The Idea of a Language of Thought | The Substrate of Intelligence, Part 3

From Aristotle and Ockham to Descartes, Whorf, Vygotsky, and Fodor’s mentalese, the two-thousand-year argument over whether thinking needs words.

tl;dr — Around 350 BCE Aristotle said a Greek and a Persian make different sounds but share one thought underneath. Two thousand years later Humboldt looked at the same two speakers and said they do not think the same thing at all, because each language carves the world its own way. That is the whole fight, fully formed before anyone built a transistor. This part traces its two bloodlines: the thought-first line from Aristotle’s affections of the soul through Ockham’s medieval mental language and Leibniz’s dream of settling disputes by calculation to Fodor’s mentalese, and the language-forms-thought line from Herder and Humboldt through the behaviorists to Sapir and Whorf. Chomsky’s 1959 review of Skinner is the hinge that reopened the mind; Vygotsky against Piaget is the developmental argument still unresolved. The verdict the twentieth century actually reached is that the strong claims died, and what survived is language as a powerful tool that scaffolds a thinking system it did not create, sitting on a compositional core that was already running before any words arrived. One old objection still cuts, Wittgenstein’s point that a private mental code cannot mean anything on its own, and it forces that core to be grounded. The fight you are having about large language models is the one Ockham and Humboldt were already having, and knowing that is the difference between repeating it and advancing it.

Complimentary Reading Link

In the 1680s, a German polymath had a dream that would sound insane for another two hundred and fifty years. Gottfried Wilhelm Leibniz proposed that all human reasoning could be reduced to a formal symbolic calculus, so that when two people disagreed, they would not argue. They would compute. “Let us calculate,” he wrote, and the dispute would settle itself the way an arithmetic problem does, by mechanical procedure over the right notation. He called the notation the characteristica universalis, a universal character in which every concept had a precise symbolic form and every valid inference was a rule you could apply without insight, the way a clerk adds a column of figures.

Leibniz never built it. He could not have. But that dream, that thinking is computation over structured symbols and could in principle run on a machine, is the direct ancestor of symbolic artificial intelligence, of formal logic, and of the argument now consuming the frontier of AI. When people fight today about whether a large language model that produces a fluent chain of reasoning is actually reasoning or only arranging tokens that look like reasoning, they are re-fighting a battle that Leibniz opened and that Aristotle opened before him. The combatants mostly do not know this. They think the question was born with the transformer.

It was not.

This is the third installment of a series about where intelligence actually lives and in what form, and this part makes a single historical claim: both of the camps now spending hundreds of billions of dollars on rival architectures were fully formed long before neuroscience or computing arrived to test them.

Part 1 cleaned up the vocabulary, separating the format of conscious experience from the format of communication from the substrate the brain actually computes over. Part 2 laid the modern field out as a map, eleven schools plotted on two axes. This part supplies the genealogy underneath that map. Two lineages run through two thousand years of Western thought about the mind, and almost every current position is standing in a spot that Ockham or Humboldt cleared centuries ago.

Two lineages, one argument

The disagreement has a clean shape once you name the sides.

The thought-first lineage holds that thinking is prior to and independent of any particular spoken language. On this view, a thought has structure and content before it is dressed in words, and the words of English or Arabic or Mandarin are a local encoding laid over a universal internal system. Language expresses thought. It does not constitute it.

The language-forms-thought lineage runs the other way. It holds that language is the loom a thought is woven on rather than a garment thrown over a finished one. Learn a different language and you carve up reality differently, remember differently, reason differently. In its strongest form, this becomes the claim that you cannot think what your language gives you no words for.

Both of these were argued to a high polish before anyone could put a person in a scanner or train a network on a corpus. That is the point worth sitting with. The evidence changed. The positions did not. When I described in Part 1 the two questions this series turns on, whether the substrate matters and whether there is a critical mass, I was naming distinctions that the ancients and the moderns were already circling. The whole fight between the thought-first line and the language-forms line is, underneath, a fight about spine question one: is intelligence tied to the format of language, or is language one channel feeding something prior to it. Hold that thought. The history states the question with unusual clarity precisely because it had no data to hide behind.

