What Is AGI?

A narrow AI system may perform extremely well at a specific activity, such as classifying images or playing a game. AGI would need to handle a much wider range of tasks and adapt when the setting changes.

Think of a system that learns an unfamiliar scientific tool, plans an investigation, interprets results and communicates its findings. The challenge is not just producing plausible text about each step; it is carrying out the work reliably when the right approach is not already supplied.

Modern general-purpose models complicate a simple narrow-versus-general distinction. They can serve many uses, yet breadth of use alone does not establish robust, human-level general intelligence.

Transfer means applying knowledge or skills beyond the setting in which they were acquired. A general system should do more than reproduce familiar patterns: it should learn a new task, use relevant concepts and notice when an old approach no longer works.

For example, a system might apply a statistical idea learned in forecasting to a new kind of experiment. The useful test is whether it adapts the idea to the new setting, rather than applying a familiar answer indiscriminately.

This is difficult to assess with a finite benchmark. A high score may reflect genuine competence, familiarity with similar training examples or a mixture of the two. Testing unfamiliar situations and failure cases is important alongside measuring average performance.

People have uneven abilities. An expert chemist and an expert negotiator set different standards, and an average person's performance is not the same as the best human performance. Definitions also differ on whether autonomy, efficiency or the ability to learn new skills is required.

One research framework separates generality—the breadth of tasks—from performance—how well a system does them. This is helpful because a system can be broad but unreliable, or extremely capable but specialised.

No single exam settles every version of the question. A useful AGI claim states its definition first and presents evidence that fits it.

AGI is not an agreed status conferred by a single institution. Claims about whether an existing system qualifies depend on the definition and remain contested. This guide does not treat a company's announcement or a benchmark result as settled proof of AGI.

Look for independent evaluation, performance across unfamiliar domains, consistency over repeated attempts and the amount of human assistance required. Reliable long-horizon work is a different test from answering a question in isolation.

Possibly, but not by definition. Superintelligence would substantially exceed the best human cognitive abilities across virtually all relevant domains. AGI, under a human-level definition, would not yet meet that standard.

Research automation and self-reinforcing improvement are proposed routes from one to the other. Each requires assumptions about how much further improvement is possible and what limits it.

Forecasts depend on different views of what current methods can accomplish, which capabilities are missing and how quickly bottlenecks can be overcome. A prediction about a loosely defined milestone is especially hard to interpret.

A date is therefore less informative without a definition, assumptions and a way to assess the prediction. Uncertainty does not mean that progress stops; it means the strength of a forecast should not exceed the evidence behind it.