AGI explained: what it means and what it does not

Learn in your browser · Chapter 01 of 06

Everyone uses the letters AGI. Not everyone means the same thing by them.

AGI stands for artificial general intelligence. The word 'general' is the key. It points to AI that is broadly capable across many kinds of tasks, instead of being built for one. Beyond that, agreement runs out. Researchers, companies and commentators have proposed several different definitions, and each one stresses something different.

Scroll sideways to see every column.

Two things follow. First, the definitions measure different things, so the same system can count as AGI under one and not under another. Second, some definitions describe what a system can do, while others describe what it is. Most definitions in the research literature focus on capability, because capability can at least be tested, while consciousness cannot yet be settled.

Learn in your browser · Chapter 02 of 06

Instead of a yes or no answer, one published paper proposes levels, so that people can say more precisely where a system stands.

In November 2023, Meredith Ringel Morris and seven co-authors, all researchers at Google DeepMind, posted a paper titled 'Levels of AGI: Operationalizing Progress on the Path to AGI'. It was later presented as a position paper at the 2024 International Conference on Machine Learning. It has since been revised, and the framework below follows the September 2025 revision.

The authors set out six principles that a good definition should meet, including focusing on capabilities rather than on how a system works, on potential rather than deployment, and on the path to AGI rather than a single endpoint. They then classify systems along two dimensions: performance (depth: how a system compares with people at a task, measured against skilled adults above Level 1) and generality (breadth: a narrow set of tasks, or a wide range).

Scroll sideways to see every column.

For general systems, the same levels apply across a wide range of non-physical tasks, including learning new skills. Level 1 general is called Emerging AGI, and Level 5 general is called artificial superintelligence, or ASI. Level 4 was originally called Virtuoso and was renamed Exceptional in a later version. As of September 2023, the authors judged that the leading language models were Competent at some tasks but Emerging for most, which made them Emerging AGI overall. That was one judgement at one date, not a current ranking, and the authors say such placements are approximate.

The paper also separates capability from autonomy, with six levels that run from no AI, through AI as a tool, a consultant, a collaborator and an expert, to AI as an agent. Its point is that a highly capable system does not have to be given free rein: people can choose how much autonomy to allow.

Learn in your browser · Chapter 03 of 06

Most AI is narrow. The interesting cases are the ones that are broad but uneven.

Narrow AI is built or trained for one task, or a clear set of tasks. It can be far better than any person at that task and useless outside it. General AI handles a wide range of tasks, including ones it was not specifically prepared for. Modern chatbots sit in an unusual place. They are broad, and they can be excellent at some tasks and unreliable at others, sometimes within the same conversation.

That unevenness is one reason claims are contested. There are others.

• No agreed definition or test, so the argument is often about words.
• Incentives: companies, investors, critics and the media each have reasons to push a view. That does not make them wrong, but it is worth noticing.
• Benchmarks may not reflect real-world reliability, as the workbook Which AI model should I use? explains.
• Shifting goalposts: when AI masters something once thought to need intelligence, people often decide it was never the real test, so the target drifts.
• Uneven ability makes a simple yes or no hard to give.
• Different concerns: some people care about capability, some about consciousness, some about autonomy or economic impact.

Learn in your browser · Chapter 04 of 06

Headlines about AGI mix four different kinds of statement. Learn to separate them and most of the confusion clears.

Scroll sideways to see every column.

Timelines deserve special care. Forecasting technology is hard, informed people disagree widely, and a date means nothing without a definition attached. Treat any single date as one opinion among many, and be wary of anyone who is certain, in either direction.

1. Who is speaking, and what do they stand to gain?
2. Which definition of AGI is being used?
3. Is this a measured result, an interpretation, a forecast or a definition?
4. What exactly did the system do, and has anyone independent checked it?
5. Is the claim about capability, about deployment or about risk?
6. What evidence would change the speaker's mind?

Learn in your browser · Chapter 05 of 06

A few neighbouring words are constantly confused with AGI. Then we turn to what matters whatever label is used.

Scroll sideways to see every column.

Whether or not a system is called AGI, four questions matter. Capability: what can it actually do on real tasks? Reliability: how consistently, including on unusual input? Safety: what harm could it cause through mistakes, misuse or misplaced trust? Governance: who is accountable, what rules apply and who checks? A system does not need to be AGI to be valuable, or to cause harm.

Regulation already exists, and it works by risk and use rather than by asking whether something is AGI. The European Union's AI Act has been rolling out in stages. Its high-risk obligations were deferred by a simplification package known as the AI Omnibus, politically agreed in May 2026 and in force since 27 July 2026.

Scroll sideways to see every column.

Learn in your browser · Chapter 06 of 06

A working definition is a tool for thinking. It lets you say precisely what you mean, and test it.

Pick the parts that matter for your purpose. How broad must the tasks be? Compared with whom, and at what level? How reliable? How much help or supervision is allowed? Then say how you would test it. Here is a template.

For my purposes, a system counts as general intelligence when it can
do [which tasks]
at least as well as [which people, at what level]
reliably enough that [how I would test it]
without needing [what help or supervision].
It does NOT require [what I am leaving out, for example consciousness].

Well done. You can now read AGI claims with care, without hype or panic. Next, the AI agents workbook looks at systems that act on your behalf, and the workbooks on reasoning models and on evaluating frontier models show how to judge what new systems can really do.