The thought-first line: from affections of the soul to mentalese

Start where the West starts, with Aristotle.

In On Interpretation, he separated spoken and written words from what he called the affections of the soul, the mental contents that words signify, and he made an observation that is doing quiet work at the center of this entire series [1]. Spoken words differ from one language to the next. The affections of the soul they signify are the same for everyone. Aristotle was drawing a line between a public symbol, which is conventional and local, and the internal thing it points at, which is shared across all people who have minds. That is the earliest clear statement of an amodal, universal substrate under the variety of tongues, and it is why he keeps appearing in this series like a founding witness.

The medieval logicians took the hint and built it out. William of Ockham, in the fourteenth century, developed a full theory of mental language, in Latin oratio mentalis, an internal discourse whose terms and propositions are prior to any spoken language and common to every rational mind [2]. Ockham’s move was sharper than Aristotle’s. He argued that the mental terms have natural signification, they mean what they mean because of how the mind is, while spoken and written words signify only by convention, because a community agreed to use them that way. The semantics of public language, on this account, is borrowed from a deeper mental language that does the real referring. If that sounds like a medieval anticipation of a modern hypothesis, it is. Fodor’s mentalese, seven centuries later, is Ockham’s oratio mentalis with a computer metaphor bolted on.

The early modern rationalists turned the intuition into programs. René Descartes drove a hard wedge between the mechanical body and the thinking substance and treated rational thought as a faculty not reducible to sensory input or to the words of any language [3]. Leibniz, as we have seen, dreamed the characteristica universalis and a calculus ratiocinator, an idealized formal language plus a mechanical procedure for reasoning in it, which foreshadowed the computational theory of mind so exactly that it reads less like philosophy and more like a design document written before the hardware existed [4], [5]. The Port-Royal Grammar of 1660, written by Antoine Arnauld and Claude Lancelot, analyzed the grammar of actual languages as the surface expression of universal logical structures of thought, an early claim that beneath the differences among French and Latin and Greek lies one rational architecture the grammars are all tracking.

Even the empiricist wing kept the ordering. John Locke, in the Essay Concerning Human Understanding, made words central to knowledge and devoted a whole book to language, yet he held the line that words are arbitrary marks for internal ideas, and that the ideas come first, drawn from sensation and reflection [6]. Words label ideas. They do not manufacture them. Locke worried a great deal about how sloppy words corrupt clear ideas, which only makes sense if the ideas have a prior existence the words can fail to capture.

This whole line comes to its modern head in one book. In 1975, Jerry Fodor published The Language of Thought and argued that cognition runs on an innate, internal, combinatorial symbol system, mentalese, that operates independently of any spoken language [7], [8]. Fodor’s most provocative argument was about learning. To learn a natural language, he claimed, you have to form and test hypotheses about what its words mean, and forming a hypothesis requires a medium to frame it in. So you cannot learn your first representational system from scratch, because learning already presupposes one. The medium you frame the hypotheses in is mentalese, and mentalese therefore cannot itself be learned. It has to be there from the start. You can accept or reject that argument, and many have rejected it, but notice what it does: it makes an amodal language of thought a precondition for acquiring the spoken kind, rather than a mere convenience.

Fodor, working with Zenon Pylyshyn, then gave the position its most durable empirical criterion in a 1988 paper aimed at connectionism [9]. Their claim rested on two properties of thought. The first is systematicity, the fact that a mind able to think “John loves Mary” is thereby able to think “Mary loves John,” because the same constituents can be recombined by the same rules. You do not meet a person who grasps the first and finds the second literally unthinkable. The second is productivity, the capacity to generate an unbounded number of novel thoughts from a finite stock of primitives and rules, most of which the mind has never encountered before. As Fodor and Pylyshyn put it, “the ability to produce and understand some sentences is intrinsically connected to the ability to produce and understand certain others.” Their argument was that these two properties fall out for free from a classical architecture with genuine compositional constituent structure, and that a network lacking such structure either fails to be systematic or succeeds only by quietly implementing a classical symbol system underneath. Structure is the thing, they said, and structure does not care whether its pieces arrived as sounds, sights, or signs. That is the thought-first line, twenty-four centuries after Aristotle, wearing a lab coat.

