“Artificial general intelligence” is used to mean everything from a better chatbot to a world-changing superintelligence. In November 2023, a team at Google DeepMind proposed a way to make the term more precise. Their paper, Levels of AGI for Operationalizing Progress on the Path to AGI, borrows an idea from self-driving cars: instead of asking whether a system is AGI, rate how capable and how general it is, on a ladder of levels.

Why a framework?

The authors reviewed many existing definitions of AGI — the Turing test, consciousness-based definitions, “most economically valuable work”, flexible learning ability and others — and found each lacking in some way. A shared vocabulary, they argued, would help researchers compare models, policymakers assess risks and the public understand claims.

The inspiration was the SAE levels of driving automation, which describe cars from Level 0 (no automation) to Level 5 (full automation) and became a common language for the autonomous-vehicle industry.

Six principles

The paper proposes that a useful definition of AGI should:

  1. Focus on capabilities, not processes (how the system works internally, or whether it is conscious).
  2. Consider both generality and performance.
  3. Focus on cognitive and metacognitive tasks, not physical ones such as robotic manipulation.
  4. Look at potential, not deployment — what a system can do, not whether it is used.
  5. Emphasize ecological validity: tasks that people actually value.
  6. Focus on the path to AGI, not a single endpoint.

The levels

The framework crosses two dimensions. Performance describes how well a system does compared with skilled adults:

  • Level 0 — No AI
  • Level 1 — Emerging: equal to or somewhat better than an unskilled human
  • Level 2 — Competent: at least as good as the 50th percentile of skilled adults
  • Level 3 — Expert: at least the 90th percentile
  • Level 4 — Virtuoso: at least the 99th percentile
  • Level 5 — Superhuman: outperforms 100% of humans

Generality distinguishes narrow systems, which excel at a specific task, from general systems, which perform across a wide range of non-physical tasks.

This produces examples on both sides. Narrow superhuman systems already exist — AlphaFold for protein structure prediction, or AlphaZero for chess and Go. In the 2023 paper, the authors classified the leading chatbots of the time as “Emerging AGI”: general, but performing at Level 1 across many tasks.

Autonomy is separate

The paper adds a second ladder for autonomy: from AI as a tool controlled by people, to a consultant, collaborator, expert and fully autonomous agent. A capable system does not have to be deployed autonomously, and the risks depend on both capability and how much independence it is given.

Benchmarks needed

The authors stress that their levels require measurement. An ideal AGI benchmark would cover a broad, diverse and evolving set of cognitive tasks, including those requiring learning new skills, and should be “living” so it can be updated as systems improve.

Criticism and debate

  • Human percentiles are hard to measure for many tasks, and “skilled adults” vary widely by field.
  • Task selection determines the result: which tasks count as part of “general” intelligence?
  • Excluding physical tasks leaves out abilities many consider essential to intelligence.
  • Levels can create false precision, giving a scientific look to judgments that remain subjective.

Still, the framework has influenced how labs and commentators talk about progress, shifting debates from “is it AGI?” to “how capable, how general, and how autonomous?”

Key takeaways

  • Google DeepMind proposed rating AI by performance and generality instead of a yes-or-no AGI label.
  • Performance levels run from Emerging (Level 1) to Superhuman (Level 5), measured against skilled adults.
  • Narrow systems such as AlphaFold can already be superhuman; general chatbots were rated “Emerging AGI” in 2023.
  • Autonomy is assessed separately, because risk depends on how independently a system is deployed.
  • The framework still depends on benchmarks that do not yet fully exist.

Frequently asked questions

Which level are today’s chatbots?

In the 2023 paper, the authors placed leading chatbots at Level 1, “Emerging AGI”. Placement depends on measurements the framework calls for but does not fully provide, so any current rating is a judgment rather than a certified score.

Why compare AGI to self-driving levels?

Because the SAE levels gave the automotive industry a shared vocabulary for gradual progress. The DeepMind authors wanted a similar common language for AI, replacing a single yes-or-no AGI threshold.

Does a higher level mean more risk?

Not by itself. The paper treats autonomy separately: a highly capable system used as a tool under human control may pose fewer risks than a less capable system acting autonomously.