Artificial general intelligence (AGI)

What the term means

Artificial general intelligence (AGI) is the label used for AI systems that would match or exceed human beings across a broad range of tasks, rather than doing one narrow job well. There is no single agreed definition of the term, which is why it means very different things in research, policy and marketing^[1]. That ambiguity is the starting point for most attempts to make the idea precise: before asking whether a system "is AGI", one has to say what would count.

A framework built from existing definitions

In November 2023, Morris and colleagues published a framework, later published at ICML 2024, that classifies AGI models and their precursors by levels of performance, generality and autonomy^[1]. To build it, the authors analysed existing definitions of AGI and distilled six principles that a useful ontology for AGI should satisfy^[1]. The aim was not another definition but a structure in which different systems and different claims can be placed and compared.

Two axes: depth and breadth

The levels are defined along two axes. Depth means performance relative to skilled adults; breadth means generality across tasks^[1][2]. The two axes can come apart: a system may be extremely strong at a narrow job without being general, or general across many tasks without being very capable at any of them. The matrix formed by the two axes is what lets the framework describe systems that are on the way to AGI, not only those that have arrived.

Six performance levels

Along the depth axis the framework defines six performance levels: No AI, Emerging, Competent, Expert, Exceptional and Superhuman^[2]. Level 1 Emerging means equal to or somewhat better than an unskilled human; Level 2 Competent means at least the 50th percentile of skilled adults; Level 3 Expert the 90th percentile; Level 4 Exceptional the 99th percentile; and Level 5 Superhuman outperforms 100% of humans^[2]. The scale is anchored in comparison with people rather than in test scores, so a level is a statement about where a system's output sits relative to what a skilled adult can do. In a later revision of the paper, Level 4 was renamed from Virtuoso to Exceptional^[1].

Which cells count as AGI

The matrix crosses those levels with breadth. Only the general column of the matrix constitutes AGI, running from Emerging AGI through Competent AGI, Expert AGI and Exceptional AGI to artificial superintelligence^[2]. A system that is Expert level at one narrow task therefore does not qualify; the general label applies to systems whose competence holds across tasks.

Where today's chat models sit

The authors placed frontier chat models such as ChatGPT, Bard, Llama 2 and Gemini at Level 1 General, which they label Emerging AGI, because those systems reach only Emerging level on most tasks^[2]. That placement is a statement about the scale, not a claim that the problem is solved: these systems sit at the bottom of the general column.

Benchmarks, deployment and limits

The framework also discusses what future benchmarks would need to measure these levels, and how the levels interact with deployment considerations such as autonomy and risk^[1]. Autonomy is treated as a dimension separate from capability: how much a system does on its own is not the same question as how well it performs. The main limitation is inherited from the term itself. Because AGI has no single agreed definition, the levels mean what this framework says they mean, and other definitions draw the line elsewhere^[1].