The counter-line: language as the maker of thought

The modern version begins in the German Enlightenment. Johann Gottfried von Herder argued in his 1772 treatise on the origin of language that speech is the organ of thought itself, that reflection and language arise together, and that a wordless human reason is a fiction [10]. Wilhelm von Humboldt radicalized this into the idea that has organized the debate ever since. Each language, Humboldt held, embodies a distinct Weltansicht, a worldview, a particular way of carving up and construing reality, so that to acquire a language is to inherit a way of seeing [11]. His famous formulation casts language as a continuous generative activity, an energeia, rather than a finished product, an ergon, and holds that thought and language are so intertwined that you cannot cleanly pry them apart. That is the taproot of everything that follows on this side.

Gottlob Frege complicates the map, and I want to flag the complication honestly rather than flatten it, because my two research sources actually disagree about which side he belongs on. Frege held that thoughts, in his technical sense the Gedanken, are objective and language-independent, existing in a third domain beyond the physical world and the private mind, the same thought graspable by anyone regardless of the language they think in [12]. That reads as thought-first. But Frege also insisted that natural language systematically misleads us about the structure of thoughts, and that to grasp a thought clearly you must regiment it in a purpose-built logical notation, his Begriffsschrift. So he made a formal language the necessary instrument for seeing thought as it really is. He is claimed by both camps because he genuinely faces both ways: the content of thought is language-independent for Frege, while our access to that content runs through the discipline of a logical language. Casual histories pick one face and drop the other. The honest version keeps both.

Ludwig Wittgenstein then traveled the entire distance between the two lineages inside a single career, which is why invoking “Wittgenstein” as a name for one position is always a mistake. The early Wittgenstein of the Tractatus Logico-Philosophicus wrote that the limits of my language mean the limits of my world, and built a picture on which language and reality share a logical structure, a view close in spirit to a strong isomorphism between the sayable and the thinkable [13]. The later Wittgenstein of the Philosophical Investigations turned against his own early self and produced the private-language argument, the claim that there could be no meaningful language, and by extension no meaningful thought-code, intelligible in principle to only one person, because meaning depends on public, correctable, rule-governed use [14]. Both Wittgensteins stay relevant to this series for different reasons. The first speaks to representational format. The second speaks to how any symbol, internal or external, could acquire a norm, which is a live problem for the idea of a purely private mentalese and connects straight to the symbol grounding worry from Part 1.

Then the behaviorists tried to settle the whole thing by making thought physical and observable. John Watson, in launching behaviorism in 1913, proposed that thinking simply is subvocal speech, tiny incipient movements of the larynx and vocal apparatus, so that what feels like silent reasoning is covert talking with the volume turned to zero [15]. B. F. Skinner pushed the program to its most ambitious form in Verbal Behavior in 1957, an attempt to explain all of language, and with it the appearance of thought, as operant behavior shaped by reinforcement, with no internal representations invoked at any point. On this view there is nothing under the words to explain. There are only histories of conditioning. The language-forms-thought line, taken to its behaviorist limit, makes a claim far stronger than that language shapes thought. It says the talk is all there is.

The line reached its cultural peak with linguistic relativity. Edward Sapir, in 1929, argued that the “real world” is to a large extent built up on the language habits of a community, and his student Benjamin Lee Whorf pressed this into the strong claim now called the Sapir-Whorf hypothesis, that the grammatical and lexical categories of your native language shape, and in the hard version determine, the categories of your thought. Whorf’s writings, collected in Language, Thought, and Reality in 1956, made this the emblematic twentieth-century form of the counter-lineage. If your language lacks a grammatical resource, the strong version says, the corresponding distinction is unavailable or effortful for your mind. This is spine question one answered as hard as it can be answered on the language side: the substrate is language, all the way down, and different languages are different substrates.

