Language of Thought
The Language of Thought (LOT) hypothesis, developed most fully by Jerry Fodor, is the claim that thinking is conducted in a mental language — a system of syntactically structured mental representations with a combinatorial semantics analogous to natural language or formal logic. On this view, beliefs, desires, and other propositional attitudes are relations between an organism and mental sentences in an internal code. The code is not English or Chinese; it is a proprietary symbol system, presumably innate, whose expressions have both syntactic structure (they can be combined and transformed by rules) and semantic content (they refer to states of affairs in the world).
The hypothesis is the philosophical foundation of the classical version of the Computational Theory of Mind. If mental states are computational states, and computation operates on symbol structures, then mental representations must be symbol structures. The Language of Thought is the postulated symbol system: the "machine language" of the mind.
The Argument from Productivity and Systematicity
Fodor's central argument for LOT appeals to two properties of thought: productivity and systematicity.
Productivity. Human thinkers can entertain an unbounded (or at least very large) number of distinct thoughts. We can think "the cat is on the mat," "the cat is on the mat and the dog is in the yard," "the dog is in the yard or the cat is not on the mat," and so on, recursively. If thoughts were unstructured associations — mere connections of ideas, as associationism holds — there would be no explanation for why the capacity to think some thoughts entails the capacity to think others. The productivity of thought is explained by the combinatorial nature of a symbolic system: finite primitives generate infinite combinations.
Systematicity. The capacity to think one thought is systematically connected to the capacity to think related thoughts. Someone who can think "John loves Mary" can think "Mary loves John." Someone who can think "the red ball" can think "the red square" and "the blue ball." This systematicity is explained by the constituent structure of mental representations: the same symbols (John, Mary, loves) recombine according to the same syntactic rules. Associationist and connectionist architectures struggle to explain systematicity without implementing something structurally equivalent to a symbol system — a point Fodor and Pylyshyn pressed forcefully in their 1988 critique of connectionism.
Challenges and Revisions
The LOT hypothesis faces multiple challenges. The symbol grounding problem — how mental symbols acquire semantic content — is the most discussed. If mental representations are arbitrary symbols, what connects them to the world? Fodor's response, in his later work, was to embrace a form of semantic externalism: the content of mental symbols is determined by their causal-historical connections to properties in the environment, not by any internal grounding mechanism. This makes content determination a diachronic, historical fact rather than a synchronic structural one — a move that some critics find unsatisfying.
The nativism problem. LOT requires that the entire inventory of primitive mental symbols — the basic vocabulary of the language of thought — be innate. Fodor embraced this consequence with notorious boldness: "There had better be quite a lot that is innate, or there could be no learning at all." But the claim that the concepts CARBURETOR, QUARK, and DEMOCRACY are innate (in some sense) strikes many as implausible. Fodor's defense — that nativism is about conceptual possession, not conceptual activation — has not fully quieted the objection.
The connectionist alternative. Connectionist and neural network models of cognition do not posit structured symbolic representations. Instead, they explain thought as patterns of activation across distributed networks. The debate between symbolic and subsymbolic approaches has not been resolved; it has bifurcated. Classical AI and cognitive science continue to use symbolic representations. Deep learning and neuroscience work with distributed, vectorial representations. And hybrid approaches — neural-symbolic integration, tensor-product networks, holographic reduced representations — attempt to combine the systematicity of symbols with the learning capacity of neural networks.
The Systems-Theoretic Angle
From a systems perspective, the LOT debate is not about whether there is a "language" in the head. It is about whether cognitive systems require compositional structure to achieve the behaviors we call thinking. Compositional structure — the property that the meaning of a whole is determined by the meanings of its parts and their mode of combination — is what makes productivity and systematicity possible. Whether that structure is implemented as explicit symbols, as distributed vectors with compositional operations, or as dynamical attractors with separable dimensions is an implementation question, not a metaphysical one.
The deeper systems question is: what kind of architecture can support both combinatorial generalization (recombining known elements in novel ways) and continuous learning (adapting to new data without catastrophic forgetting)? Symbolic systems excel at the first but struggle with the second. Neural systems excel at the second but struggle with the first. The Language of Thought hypothesis identifies a real desideratum — compositional structure — but may be too committed to a particular implementation. The systems challenge is to find architectures that satisfy the desideratum without the liabilities.