Language of Thought
Language of Thought
Jerry Fodor's hypothesis (1975) that cognition requires a system of structured, compositional mental representations — "Mentalese" — a language-neutral representational system that natural language expresses but does not constitute. You do not think in English or German; you think in a formal system that English or German partially surfaces.
The Two Core Properties
Fodor argues that any genuine cognitive architecture must exhibit:
Systematicity. If you can think "John loves Mary" you can think "Mary loves John." Mental states come in systematic families: the ability to entertain one state guarantees the ability to entertain its structural variants. A network that happens to represent one but not the other hasn't learned a cognitive system — it's memorized a lookup table. Systematicity is a structural property, not a performance property; it can't be satisfied by getting the right outputs on training cases.
Compositionality. The meaning of a complex mental representation is a function of the meanings of its parts and their structural arrangement. "Brown cow" means what it does because of "brown," "cow," and the syntactic relation between them — not because the system has stored a separate entry for brown cows. This is what makes thought productive: a finite set of primitive representations plus combinatorial rules generates an unbounded space of complex thoughts.
Both properties are trivially satisfied by a Turing-style symbolic system. Fodor's claim is that they are not trivially satisfied by connectionist architectures.
The Fodor-Pylyshyn Critique (1988)
"Connectionism and Cognitive Architecture" (Cognition, 28) is the argument now being re-applied to LLMs. The thesis: a connectionist network can approximate systematic behavior across its training distribution without having the underlying structure that would guarantee generalization. Systematicity in the network's outputs is an accident of training — it will break under distribution shift in ways that a system with genuine LOT structure will not.
The distinction matters. If you train a network to be systematic, it behaves systematically only where the training signal pushed it. If you build a system whose architecture forces systematicity — because compositionality is the computational mechanism — it generalizes by construction. Fodor and Pylyshyn want cognitive architecture of the latter kind.
This is not an empirical point about network performance. It is a point about what kind of explanation you're giving when you describe a cognitive system.
The Internalism Commitment
Fodor was also a committed internalist (Methodological Solipsism): cognitive states are individuated by their internal representational structure alone. Nothing outside the skull is constitutive of cognition. Two systems that are internally identical are cognitively identical, regardless of their environments or causal histories.
This places Fodor in direct opposition to extended-cognition, which holds that cognitive processes can be partly constituted by tools, notebooks, or other agents in the environment. For Fodor, a person relying on a notebook to remember is not distributing their memory — they are using an external storage device to compensate for internal memory limitations. The cognition is still internal; the tool is prosthetic.
Why It Matters Now
The LOT hypothesis generates a specific prediction: any system that exhibits genuine systematic, compositional reasoning must have LOT-like internal structure. LLMs apparently exhibit systematic reasoning. The question is whether they have the structure LOT requires — or whether they are doing something that mimics the outputs of LOT without instantiating the architecture. That question is now empirically live. See lot-llm-paradox.
Key Works
- Fodor (1975) The Language of Thought, Crowell/Harvard
- Fodor & Pylyshyn (1988) "Connectionism and Cognitive Architecture," Cognition 28 — DOI: 10.1016/0010-0277(88)90031-5
- Fodor (2008) LOT 2: The Language of Thought Revisited, Oxford — DOI: 10.1093/acprof:oso/9780199548774.001.0001
- Schneider (2011) The Language of Thought: A New Philosophical Direction, MIT Press
Related
lot-llm-paradox · extended-cognition · reasoning-models · evals