The hinge that broke behaviorism

In 1959, Noam Chomsky reviewed Skinner’s Verbal Behavior in the journal Language, and the review did more damage than most original papers ever do. Chomsky argued that the productivity and structure of language could not be captured by stimulus, response, and reinforcement, because speakers routinely produce and understand sentences they have never heard, an unbounded set generated from finite means, which conditioning histories cannot explain. Terms like “stimulus” and “response,” he showed, either kept their laboratory meanings, in which case they plainly failed to cover verbal behavior, or were stretched so far to fit that they became empty. The review did more than wound Skinner’s book. It helped end strong behaviorism as a research program and reopened the mind, internal representations and all, as a legitimate object of scientific study. The cognitive revolution has many parents, but this review is one of its birth certificates, and notice the irony: the argument that killed the behaviorist wing of the language-forms line, productivity and structure, is the very same argument Fodor and Pylyshyn would turn on connectionism thirty years later. The weapon does not care who wields it.

Here is the whole terrain in one view, the two lineages and where their modern heirs stand.

+---------------------------+---------------------------+
| THOUGHT-FIRST LINE        | LANGUAGE-FORMS-THOUGHT    |
| (thought precedes words)  | (language shapes thought) |
+---------------------------+---------------------------+
| Aristotle                 | Herder                    |
|   affections of the soul  |   speech as organ of      |
|                           |   thought                 |
+---------------------------+---------------------------+
| Ockham                    | Humboldt                  |
|   oratio mentalis         |   Weltansicht, energeia   |
+---------------------------+---------------------------+
| Descartes                 | Wittgenstein (early)      |
|   reason before speech    |   limits of language      |
+---------------------------+---------------------------+
| Leibniz                   | Watson                    |
|   characteristica         |   thought as subvocal     |
|   universalis             |   speech                  |
+---------------------------+---------------------------+
| Port-Royal, Locke         | Skinner                   |
|   grammar / ideas as      |   verbal behavior as      |
|   source                  |   operant conditioning    |
+---------------------------+---------------------------+
| Fodor, Pylyshyn           | Sapir, Whorf              |
|   mentalese,              |   linguistic relativity   |
|   systematicity           |                           |
+---------------------------+---------------------------+
| Frege sits across both: thoughts are                  |
| language-independent, but grasped through             |
| a logical language.                                   |
+-------------------------------------------------------+
| Bridge: Vygotsky. Inner speech as converged           |
| verbal thought, taken up in the next section.         |
+-------------------------------------------------------+

The developmental hinge: Vygotsky against Piaget

The sharpest version of the whole argument is fought over children, because a child is the one place you can watch thought and language arrive and see which one shows up first. Two figures staked out the poles, and the fault line between them still organizes developmental psychology.

Jean Piaget gave weight to the priority of cognition. In his account, infants build a rich sensorimotor intelligence long before they have language, constructing an understanding of objects, space, and causation through action on the world, and language when it comes is grafted onto cognitive structures that action already built. For the early Piaget, the young child’s egocentric speech, the running out-loud narration you hear from a three-year-old playing alone, is a symptom of immature, self-centered thinking that fades as the child becomes able to take other perspectives. Speech, on this telling, expresses a stage of thought. It does not drive it.

Lev Vygotsky, in Thought and Language in 1934, argued the opposite about that same egocentric speech, and the disagreement is precise enough to be productive. Vygotsky held that thought and speech have different developmental roots, a prelinguistic intelligence and a preintellectual vocalization, which converge around the age of two into something new, verbal thought. And he read the egocentric speech that Piaget saw withering as something that instead matures and turns inward. It “goes through an evolution… In the end, it becomes inner speech.” The out-loud narration goes underground and becomes the silent inner voice that adults use to plan, to self-cue, to reason through a hard problem. On this account language becomes the medium of a distinctively human kind of self-directed cognition, a tool the mind builds and then internalizes until it feels like the seat of thought itself.

