AGI (artificial general intelligence)

Definition

AGI (artificial general intelligence) refers to an AI able to learn and succeed at almost every intellectual task a human can do, without being designed for each one. It is the opposite of “narrow” AI: an unbeatable chess program cannot fill in a tax return.

The word is simple; the definition is not. Each player has its own:

• OpenAI (2018 charter): highly autonomous systems that outperform humans at most economically valuable work. A definition centred on work.
• Google DeepMind (2023): a scale with several levels, from “emerging” to “superhuman”, that crosses performance with the breadth of tasks. For its authors, AGI is not a switch but a progression.
• Dan Hendrycks and a large group of researchers, including Yoshua Bengio (2025): an AI that matches the cognitive versatility and proficiency of a well-educated adult, measured on ten abilities inspired by human psychology tests. On this grid, GPT-4 scored 27% and GPT-5 57%, with a marked weakness in long-term memory.

Why everyone is talking about it

AGI is the stated goal of several big labs, including OpenAI (it is written into its charter) and Google DeepMind. It serves as a compass for massive investment in data centres. Its supporters see it as a historic break: digital “workers” available at scale and faster scientific discovery. That is also what worries people: a general, autonomous AI that is poorly aligned with our intentions would pose risks of a different order.

Where things stand (October 2026)

• Claims are multiplying. On September 3, 2026, presenting GPT-6 Astra, OpenAI president Greg Brockman said he thought this model might mark the arrival of AGI, and closed with “Welcome to the AGI era”. He did, however, leave it to users to judge.
• Independent checks are more nuanced. The ARC Prize Foundation, whose ARC-AGI-3 test was cited in support, points out that even maxing out its test would not prove AGI. On its neutral test setup, Astra's score drops sharply: 62.7%, versus 99.9% with the tooling OpenAI built for its model. Other evaluators see no across-the-board leap over competitors.
• Timelines are shrinking. In a Wall Street Journal interview reported on October 21, 2026, Demis Hassabis (Google DeepMind) estimated that AGI could arrive within two to three years, against three to five earlier in the year. Other researchers, such as Yann LeCun, think today's language models will not get there.
• The limits are still visible. Current models excel at code, mathematics and knowledge, but remain uneven: long-term memory, continual learning, reliability on long tasks in the real world.

In other words: progress is real and fast, but there is no scientific consensus that AGI has been reached. When a company says it has, it is judging by its own definition.

How to read an “AGI” announcement

• Which definition? The company's own, or an independent grid?
• Which test, and run by whom? A score obtained by the vendor, with its own tooling, is not worth an independent evaluation.
• In the real world? Solving puzzles or passing exams does not guarantee running a multi-week project alone without mistakes.
• Who benefits from saying it? The word attracts investors, customers and talent: a company announcement is not scientific proof.

Don't confuse

• AGI and superintelligence: AGI matches humans in most areas; superintelligence would far exceed them in almost all.
• AGI and AI agents: an agent chains actions autonomously on a given task. It can be very useful without being “general”.
• AGI and consciousness: a general AI would not necessarily be conscious. These are two separate questions.