In plain terms
Today's AI is impressive in many areas and unreliable in others; a model can pass a law exam and then miscount the objects in a picture. AGI is the name for a system without that unevenness: one that could learn and do any intellectual job a capable person could. Whether it is near, far or partly here already is argued about, largely because people mean different things by the term.
Why it matters
For a business the term matters mainly as context. It drives the investment, the strategy of the leading labs and much of the regulatory debate, so it shapes the market you buy in. It should not drive your planning. Decisions are better based on what current models demonstrably do in your own tests than on forecasts of arrival dates, which range from a few years to never and come from people with different definitions and different interests.
Example
Two headlines appear in one week: “AGI by 2027, says lab chief” and “AGI is decades away, says leading researcher”. Both can be sincere. The first speaker defines AGI as systems that outperform people at most economically valuable work. The second requires human-like learning from a few examples and dependable reasoning about the physical world. They disagree about the definition as much as about the date.
Most often confused with
AGI vs. Artificial intelligence (AI)
AI is here and in daily use; AGI is a goal. The assistants and agents on sale today are broad in scope without being consistently at human level. Treat any product claim of “AGI” as marketing unless the definition in use is stated.
Origin: The term was popularised in the 2000s by the researchers Ben Goertzel and Shane Legg.
Under the hood
There is no agreed definition or test. Definitions in circulation: performing any cognitive task a human can; outperforming humans at most economically valuable work (the wording of OpenAI's charter); and graded frameworks such as Google DeepMind's levels of AGI, which rate performance and generality separately. Proposed measures include broad benchmark suites, tests of learning new skills from a few examples such as ARC-AGI, and the length of task an AI system can complete on its own. Questions researchers disagree on: whether scaling up current architectures is enough, and how much continual learning, long-term memory, robust reasoning and grounding in the physical world matter. Related terms: artificial superintelligence (ASI), meaning far beyond human level; and transformative AI or powerful AI, which some labs prefer because they describe effects and avoid the dispute over definitions. The prospect of AGI is the main motivation for work on alignment and AI safety.