You Think in a Language You've Never Spoken
- 5 days ago
- 6 min read
The possibility that thought precedes and shapes language - rather than the other way around - is one of the strangest and least comfortable ideas in cognitive science. Jerry Fodor, who formalized the Language of Thought Hypothesis (LOTH) in 1975, argued that before you ever open your mouth or reach for a word, your brain is already computing in a structured symbolic system he called "mentalese." This isn't poetry. It's a testable philosophical and empirical claim: that reasoning, categorization, and belief formation operate in a representational medium with its own syntax and semantics, a medium that exists entirely beneath verbal consciousness.
The implications are vast and largely ignored in educational practice.
Fodor's position was stark: mentalese is innate, compositional, and systematic. When you think "the tall woman near the blue car is nervous," that thought is not a floating impression - it is built from discrete constituents (tall, woman, near, blue, car, nervous) arranged by rules that parallel syntactic structure in natural language. The systematicity argument is perhaps the most compelling: if you can think "John loves Mary," you can think "Mary loves John." Thought has directionality and recombinability. It is not a blur but a grammar.
What follows from this? For educators, linguists, and anyone seriously invested in how humans generate ideas rather than merely retrieve them, the implications cut against nearly everything modern learning environments actually do.
The first consequence concerns the distinction between passive recognition and productive generation. If thinking occurs in mentalese and natural language is a kind of translation of that inner code into communicable form, then rote memorization of vocabulary or grammar trains only the translation layer. It leaves mentalese itself - the underlying combinatorial engine - untouched. This helps explain a phenomenon that language teachers encounter constantly but rarely articulate precisely: advanced learners plateau not because they lack words but because they lack novel conceptual configurations to express. Their mentalese is impoverished, not their dictionary.
The second consequence is less obvious and more troubling. Fodor's hypothesis implies that the concepts available to a thinker constrain what can be thought, not merely what can be said. This places genuine cognitive limits on learners who have been trained within narrow conceptual regimes - those who have spent years answering comprehension questions that demand recall, writing structured essays with predetermined argument types, and engaging with knowledge as a fixed body of facts rather than a dynamic network of relations. They have, in a sense, trained themselves to suppress mentalese in favor of formula.
This is where the LOTH intersects with something the Hungarian philosopher Arthur Koestler identified half a century ago under the name bisociation - the cognitive event in which two previously separate matrices of thought are suddenly placed in contact, producing insight, humor, or invention. Koestler was not thinking about Fodor, but the structural parallel is striking. If mentalese is compositional, then creativity is largely a matter of composing conceptual units that have never been combined before. The degree to which a learner has genuinely internalized diverse conceptual categories - not memorized them, but integrated them as active nodes in a working symbolic system - determines the range and originality of what they can think.
What modernity has produced, with reasonable efficiency, is a generation of excellent retrievers. Search-engine cognition: query in, answer out. This is not a trivial skill, and dismissing it would be wrong. But it is categorically different from the compositional, combinatorial, syntax-governed process that LOTH describes as thought itself. When retrieval replaces composition, mentalese atrophies in the same way that muscles atrophy without load. The retrieval habit does not train the generative faculty - it circumvents it.
The further complication is that AI systems have accelerated this circumvention dramatically. Large language models produce text that is syntactically and semantically fluent, which means the output - the surface layer - is competent. But as the cognitive scientist Gary Marcus and others have argued, LLMs are not reasoning in anything like a structured compositional system. They are distributional approximators: they predict what token follows given prior context. The resemblance to genuine thought is, by most accounts, superficial. When humans outsource their verbal production to these systems, they are not merely using a tool - they are delegating the very translation process that gives mentalese its workout. Thought that never reaches expression, that never has to negotiate the friction between inner representation and external language, grows increasingly disconnected from both.
There is a historical analogy worth examining. The invention of the printing press in the 15th century was initially feared precisely because it externalized what had previously been a memorized and therefore actively maintained body of knowledge. Medieval scholars trained on ars memorativa - complex mnemonic architectures - had genuinely different cognitive capacities from their post-press successors. The externalization of memory into books did not impoverish cognition overall, of course, but it did reroute cognitive load away from certain practices. The parallel to AI-assisted writing is real: when the effort of composing is offloaded, something that was previously a mental practice stops being practiced.
