Introduction to Cognitive Science

The big picture

Take a second to notice what you are doing right now. Photons are striking your retina, and out of that flickering mosaic you are effortlessly extracting letters, words, and meaning. You are holding the beginning of this sentence in mind while its end arrives. Somewhere in the background you may be monitoring a conversation, planning dinner, and feeling mildly curious or mildly skeptical. Every bit of this is being done by roughly three pounds of tissue with the consistency of soft tofu, running on about 20 watts, the power of a dim light bulb. Nobody fully understands how. The scientific attempt to find out is called cognitive science.

Here is the plan for today. First we will define the field and its central bet: that thinking is a form of information processing that can be studied scientifically. Then we will meet the six disciplines that founded it, and see what each one contributes that the others cannot. We will work one everyday act, understanding a sentence, through all six lenses. And we will finish by being honest about what the field is not: it is not neuroscience with better marketing, and it is not a self-improvement program. If you leave with one idea, let it be this: no single method can crack the mind, which is why cognitive science was built, on purpose, as a coalition.

One question nobody owns

Cognitive science is the interdisciplinary study of mind and intelligence. The question it asks is old: how does thinking work? What is new, historically speaking, is the way it asks. Instead of treating the mind as a philosophical mystery or a poetic metaphor, cognitive science treats mental activity as information processing: the mind takes in information from the senses, transforms it, stores it, retrieves it, and uses it to guide behavior. Perceiving, remembering, speaking, deciding, and imagining are all, on this view, computations of a kind, and computations can be studied, modeled, and tested.

Notice what this commitment does and does not say. It does not say your mind is literally a laptop, and it does not say feelings are fake. It says that when you recognize your friend's face, something in you is taking retinal input and turning it into an identity, and that this transformation has a describable logic that science can uncover. That is a hypothesis, not a certainty, and parts of this course will test its limits. But it has been an extraordinarily productive hypothesis, and every discipline in the field shares it.

Key idea: Cognitive science is the interdisciplinary science of mind, built on one working bet: mental activity is information processing, so it can be described, modeled, and tested rather than only introspected about.

The hexagon

In 1978 the Sloan Foundation commissioned a report on this young field, and the report contained a diagram that became its unofficial logo: a hexagon with six disciplines at the corners, connected by lines marking their collaborations. The corners were psychology, neuroscience, linguistics, philosophy, artificial intelligence, and anthropology. The diagram made a claim: the mind sits in the middle, and every corner sees a face of it the others miss.

Discipline	What it asks about the mind	Signature evidence
Psychology	What can minds do, and under what conditions do they fail?	Controlled experiments on behavior: reaction times, error patterns, illusions
Neuroscience	How does brain tissue implement mental activity?	Lesion studies, single-cell recordings, EEG, fMRI
Linguistics	What is the structure of language, the mind's most revealing output?	Grammaticality patterns, child language, cross-language comparison
Philosophy	What are the concepts (representation, consciousness, knowledge) the science relies on?	Arguments, thought experiments, conceptual analysis
Artificial intelligence	Can we build the process? What does it take to make a machine do this?	Working computational models, from logic programs to neural networks
Anthropology	Which parts of cognition are universal, and which are shaped by culture?	Fieldwork and cross-cultural experiments

Each corner also has a famous blind spot. Psychology measures behavior with beautiful precision but cannot, by itself, see the machinery. Neuroscience sees machinery but risks producing what critics call a map without a legend: activity everywhere, meaning unclear. Linguistics has deep theory but about one domain. Philosophy sharpens questions but does not run studies. AI builds working systems that may work in thoroughly un-humanlike ways. Anthropology finds variation but rarely isolates causes. The hexagon exists because the blind spots do not overlap.

Key idea: The six founding disciplines are psychology, neuroscience, linguistics, philosophy, AI, and anthropology. Each contributes a kind of evidence the others cannot generate, and each has a blind spot another corner covers.

A worked example: understanding one sentence

Abstract diagrams become real when you push an example through them. Take the sentence: "The coffee is too hot to drink." You understood it instantly. Now watch six disciplines take it apart.