I think this is the most sophisticated argument the language-central camp has, and it deserves respect precisely because it is developmental and mechanistic rather than a slogan. It does not claim language is thought from the start. It claims that a specific, powerful layer of adult cognition is language, taken inward. And it connects directly to the findings from Part 1 on people who lack an inner voice, the condition called anendophasia, who reason competently without the internalized channel Vygotsky described, which is exactly the kind of case that tells you how load-bearing that channel really is.

Set the two side by side and the modern shape of the field appears. Piaget’s heirs run toward core knowledge and the claim that structured cognition precedes and underlies language. Vygotsky’s heirs run toward the claim that language augments and partly reconstructs cognition through development. Both survived. What did not survive was the pretense that you have to choose all or nothing.

Why a medieval monk is in your model evaluation

Take the Fodor and Pylyshyn criterion, systematicity and productivity, and read it as a specification for testing a model. Their 1988 challenge was aimed at the neural networks of their day, but it is the same challenge a serious evaluation puts to a large language model now, whether or not the evaluators know they are quoting a philosophy paper. When you test whether a model that handles “load the red block onto the blue block” also handles a novel recombination it never saw in training, “load the blue block onto the red block,” you are running the systematicity test. When you probe whether performance generalizes to compositions of known parts or collapses the moment the combination is unfamiliar, you are asking Fodor and Pylyshyn’s question in code. The whole modern research literature on compositional and systematic generalization in deep learning is the 1988 argument reborn with benchmarks, and the answer is genuinely unsettled, which is why the evaluation genuinely bites. A model that only appears systematic, that has memorized enough combinations to fake the property without the underlying structure, is a specific and diagnosable failure mode, and knowing that the property has a name and a two-hundred-year pedigree makes you a better tester of it.

Leibniz cashes out too. His dream of reasoning as computation over a formal notation is the ancestor of symbolic AI, and the reason the neuro-symbolic middle keeps returning, the reason people keep trying to bolt explicit structure onto learned networks, is that the thought-first intuition about compositional structure refuses to die even as the connectionist systems get more capable. If you are allocating capital or writing policy, the useful thing to carry out of this history is that the choice between a system that learns everything from a rich channel and a system that is given structured, compositional representations to compute over is not a new engineering fashion. It is the oldest question in the field, and treating it as settled in either direction is how you get surprised.

There is a governance edge to the history as well. When a language model emits a step-by-step chain of reasoning, the temptation is to read that visible chain as the thought, the way Watson read subvocal movements as the thought and the way you read your own inner monologue as your thinking. Part 1 warned against exactly this conflation of the display with the substrate. The history sharpens the warning: behaviorism died in part because it mistook the observable trace of a process for the process. If you certify or audit these systems on the basis of their rendered reasoning traces, you are making a bet that the trace is the computation, and that is precisely the bet a century of cognitive science learned to distrust.

What survived the century

Two extreme positions died in the twentieth century, and they died for good reasons. Strong behaviorism died, killed by Chomsky’s review and by the sheer structured productivity of language and thought that stimulus and response could never generate. Strong Whorfianism, the hard claim that language determines and bounds what you can think, also died, and it died on evidence I am going to lay out in Part 4, where infants and animals turn out to hold concepts they have no words for and speakers of radically different languages turn out to share far more cognition than the strong thesis predicts. Neither of those deaths is controversial among people who work on this now, whatever the popular science headlines suggest.

What survived on the thought-first side is the load-bearing claim of the whole lineage: cognition is computation over structured, compositional representations, and there is a distinction worth keeping between public language and the internal system that does the reasoning. That claim runs unbroken from Aristotle’s affections of the soul through Ockham’s mental language to Fodor’s mentalese, and it is the deep root of the amodal computational core this series is threading. I want to be honest that “survived” means “remains the strongest available framework,” not “was proven.” The thought-first line earned its standing by explaining systematicity, productivity, and the dissociation of reasoning from language that later neuroscience would document [16], [17]. It did not earn the right to specify that the core must be classical symbols rather than some structured distributed format, and that specific question stays genuinely open.

