In The Language of Thought (1975), Jerry Fodor reasoned as follows: our best theories of perception, concept learning and decision describe computational processes; all computation requires a medium of representation; so there is a system of mental representation, mentalese. It cannot be English or Spanish, because infants and other animals think too. It is a sister idea to symbolic AI: the mind as a physical symbol system.
What defines the language of thought is its structure. Its representations have constituents, the meaning of the whole depends on that of the parts, and mental processes are sensitive to that structure: constituents play causal roles (Millière & Buckner, 2024a, sec. 3.1). This would explain the productivity of thought and its systematicity: whoever can think that John loves Mary can think that Mary loves John (Fodor & Pylyshyn, 1988) (see compositionality). The hypothesis also anchored a radical nativism: Fodor went as far as holding that concepts like CARBURETOR are innate (Buckner, 2024, p. 14). For Buckner (2024) this is one axis of the debate between rationalists, who place thought in a symbolic medium, and empiricists, who see it operating on representations of perceptual origin (2024, p. 42).
Neural networks looked like the alternative, and Fodor and Pylyshyn's dilemma awaited them (see connectionism). Today the dispute is empirical. If transformers generalize compositionally, they may be implementing a language of thought or using a non-classical structure; behavior alone cannot decide (Millière & Buckner, 2024a, sec. 3.1). Interpretability offers clues both ways: induction heads store a token in a subspace that works like a variable (Millière & Buckner, 2024b, sec. 2.5.1), and McCoy et al. (2026) find implicit symbolic structures in vectors. Quilty-Dunn et al. (2023) argue that, across several cognitive sciences, the hypothesis is once again the best game in town.
Language models turn Fodor's picture inside out: outside, in their chain of thought, they "think" in natural language; inside, there are only continuous vectors. Where, if anywhere, would their language of thought be?