The psychologist asks what you did and how long it took. Eye-tracking shows your gaze landing on content words, skipping short function words, regressing when a sentence turns ambiguous. Priming experiments show that reading "coffee" makes you faster, for a few hundred milliseconds, at recognizing "tea": meaning spreads. The linguist asks how the sentence works at all. Notice that "too hot to drink" leaves out who is drinking and what is drunk, yet you filled both in without noticing. Compare "The coffee is too hot to drink" with "The chef is too tired to cook": the missing pieces get filled differently, following structural rules you were never taught and cannot state. The neuroscientist asks where and when: EEG shows a distinctive brain response about 400 milliseconds after a word that violates meaning ("He spread his toast with socks"), and damage to certain left-hemisphere regions can remove the ability to use grammar while leaving single words intact.

The AI researcher asks what it takes to build a system that fills in those missing pieces, and discovers the hard way that it requires enormous background knowledge: knowing that people drink coffee, that hot liquids burn, that "too X to Y" signals prevention. Decades of hand-coding that knowledge mostly failed, and modern language models that learn it statistically raise their own questions, which we will treat carefully in Module 6. The philosopher asks what "understanding" even means here: if a machine fills in the gaps correctly, does it understand, or merely behave as if it does? And the anthropologist asks what varies: languages differ in how they carve up space, time, and even the word "hot," and those differences let us test which parts of comprehension are universal equipment and which are cultural inheritance.

One two-second act; six research programs. That is the field in miniature.

Why converging evidence is the method

You might worry that interdisciplinarity is a polite word for vagueness. In practice it is a discipline of its own, and its logic is converging evidence: a claim about the mind earns trust when independent methods, with independent weaknesses, point to the same conclusion. Consider the claim that short-term memory and long-term memory are different systems. Psychology contributes the behavioral signature: in free recall, people remember the last few items well only if tested immediately, and the first few items well regardless. Neuroscience contributes patient H.M., who after surgery could hold a conversation (short-term intact) but could not form new lasting memories, and other patients showing the reverse pattern. AI contributes working models showing that a fast, small buffer plus a slow, huge store is an effective design for a learning system. No one strand is decisive; the braid is strong. You will see this pattern in every module, and you should demand it from every confident claim about the mind you meet outside this course.

Key idea: Cognitive science runs on converging evidence: independent methods with different weaknesses supporting one conclusion. A finding backed by behavior, brain data, and a working model outranks any single striking study.

What cognitive science is not

Three clarifications will save you confusion later. First, cognitive science is not simply neuroscience. Knowing where something happens in the brain is not the same as knowing what is happening, a point we will sharpen next lesson into Marr's levels of analysis. A perfect wiring diagram of the brain would still leave you needing a theory of what the wiring computes. Second, cognitive science is not the commercial "brain training" industry. Reviews of the evidence, including a 2016 consensus evaluation in Psychological Science in the Public Interest, find that training on puzzle apps makes you better mainly at those puzzles, with little transfer to general intelligence or everyday memory. Third, cognitive science is not introspection. One of the field's founding lessons, which you will meet repeatedly, is that the mind does not come with a window on its own workings: people confidently report reasons for choices that demonstrably were not the causes. That is precisely why we need reaction times, patients, models, and cross-cultural data rather than testimony alone.

A note on the word "cognitive." In this field it does not mean "cold reasoning as opposed to emotion." Emotion, motivation, and social life are information-processing problems too, and modern cognitive science studies them as such. The border of the field is not a border around a part of your mental life; it is a commitment about method.

Key idea: Cognitive science is not neuroscience alone, not brain-training marketing, and not introspection. It is a method: treat mental activity as information processing and test claims with converging evidence.

Where the field came from, in brief

The hexagon did not assemble by accident. In the mid-1950s, researchers in psychology, linguistics, and the brand-new field of computing began meeting around a shared frustration with behaviorism, the school then dominating American psychology, which held that science should speak only of stimuli and responses, never of inner processes. A famous symposium at MIT on September 11, 1956 put a young Noam Chomsky, the memory researcher George Miller, and the AI pioneers Allen Newell and Herbert Simon on the same stage, and Miller later wrote that he left with the conviction that these separate fields were fragments of one larger science. The label arrived in the 1970s: the journal Cognitive Science was founded in 1977, the Cognitive Science Society in 1979, and the Sloan report drew its hexagon in 1978. Next lesson we tell that founding story properly, because the argument that killed behaviorism, and the computer metaphor that replaced it, are still the field's load-bearing walls.