What survived on the language side is more limited and more interesting than the slogans. Language turned out to be an extraordinary tool for thought rather than its engine: a compressor of the world’s regularities, a scaffold for concepts a mind could not stabilize alone, the medium that makes knowledge portable, inspectable, and shareable across people and generations, and, in Vygotsky’s sense, an internalized channel that reshapes a thinking system it did not create. The moderated relativism that survived says language biases attention, categorization, and memory in selected domains, especially where a culture has stabilized a coding scheme for distinctions that would otherwise stay slippery. That is a real effect, and Part 4 takes it into the laboratory. It is a long way from the claim that language is the substrate.

So the durable inheritance from the century, the thing I think is worth carrying forward, is the separation between language as a tool for thought and language as the engine of thought, with the evidence favoring the first reading. Intelligence looks older and deeper than any single public code. The core knowledge that Elizabeth Spelke and others document in prelinguistic infants, structured expectations about objects, number, space, and agents, is the modern descendant of the thought-first line’s oldest bet, that a mind arrives with structure before any language teaches it much of anything [18].

I am taking you, the reader, through the history that without asserting it as settled, and key questions didn’t really got answered in this part. The ancients and the moderns were arguing about spine question one, whether the substrate is language or something prior to it, and the argument is what I have traced, but the adjudication lives in the evidence, and the evidence is what the next parts are for.

Part 4 takes the surviving relativism and the core-knowledge claim into developmental and cross-cultural data, where children and cultures test what actually exists before and without language.

That is where the two-thousand-year argument finally meets a control group.

Frequently Asked Questions.

Is the “language of thought” the same as the language I think in? No, and the whole tradition turns on keeping them apart. The language of thought, mentalese in Fodor’s sense, is a hypothesized internal, amodal, combinatorial system of representations that is not English or any other spoken language and that operates below awareness [7]. The language you seem to think in, the inner voice, is a conscious phenomenon that Part 1 placed at the display level, and Vygotsky argued it is spoken language taken inward during development. One is a claim about the computational substrate. The other is a claim about conscious experience. Confusing them is the error the series exists to correct.

If both camps are two thousand years old, hasn’t the argument just gone in circles? The positions are old. The evidence is not, and that is what breaks the circle. Aristotle and Humboldt argued from introspection and grammar because that is all they had. We now have infant looking-time studies, brain imaging that dissociates reasoning from the language network, deprivation cases, and models we can probe. The philosophical options were mostly enumerated centuries ago. The job of the last few decades has been to find out which ones survive contact with data, and some clearly have not.

Didn’t Chomsky’s review of Skinner just win the argument for the thought-first side? It ended one specific program, strong behaviorism, by showing that conditioning cannot explain the productivity of language [15]. That cleared the ground for cognitivism and for Fodor. But it did not settle the deeper question of whether the internal system is classical symbols, a structured distributed representation, or something else, and it did not touch the moderated version of linguistic relativity, which is alive and defensible. Winning against the strongest form of a rival is not the same as proving your own architecture, a distinction I try to keep throughout.

Where does Wittgenstein actually stand? In two places, because he changed his mind and the change matters. The early Wittgenstein of the Tractatus is close to a strong link between language and world [13]. The later Wittgenstein of the private-language argument holds that meaning depends on public, rule-governed use, which is a direct problem for the idea of a purely private mental code [14]. If someone cites “Wittgenstein” as if he had one view on this, they are almost certainly collapsing two positions that contradict each other.

