What would an AI have to surpass to deserve the word “general”? If the answer is a collection of tests, somebody still needs to explain what that collection represents. If it is human performance, which humans, doing which tasks and with what assistance? The debate stops being merely semantic when a classification is used to sell a product, justify an investment or demand that institutions change their rules.
Morris and colleagues propose separating breadth and performance, while distinguishing both from deployment autonomy. Their levels-of-AGI framework makes claims comparable that a single label mixes together. The available version was revised on 24 September 2025. It is a proposal for classification and evaluation, rather than a universally agreed definition or an announcement that the destination has been reached. Source: arXiv
The examination has a syllabus too
The distinction helps with a hypothetical tool that drafts contracts brilliantly but interprets images poorly. Averaging those scores may conceal what a buyer needs to know. A law firm is asking about a particular use; somebody claiming to have built general intelligence must justify the breadth of its capabilities. One score can be relevant to the first question and wholly insufficient for the second.
Results also depend on examination conditions. Time, tool access, prior information and opportunities to repeat attempts form part of the capability demonstrated. NIST recommends pairing accuracy measurements with realistic tests representing intended use and documenting the method. Source: NIST The practical consequence is that a comparison loses meaning if one contestant answers once while another has an entire organisation available to correct its work.
That does not invalidate assistance. An institution may prefer a system combining models, search and checks because the complete service works better. It should describe that combination, its costs and its dependencies. If performance disappears when a tool becomes unavailable, that is as important to know as the successful result. An examination can legitimately assess a team; presenting its performance as the attribute of one participant changes what has been demonstrated.
Capability does not confer authority
Moving from solving problems to deciding for other people also requires justification for exercising that authority. A system may compare many alternatives better than an official without being authorised to choose the objectives of public policy. Classifying capabilities helps identify available assistance. Deciding who may act requires knowing mandates, rights, responsibilities and opportunities to challenge an outcome. These remain relevant even when the technical comparison is overwhelmingly favourable to the machine.
This also limits the regulatory use of a scale. Two systems with similar performance can have very different consequences if one recommends while the other executes without review. A modest system can seriously affect many people when it enters a compulsory procedure. Generality matters for anticipating uses; concrete power depends on where a tool is installed, which actions it enables and what other options affected people retain.
The framework's value lies in making explicit differences the debate often compresses. Its harder test will come when evaluations represent those differences and reveal failures outside familiar exercises. A classification can organise disagreement before resolving it. Readers gain an immediate demand to make of the next proclamation of AGI: attach “better” to a task, a comparator and specified conditions. Then there will be something more precise to defend than the prestige of the label.