Common misconceptions

• "Cognitive science is just a fancy name for psychology." Psychology is one corner of the hexagon. A psychology experiment cannot tell you whether a grammar is learnable in principle, whether a model can be built, or whether a concept is coherent. The field exists because those questions need linguistics, AI, and philosophy in the room.
• "Studying the mind scientifically means reducing everything to brain scans." Brain data is one evidence stream among several, and by itself it often cannot distinguish rival theories of what the mind is doing. Behavior, models, and language structure carry equal weight.
• "The computer metaphor means scientists think you are a laptop." The claim is that mental processes are describable as information processing, not that neurons are silicon or that minds run Windows. How far the metaphor stretches is an open research question, not a dogma.
• "Cognitive means unemotional." Emotion and social cognition are central research areas. The word marks a method (information processing plus converging evidence), not a cold subject matter.
• "Brain-training apps are applied cognitive science." The consensus evidence shows narrow practice effects with little general transfer. Real applied cognitive science looks like aviation checklist design, eyewitness interview reform, and speech therapy, all of which you will meet in this course.

Recap

Cognitive science is the interdisciplinary study of mind and intelligence, founded on the working bet that mental activity is information processing. Its six founding disciplines, psychology, neuroscience, linguistics, philosophy, artificial intelligence, and anthropology, each contribute a kind of evidence the others cannot: behavioral signatures, neural implementation, linguistic structure, conceptual rigor, working models, and cultural variation. The field's method is converging evidence, in which independent approaches with independent weaknesses must agree before a claim is trusted. It is not neuroscience alone, not introspection, and not brain-game marketing. It was born in the 1950s when several fields discovered they were asking one question, and the next lesson tells that story: how the mind became a legitimate scientific object again.

Sources

1. Thagard, P. (2023). Cognitive science. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
2. Encyclopaedia Britannica. (2024). Cognitive science. Encyclopaedia Britannica. britannica.com
3. Wikipedia. (2025). Cognitive science. Wikimedia Foundation. en.wikipedia.org
4. MIT OpenCourseWare. (2011). 9.00SC Introduction to Psychology. Massachusetts Institute of Technology. ocw.mit.edu

The big picture

Every science has a founding story it tells its students, and this is ours. For roughly the first half of the twentieth century, American psychology officially did not study the mind. That sounds absurd, but it was a principled position with real arguments behind it, and understanding those arguments is the only way to appreciate what came next. Between about 1950 and 1970, a coalition of psychologists, linguists, and computer scientists overthrew that position in what is now called the cognitive revolution. Out of it came the field you are studying and its most important thinking tool, which we will spend the second half of this lesson learning to use: David Marr's three levels of analysis.

A warning before we start: founding stories get simplified in the retelling, and this one is no exception. The revolution was slower, messier, and less total than the legend says, and behaviorism left behind methods and findings the field still uses daily. We will tell the honest version.

Why serious scientists banned the mind

Recall from psychology's early history that introspection, the trained self-observation of experience, collapsed because different laboratories reported different experiences of the same stimulus and nothing could settle who was right. John Watson's response in 1913 was surgical: if reports of inner experience cannot be verified, science should stop relying on them. Study what can be observed by anyone: stimuli and responses. B. F. Skinner built this into a powerful research program, operant conditioning, showing with beautiful precision how consequences shape behavior. Reward a hungry pigeon for pecking a lit key and the pecking rate climbs; vary the schedule of reward and the behavior follows lawful, replicable curves. No talk of what the pigeon believes or wants was needed, and Skinner argued the same would hold for people.

Do not condescend to this view. It was a serious attempt to make psychology rigorous, it produced technologies that still work (behavioral therapies for phobias, animal training, the token economies used in some clinics), and its statistical descendant, reinforcement learning, sits inside modern AI. The question is not whether behaviorism discovered real things. It did. The question is whether stimulus and response are enough to explain all behavior. Three cracks said no.