Is Whorf simply wrong, then? The strong version, that language determines thought and bounds what you can conceive, is wrong, and it fails against the evidence in Part 4. A weak version survives and is well supported: language influences attention, memory, and categorization in specific domains, especially where a linguistic distinction is habitual and a nonlinguistic one would be effortful. The honest summary is that Whorf overreached by an enormous margin and still identified a real, smaller effect. Both halves of that sentence are true, and dropping either one distorts the record.

Why should an AI engineer care about Ockham or Leibniz? Because they named the bet you are placing. Ockham’s mental language and Leibniz’s characteristica are the intuition that thought is structured, compositional, and language-independent, which is exactly the property modern evaluations test when they probe whether a model generalizes to novel combinations of known parts [9]. When you find yourself debating whether to add explicit structure to a learned system, or whether scale alone will induce that structure, you are choosing between the thought-first line and its rivals with a budget. Knowing the lineage will not tell you which is right. It will stop you from thinking the question is new.

What is the difference between the thought-first line and a “world model”? They answer different questions and can both be right. The thought-first line is about representational format, whether cognition runs on structured, compositional, amodal representations. A world model is about content, whether a system has learned the dynamics and structure of an external world well enough to predict and plan. You can hold that the format is amodal and compositional, the thought-first bet, while also holding that the content has to be grounded in real contact with a world, which is the grounded camp’s surviving claim from Part 1. The series thesis actually takes both, and Part 10 is where the world-model debate gets its own treatment.

Does this history take a side on whether large language models understand? Not directly, and I am wary of anyone who says the history settles it. What the history offers is the right question. Frege and the private-language argument press whether internal symbols can mean anything without public grounding. Fodor and Pylyshyn press whether a system without genuine compositional structure can be reliably systematic. Both questions apply to modern models, both are being actively researched, and both are still open. The genealogy sharpens the interrogation. The verdict comes from evidence I take up in the parts on neural encoding, on grounding, and on the industry’s wager.

Glossary

Affections of the soul. Aristotle’s term in On Interpretation for the mental contents that spoken and written words signify. He held that words differ across languages while these mental contents are the same for all people, an early statement of a universal substrate beneath linguistic variation.

Behaviorism. The early-twentieth-century program, associated with Watson and Skinner, that restricted psychology to observable behavior and tried to explain thought and language through conditioning without appeal to internal representations. Its strong form was ended by Chomsky’s 1959 critique.

Characteristica universalis. Leibniz’s proposed universal formal notation in which every concept would have a precise symbolic form, paired with a mechanical procedure for reasoning, so that disputes could be settled by calculation. It is a direct conceptual ancestor of symbolic AI and the computational theory of mind.

Egocentric speech. The out-loud self-directed narration typical of young children. Piaget read it as a symptom of immature thought that fades with development; Vygotsky argued it goes underground and becomes inner speech rather than fading.

Inner speech. The silent verbal narration adults use to plan, self-cue, and reason. Vygotsky argued it is internalized egocentric speech, making a specific layer of adult cognition into language taken inward.

Language of thought (mentalese). Fodor’s hypothesis that cognition runs on an innate, internal, combinatorial symbol system that operates independently of any spoken language and supplies the compositional structure reasoning requires. It is the modern descendant of Ockham’s mental language.

Language-forms-thought lineage. The tradition, running from Herder and Humboldt through behaviorism to Sapir and Whorf, holding that language constitutes or strongly shapes thought rather than merely expressing it.

Oratio mentalis (mental language). Ockham’s fourteenth-century theory of an internal discourse, prior to any spoken language and common to all rational minds, whose terms signify naturally while spoken words signify only by convention. The semantics of public language, on this view, is borrowed from the mental one.

Private-language argument. The later Wittgenstein’s argument that there can be no meaningful language, and no meaningful thought-code, intelligible in principle to only one person, because meaning depends on public, correctable, rule-governed use. It pressures the idea of a purely private internal symbol system.

Productivity. The capacity to generate an unbounded number of novel thoughts or sentences from a finite set of primitives and rules. Fodor and Pylyshyn treated it, with systematicity, as a hallmark of a compositional architecture.

