cognitive science and artificial intelligence: one intertwined history
Cognitive science asks how minds work. Artificial intelligence asks how to build one. Told separately, each history is a tidy story of progress. Told together, they turn out to be one story: two disciplines born in the same year, from the same idea, that have spent seventy years trading places as teacher and student.
The idea was this: that thinking might be a kind of computation — so that a machine which computed the right way would think, and a mind that thought was computing in some way we could discover. From 1956 on, a program could be a psychological theory and a psychological theory could be a program. The perceptron, ancestor of every neural network, was published in a psychology journal. So was the first general problem-solving program, and the first learning theory of what became reinforcement learning. When the mind sciences changed their picture of thought — from rules to networks, from networks to bodies, from bodies to prediction — AI changed with them, or against them. Then, quite recently, the current reversed: the machines became the best available models of the brain, and the mind sciences turned their instruments on the machines.
The left strand is the study of minds: philosophy, psychology, linguistics, neuroscience. The right strand is the making of machines: logic, computing, AI. Where an event sits on the spine, it belonged to both at once. The colored threads follow single ideas as they cross from one strand to the other and back. Pick a thread to isolate it; tap any event to read more. The layout is not to scale: each event gets one row, so the crowded recent decades stretch and the sparse early centuries compress.
Seven patterns appear when the two histories are laid side by side.
1879: psychology gets a laboratory and logic gets a notation. 1943: a neurophysiologist and a logician describe the neuron as a logic gate. 1956: the same program is shown to computer scientists in June and to psychologists in September. 1986: the central debate of cognitive science is conducted between two AI architectures. 2014: an AI system becomes the best available model of visual cortex. The knots on the spine are not coincidences. They are the actual structure of the field.
For most of this history, the psychology journal of record was where AI published its architectures. Rosenblatt’s perceptron (1958), Newell, Shaw and Simon’s problem solver (1958), Sutton and Barto’s prediction learning (1981), and Landauer and Dumais’s latent semantic analysis (1997) all appeared in Psychological Review; Elman’s recurrent network (1990) in Cognitive Science; the PDP volumes were subtitled “explorations in the microstructure of cognition.” The artifacts were the theories.
Psychology borrowed storage, retrieval, encoding, buffer, capacity, program from computing. Computing borrowed neuron, learning, attention, memory, replay, curriculum, forgetting from psychology — and lately biology, dissection, circuit. Each field uses the other for legitimacy: the computer made inner states scientific again after behaviorism; the brain made statistical pattern-matchers sound like minds.
A learning rule proposed for cats in puzzle boxes (1911) became an equation for conditioning (1972), then an algorithm (1981–88), then was found in dopamine neurons (1997), then played Atari (2015). Neurons discovered in cat cortex (1959) became an architecture (1980), then a working system (1989, 2012), then the best model of those very neurons (2014). Helmholtz’s unconscious inference (1867) became Bayesian children (2004), predictive brains (2013), and next-word prediction (2020).
Behaviorism banished the inner life; cognitivism restored it as symbols; connectionism dissolved the symbols into weights; the embodied turn threw out the inner model altogether; the Bayesian turn brought the model back as prediction; deep learning brought back the weights at scale. Language models are, in one light, behaviorism’s revenge — systems trained on nothing but recorded verbal behavior — and in another, the most “cognitive” artifacts ever made, full of inner representations that interpretability now catalogs. Skinner and Chomsky each turned out to be partly right, in ways neither would have accepted.
Dreyfus said common sense could not be written down as rules; expert systems proved him right — but the remedy was not embodiment, it was learning from data at a scale nobody imagined. Searle said syntax is not semantics; the question was never settled, only renamed. Lovelace said the engine could originate nothing; Turing answered with machines that learn, and machines that learn now originate things their makers cannot account for. Every line drawn around what machines cannot do has been crossed, and each crossing taught us that the line was in the wrong place — not that there is no line.
For sixty years AI was a model of the mind: build it and you will understand it. Since about 2014 the flow has reversed. Deep networks are the leading models of visual and language cortex; psychologists run their experiments on language models; interpretability researchers dissect networks with tools borrowed from neuroscience and describe what they find in the language of biology. The student became the specimen.
What follows is one reading of this history: what it shows, what it changes, and what it does not.
The chips at the top of this page are a warning. Hydraulic spirits, clockwork, telegraph wires, storage and retrieval, weights and activations, prediction error: each was the best available technology, and each was mistaken for the mind itself. The computer, though, is a different kind of metaphor. Turing’s thesis says that anything done by following a definite procedure can be done by a universal machine. So “the mind is a computer” was never merely an analogy; it was the claim that whatever the mind does effectively, a machine can do too — and the seventy years since have been a long, uneven confirmation of it, one supposedly uncomputable capacity at a time. What remains open is not whether the mind’s doings can be computed but whether being the computation is the whole story: whether there is something it is like to run it. The history settles the first question far more than the second.
