“Levels of AGI for Operationalizing Progress on the Path to AGI” is a 2023 paper from Google DeepMind, with authors including Meredith Ringel Morris, Jascha Sohl-Dickstein, Allan Dafoe, and Shane Legg. It responds to a recurring problem: the term AGI is used constantly but rarely defined the same way twice, which makes it hard to say whether any given system counts or whether the field is making progress. The paper proposes a structured framework so that claims about AGI can be compared rather than just asserted.
The framework grades systems along two axes. One is performance, or depth - how well the system does a task compared with skilled humans. The other is generality, or breadth - how wide a range of tasks it can handle. Crossing these gives a matrix and a ladder of five levels: Emerging, Competent, Expert, Exceptional (originally called Virtuoso), and Superhuman. A narrow system can be superhuman at one task, like a chess engine, while a general system might only be “Emerging” across many tasks. By the authors’ own placement, the strongest large language models of 2023 sat around “Emerging AGI” - general but not yet reliably competent.
The paper also separates capability from deployment. It argues that how much autonomy a system is given, and at what risk, is a distinct decision from how capable it is, and it sketches matching levels of human-AI interaction. The framework has become a common reference for discussing where systems sit on the road to general intelligence without resorting to a single yes-or-no threshold.
Why business readers should care: vendor claims of “AGI” or “human-level” performance mean little without specifying which tasks and what level of reliability. A depth-and-breadth framing turns marketing language into something you can actually test against your own use cases.