Sapir-Whorf hypothesis (linguistic relativity). The claim that the categories of one’s language shape the categories of one’s thought. The strong, determinist form did not survive empirical testing; a weaker form, that language biases cognition in selected domains, is supported.

Systematicity. The property that the ability to understand one structured thought brings with it the ability to understand structurally related thoughts, because the same constituents recombine by the same rules. Fodor and Pylyshyn cited it as evidence for a compositional substrate and as a challenge to networks that lack one.

Thought-first lineage. The tradition, running from Aristotle and Ockham through Descartes, Leibniz, Locke, and the Port-Royal grammarians to Fodor, holding that thought is structured and largely language-independent, and that spoken language is a local encoding of a prior internal system.

Verbal thought. In Vygotsky’s account, the new formation that appears around age two when a prelinguistic intelligence and a preintellectual vocalization converge, producing a distinctively human, language-mediated mode of self-directed cognition.

Weltansicht. Humboldt’s term for the worldview embodied in a particular language, the specific way it carves up and construes reality, such that acquiring a language means inheriting a way of seeing. The taproot of linguistic relativity.

References

[1] Aristotle, On Interpretation (De Interpretatione), trans. E. M. Edghill, The Internet Classics Archive. [Online]. Available: http://classics.mit.edu/Aristotle/interpretation.html

[2] P. V. Spade and C. Panaccio, “William of Ockham,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/ockham/

[3] G. Hatfield, “René Descartes,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/descartes/

[4] B. Look, “Gottfried Wilhelm Leibniz,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/leibniz/

[5] M. Rescorla, “The computational theory of mind,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/computational-mind/

[6] W. Uzgalis, “John Locke,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/locke/

[7] M. Rescorla, “The language of thought hypothesis,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/language-thought/

[8] J. A. Fodor, The Language of Thought. Cambridge, MA: Harvard Univ. Press, 1975. [Online]. Available: https://books.google.com/books/about/The_Language_of_Thought.html?id=XZwGLBYLbg4C

[9] J. A. Fodor and Z. W. Pylyshyn, “Connectionism and cognitive architecture: A critical analysis,” Cognition, vol. 28, no. 1–2, pp. 3–71, 1988. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0010027788900315

[10] M. Forster, “Johann Gottfried von Herder,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/herder/

[11] J. Stam and members of the SEP, “Wilhelm von Humboldt,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/wilhelm-humboldt/

[12] E. Zalta, “Gottlob Frege,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/frege/

[13] L. Wittgenstein, Tractatus Logico-Philosophicus, trans. C. K. Ogden, Project Gutenberg. [Online]. Available: https://www.gutenberg.org/ebooks/5740

[14] S. Candlish and G. Wrisley, “Private language,” The Stanford Encyclopedia of Philosophy. [Online]. Available: https://plato.stanford.edu/entries/private-language/

[15] J. B. Watson, “Psychology as the behaviorist views it,” Psychological Review, vol. 20, no. 2, pp. 158–177, 1913, Classics in the History of Psychology. [Online]. Available: https://psychclassics.yorku.ca/Watson/views.htm

[16] E. Fedorenko, S. T. Piantadosi, and E. A. F. Gibson, “Language is primarily a tool for communication rather than thought,” Nature, vol. 630, pp. 575–586, 2024. [Online]. Available: https://doi.org/10.1038/s41586-024-07522-w

[17] H. Kean, A. Fung, P. Jaggers, J. Chen, J. S. Rule, Y. Benn, J. B. Tenenbaum, S. T. Piantadosi, R. A. Varley, and E. Fedorenko, “Evidence from formal logical reasoning reveals that the language of thought is not natural language,” Proceedings of the National Academy of Sciences, vol. 123, no. 28, e2520095123, 2026. [Online]. Available: https://www.pnas.org/doi/10.1073/pnas.2520095123

[18] E. S. Spelke and K. D. Kinzler, “Core knowledge,” Developmental Science, vol. 10, no. 1, pp. 89–96, 2007. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/17181705/