The first criterion was behavioral: Turing (1950)'s The Imitation Game replaced the question "Can machines think?" with a conversational test. Block (1981) showed its limit with Blockhead, a gigantic lookup table that would pass the test without thinking: the same behavior can have origins that would make us withdraw the attribution.
Later AI adopted another criterion, skill at hard tasks, measured with benchmarks. Chollet (2019) objects that this confuses the product with the process. A human who plays chess at master level is intelligent because she had to acquire that skill through a general ability to learn; a program such as Deep Blue did not get there that way. Hence "there is no task X such that skill at X demonstrates intelligence," unless X is the meta-task of acquiring skills (2019, sec. I.3.1): with enough prior knowledge or enough experience, skill can be "bought" without generalization.
His definition: the intelligence of a system is its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience and generalization difficulty; skill is "merely the output" of that process (2019, sec. II.2.1). He distinguishes local (robustness), broad (flexibility) and extreme generalization (2019, sec. I.3.2). To measure it he proposed ARC: 400 training and 600 evaluation tasks (400 public, 200 private), Raven-style grids with about three examples per task and declared priors close to innate human knowledge (2019, sec. III.1.1). It is the same contrast his textbook draws between the interpolation of networks and the extreme generalization of people (Chollet, 2021, sec. 5.1.2).
Since Cattell, psychometrics has separated fluid intelligence (solving new problems) from crystallized intelligence (accumulated knowledge) (Cattell, 1963). One can know a great deal and reason little; the extreme case is Funes, Borges's character who remembered everything and could therefore barely think, because «pensar es olvidar diferencias, es generalizar, abstraer» ["to think is to forget differences, to generalize, to abstract"; our translation] (Borges, 1942/1944). With LLMs the question is the same: how much of what they do is crystallized intelligence, a smoothie of everything ever written, and how much is fluid? It is another way of stating Chollet's distinction.
The proposal is not neutral: it takes the human scale as its reference and inherits the disputes of psychometrics. Buckner (2024) prefers to assess progress in AI by an ecological rationality, precisely because of those controversies (2024, pp. 77, 81). But there is agreement on one point: a purely behavioral criterion that ignores learning efficiency and internal representations cannot tell apart two systems that do the same thing in different ways (Millière & Buckner, 2024b, sec. 4.1). And every test faces Clever Hans: success may come from unintended cues or from data contamination.