Three cracks in the wall

The first crack came from behaviorism's own laboratory animals. In the 1930s and 1940s, Edward Tolman ran rats in mazes and found behavior that reward could not explain. Rats allowed to wander a maze with no food reward seemed to learn nothing, but the moment reward was introduced, their performance jumped almost instantly to the level of rats rewarded all along. The learning had happened without reinforcement and sat latent until it mattered. Rats also took novel shortcuts when a learned route was blocked, as if consulting a map rather than replaying a habit. Tolman argued they had built an internal cognitive map, a representation of the maze's layout. An unobservable inner structure was doing observable work.

The second crack was quantitative. In 1956 George Miller published one of psychology's most famous papers, "The Magical Number Seven, Plus or Minus Two," documenting that immediate memory holds a strikingly constant number of units, about seven digits, letters, or words, and crucially that the limit counts chunks, not raw information. The letter string FBICIANASA is ten letters, hopeless as ten items, trivial as three familiar chunks. But a chunk is defined by what the person knows, which meant the capacity of memory could not even be stated without referring to internal knowledge. The behaviorist vocabulary literally could not express the finding. Around the same time, Karl Lashley argued in a landmark 1951 lecture that rapid skilled sequences, speech and piano playing among them, unfold too fast for each movement to be a response to feedback from the last: the sequence must be planned internally in advance.

The third crack was the decisive one, and it came from linguistics. In 1957 Skinner published Verbal Behavior, extending conditioning to language. In 1959 a young linguist named Noam Chomsky published a review of it that became more famous than the book. Chomsky's core points: virtually every sentence you produce or understand is one you have never heard before, so language cannot be a repertoire of reinforced responses; children converge on the grammar of their community without explicit teaching, producing rule-revealing errors like "goed" that no adult modeled and no one reinforced; and the notions of stimulus and reinforcement, applied to ordinary talk, become so stretched that they explain nothing. What speakers have, Chomsky argued, is a generative system, a finite set of rules that produces an unbounded set of sentences. That system is a mental structure, and studying it is doing science about the mind whether behaviorists approved or not.

Key idea: Behaviorism fell to converging counterevidence: latent learning and cognitive maps in rats, chunk-based memory limits that could not be stated without internal knowledge, planned rapid sequences, and above all language's unbounded novelty, which no stimulus-response repertoire could generate.

The machine that made the mind respectable

Cracks alone do not make a revolution; you also need somewhere to go. The digital computer, arriving in the same decade, supplied it. Here was a physical object, indisputably obeying the laws of physics, whose behavior could only be usefully explained by talking about information: what it stored, what it computed, what its program did. Nobody explains a chess program by listing voltages. The computer proved by existence that "inner states" and "processing" could be rigorous engineering vocabulary rather than mysticism. In 1956 Allen Newell and Herbert Simon unveiled the Logic Theorist, a program that proved theorems in formal logic, and the case was made concrete: a machine was doing something that, done by a human, everyone would call thinking.

This gave psychology a new self-image, the mind as an information-processing system: the same system might be described at the level of its program (the mind) or its hardware (the brain), and the program level was scientifically legitimate on its own. Psychologists began building flowchart models with boxes for stores and arrows for processes, measuring reaction times to estimate processing stages, and collaborating with the new field of artificial intelligence. That is the intellectual environment the 1956 MIT symposium crystallized and the Sloan hexagon later mapped.

An honest caveat: the revolution narrative flatters us. Historians note that cognitive research never fully vanished in Europe (Piaget in Geneva and Bartlett in Cambridge worked on mental structure straight through the behaviorist decades), that behaviorism evolved rather than died, and that the "defeat" took twenty years of ordinary science, not one dramatic review. And the computer metaphor, productive as it was, smuggled in assumptions (discrete symbols, serial steps) that later research would challenge, as you will see when we reach neural networks in Lesson 5.

Key idea: The computer legitimized mentalism: a physical system whose behavior demands explanation in terms of information and computation. The mind could be studied as the brain's program, without apology and without magic.