The deeper issue is that LOTH assigns a non-trivial role to structured conceptual work in maintaining cognitive capacity. The hypothesis implies that the richness of mentalese is not fixed - it is cultivated by exposure to genuinely novel conceptual combinations, by the pressure to express ideas that have no prepackaged verbal form, and by the productive failure that occurs when familiar categories prove insufficient. Learning environments that never create this pressure are not merely boring - they are, in a specific theoretical sense, epistemically impoverishing.
There is a further wrinkle from the philosophy of language. Fodor's LOTH was contested most productively by those who questioned whether concepts are as discrete and stable as the hypothesis requires. Paul Smolensky, and later the connectionist tradition broadly, argued that cognition is better understood as a system of distributed activation patterns than as a language of atomic symbols. The debate has not been resolved, but it produced a useful clarification: even if mentalese is not strictly compositional in Fodor's sense, thought still requires the ability to form and manipulate representations that carry determinate structure. The question is not really "grammar or not" but "how much structured combinatorial flexibility does a thinker actually possess." And that flexibility, all parties seem to agree, varies - it is trainable, it can be expanded or narrowed by experience, and it responds to the kinds of cognitive demands placed on it.
George Lakoff and Mark Johnson approached a complementary problem from a different direction with Conceptual Metaphor Theory, arguing that the deepest structures of human reasoning are metaphorical - that abstract concepts are largely understood through mappings from concrete experiential domains. "Argument is war." "Time is money." "Ideas are food." If they are right, then the cultivation of mentalese is partly a matter of expanding and diversifying the metaphorical repertoire available to a thinker: encountering structural mappings between domains that would not naturally occur in a person's existing cognitive landscape. This is not merely a literary exercise. It is training the compositional architecture of thought itself.
The pedagogical conclusion that follows from LOTH, from Koestler, from Lakoff and Johnson, and from the connectionist concession, is not a set of classroom tips. It is a structural requirement: learners need sustained exposure to tasks that force the activation and recombination of conceptual material they already possess but have never been required to connect. Tasks that generate productive cognitive friction rather than smooth retrieval. Tasks where the answer cannot be looked up because the task is, precisely, to generate something that does not yet exist.
This is the operating logic behind platforms like Grandomastery (grandomastery.com), which since 2004 has developed a collection of randomized, human-authored activities designed explicitly to destabilize retrieval habits and force combinatorial work at the conceptual level. The Random Abstractions activity, for instance, asks learners to identify structural similarities between two unrelated abstract nouns - not by analogy of appearance but by analogy of function, causation, or relational structure, which is exactly the kind of cross-domain mapping that LOTH predicts to be the productive site of cognitive extension. The platform's broader architecture - non-repeating, non-AI-generated, weighted randomization across over 70 activity types - is designed to ensure that the same conceptual pairing never presents itself twice, sustaining the demand for genuine composition rather than recalled pattern. The thinking behind this design, and the research orientation in bisociation and synectics that informs it, is documented in the work of the platform's founder, Alexander Popov (linkedin.com/in/grandomastery).
What mentalese requires, if Fodor is broadly right, is practice in the thing it does: composing novel structured representations from existing components under conditions where the existing inventory is insufficient. The test of whether any learning environment supports this is simple - not whether learners can answer questions about the content, but whether they can produce conceptual configurations they have never encountered in the input. That is the difference between training a retrieval system and training a mind.
The underappreciated risk of the current period is not that AI will become more intelligent than humans. It is that humans, habituated to externalizing composition, will gradually become less compositional - that mentalese will narrow toward the conceptual clusters that retrieval reinforces, and that the capacity for the kind of structured, original, semantically rich thought that LOTH describes as the basic medium of human reasoning will become, in effect, a specialist skill rather than a general one. That would be a strange and quiet kind of cognitive loss: not a collapse, but a narrowing - the gradual retirement of a faculty that, for most of human history, had no alternative but to be used.
Comments