condition, or prove a theorem, it would approach each with the competence of a capable human specialist.
The definition is actively contested. Some researchers set the bar at matching the best human performance in every field. Others define it in terms of the ability to learn new tasks from minimal examples, the way humans do. Ray Kurzweil's 1999 prediction placed AGI at 2029; he has since described that estimate as conservative. Several AI laboratories including OpenAI and Anthropic have named AGI as a primary goal, though their internal definitions differ. The lack of a single agreed definition makes the debate somewhat circular: the milestone may be claimed before consensus that it has been reached.
What makes AGI qualitatively different from current AI is not raw capability but adaptability. A large language model trained today is extraordinary at the tasks in its training distribution and unreliable outside it. An AGI, by definition, would transfer knowledge and reasoning across unfamiliar domains the way humans do when they encounter genuinely new problems. That transfer is currently what neither the most capable systems nor their designers fully understand how to build reliably.
For a reader trying to follow the discussion without a technical background: AGI is the threshold after which the word 'tool' becomes inadequate as a description. Current AI is a very powerful tool. AGI would be something more like a colleague: capable of independent judgment, genuine problem-solving in new contexts, and potentially of improving its own capabilities. Whether that turns out to be good, dangerous, or simply strange is what most of the serious AI safety debate is actually about.
