The phrase comes mainly from Elizabeth Spelke and her colleagues (Spelke & Kinzler, 2007): infants do not start with a blank slate but with a few ancient systems, shared with other animals, that organize experience. Chollet (2019) sums them up as four: objects and elementary physics (objects are cohesive, persist, and act only by contact), agents and goals (some things act with intentions and efficiently), natural numbers and elementary arithmetic for small quantities, and elementary geometry and topology (distance, orientation, inside and outside) (2019, sec. II.1.3).
In AI, the theory works as a list of what is missing. M. Mitchell (2019) locates in this knowledge, intuitive physics and psychology above all, the basis of the common sense that machines lack (2019, ch. 14). Chollet built his ARC test on it: the tasks assume only these priors, with no language or acquired concepts, and quantities stay below about ten (2019, sec. III.1.2); he also proposes that general AI systems have them hard-coded (2019, sec. II.1.3) (see measuring intelligence).
Buckner (2024) contests the scope. First, the concepts the theory requests (object, agent, cause, number, space) are so general that the resulting nativism is moderate (2024, p. 16). Second, the evidence is presented as incontrovertible when it is in dispute, and the replication crisis hit developmental psychology hard (2024, pp. 29, 34). Third, controlled-rearing studies expose newborn chicks and "artificial chicks," generic convolutional networks, to the same stimuli: the networks quickly learn patterns once attributed to innate knowledge, such as object permanence (2024, pp. 290–291).
It is the old debate between empiricists and nativists, now with experiments. If a generic network fed a chick's visual diet develops object permanence, is core knowledge innate knowledge, or the predictable result of a general inductive bias in a structured world?