Vico said we truly know only what we make. Cognitive science was founded on that faith — the synthetic method: understand the mind by building one. It worked, for a while, in a peculiar way: each artifact revealed that the capacity it reproduced was not what we had thought (chess was search; fluency was prediction). But the newest artifacts broke the bargain. We built systems whose behavior we cannot derive from their design, and so we now study them as we study brains — recording units, lesioning components, tracing circuits, writing papers titled “on the biology of” a program. Braitenberg warned in 1984 that it is far easier to invent a mind-like thing than to analyze one. He was describing toy vehicles. He turned out to be describing the century.
Language models are trained on recorded behavior — text — and evaluated by behavior: benchmarks are Turing tests. Their insides are opaque enough that we had to invent a psychology for them before we could invent a neuroscience. AI has re-run the twentieth century’s argument about the mind in fast-forward: behaviorism (all we have is input and output) gave way to cognitivism (there are inner representations; interpretability finds them) in the space of a few years. This suggests the old argument was never really about ideology. It was about epistemic access. And it suggests that “does it really understand?” has the shape of “does the other person really feel pain?” — a question we answer about each other not by inspecting the inside but through a whole form of life in which we treat one another as understanding. Wittgenstein’s beetle in the box has become an engineering problem.
Descartes drew the line at language and general reason. Lovelace drew it at origination. It was redrawn at arithmetic, at chess, at Go, at conversation, each time with the assurance that this capacity was the essential one — and each time the crossing was followed by the quiet verdict that it had not been intelligence after all. The pattern is called the AI effect, and it is a symptom: we were using cognitive capacities as the ground of human worth, and cognitive capacities keep turning out to be reproducible. The most coherent place the line has come to rest is not a capacity but a condition. Living systems maintain themselves; they have stakes; things can go well or badly for them. Jonas, Maturana and Varela, Thompson, and Haugeland (“the trouble with artificial intelligence is that computers don’t give a damn”) all locate the origin of mind there — in caring, not in computing. Whether an artifact could come to have stakes is an open question. The point is that the criterion changed: from how well something thinks to whether anything matters to it.
Polanyi’s “we know more than we can tell” and Dreyfus’s account of expertise as embodied intuition were meant to mark what machines could not have; economists cited Polanyi’s paradox to predict which jobs were safe. They were right that expert skill cannot be written down as rules — the expert-systems industry died of exactly that. They were wrong about the consequence. Tacit skill turned out to be the easiest thing to reproduce, because it can be learned from examples without ever being made explicit — which is, after all, how people acquire it. What remains distinctively human is therefore not a skill. It is a stance: judging what is worth doing, answering for what was done, and being the one for whom it was done.
Predictive processing says the brain is a generative model that constantly forecasts its inputs and corrects itself by the errors; the “self,” on this view, is part of the model — a useful fiction the system maintains about its own body and history. Metzinger called it a self-model; Hume called the self a bundle of perceptions; the Buddhist philosophy Varela drew on called it anattā. The convergence is striking. So is the warning that goes with it: this is the metaphor problem again. We are reading ourselves through a prediction machine because prediction machines are the newest thing we have built. The agreement with older traditions is a reason to take the idea seriously, and the history is a reason to hold it lightly.
Clark and Chalmers argued that Otto’s notebook was part of Otto’s memory; Hutchins showed a ship’s crew and instruments doing a computation no single head held; Licklider’s man–computer symbiosis was an engineering goal in 1960. The mind has always leaked into its tools — writing, arithmetic, maps. What is new is not that we offload thinking but what we offload (the composing and reasoning we took to be central) and to whom. Otto’s notebook did not have a business model, did not model Otto in return, and did not adjust itself to keep him turning pages. Skinner’s schedules of reinforcement were rediscovered by engagement optimization; Broadbent’s attentional filter is now contested territory. The same knowledge that lets us understand ourselves lets others steer us.
For four centuries, fluent language was the strongest evidence of a mind; it was Descartes’s first test. Then two things happened at once. Machines became fluent without the other marks of mind we expected to come with it, and neuroscience produced evidence that in human brains the language network is distinct from the networks that reason, plan, and understand other people. If language and thought are separable, then fluency is neither proof of thought nor proof against it — in machines or in anyone. It also means we may finally learn what is specifically human about human cognition by contrast: the slow, embodied, social, curious, mortal way we come to know things from a thousandth of the data.
It does not follow that the mind is a computer in any deflationary sense; the history shows that computational modeling is astonishingly productive, not that it is exhaustive. It does not follow that language models understand, or that they do not; it follows that our tests were inadequate and that we are, for now, in the position of the behaviorists. It does not follow that consciousness has been explained; nothing here touches the hard problem, and the honest thing to say is that we do not know whether anything we have built has experiences, and we do not have a method for finding out. It does not follow that cognitive science was refuted by the bitter lesson; its empiricist wing is winning, its nativist wing has an unanswered point about data efficiency, and its embodied wing is waiting to see whether bodies turn out to matter. And it does not follow that the question should be dropped. The defensible conclusion is an uncomfortable pair: epistemic humility about what we are and what we have made, and practical urgency about how to live with it.