Marr's three levels: the tool you will use all course

Now to the tool. David Marr was a vision scientist at MIT who died of leukemia in 1980 at age 35, leaving behind a book, Vision, whose opening chapters gave the field its clearest statement of what explaining a cognitive ability requires. Marr argued that any information-processing system must be understood at three distinct levels, and that confusion follows whenever we mix them up.

The computational level asks: what problem is the system solving, and why? Not what happens inside, but what is the task, what counts as getting it right, and what constraints does the world impose. The algorithmic level asks: what representations and procedures does the system actually use to solve it? The implementational level asks: how is that algorithm physically realized, in neurons or silicon or anything else?

Marr's own example was a cash register. Computationally, it does addition: the task is defined by the abstract laws of arithmetic, and you can say a great deal about what any register must do (order of items cannot matter, adding nothing changes nothing) before opening the machine. Algorithmically, a particular register might represent numbers in decimal or binary and add right-to-left with carrying, or some other way. Implementationally, the algorithm might run in brass gears, transistors, or a bored cashier's head. The levels constrain each other but do not dictate each other: one computation can be served by many algorithms, one algorithm by many implementations.

Apply it to vision, Marr's real target. Computational level: from a flat, noisy, two-dimensional retinal image, recover the three-dimensional layout of surfaces that caused it, exploiting regularities of the physical world (surfaces are mostly continuous, illumination changes slowly, objects are rigid). Algorithmic level: what intermediate representations does the visual system build, edges, depth maps, object models, and by what procedures? Implementational level: how do retinal cells, the thalamus, and visual cortex carry it out? Three different research programs, all necessary, none reducible to the others.

Why does this matter so much that we will invoke it in nearly every lesson? Because it disciplines arguments. When someone says "memory is just neurons firing," they are collapsing three levels into one; the firing is the implementation, and it does not answer what problem memory solves or what algorithm it runs. When someone says a language model "thinks like a person" because its answers match human answers, they are arguing from the computational level to the algorithmic level, and that inference is invalid without further evidence: same task performance, possibly utterly different procedure. When the imagery debate of Lesson 5 asks whether mental images are pictures or descriptions, it is asking an algorithmic question that behavioral data at the computational level struggled to settle. Marr gives you the x-ray glasses for all of it.

Key idea: Marr's levels: computational (what problem, and why), algorithmic (what representations and procedures), implementational (what physical machinery). One level never settles another for free, and most bad arguments about minds and machines mix levels.

The revolution's scorecard

What did the revolution actually win? A licensed vocabulary: representation, computation, storage, retrieval. A method: infer hidden processes from behavioral signatures such as reaction times and error patterns, then triangulate with models and, later, brain data. A federation: the hexagon of Lesson 1. And a research agenda that this course now follows: perception, attention, memory, language, reasoning. What it did not win is equally instructive. It did not settle what representations are (Module 2 fights that battle), it did not banish learning from the story (statistical learning returns with force in Modules 4 and 6), and it did not prove the mind is a serial symbol cruncher (neural networks reopened that in the 1980s). Revolutions start arguments; they rarely finish them.

Common misconceptions

• "Behaviorism was obviously stupid." It was a rigorous response to introspection's real failure, it produced durable therapies and training methods, and its core insight lives on in reinforcement learning. It was wrong as a complete account, not worthless.
• "Chomsky proved language is not learned." The 1959 review showed reinforced-response learning cannot account for language. Whether and how much of grammar is innate versus learned by other means (statistical learning, social inference) is a live debate you will meet in Module 4.
• "The cognitive revolution happened overnight in 1956." The shift took roughly two decades, cognitive work had continued in Europe throughout, and behaviorist methods never left the toolkit. 1956 is a milestone, not a light switch.
• "The computer metaphor claims brains work like desktop computers." The claim is about levels of description: minds can be studied as information processing regardless of hardware. Whether cognition is serial, symbolic, or anything like a von Neumann machine is an empirical question the field still argues about.
• "Once we map the brain, the algorithmic and computational levels will be unnecessary." Marr's point is the reverse: a complete wiring diagram without a theory of the task and the algorithm explains nothing, just as reading a program's machine code does not tell you it computes payroll.

Recap

Behaviorism banned talk of the mind for defensible reasons and built real science on stimulus and response. It cracked under converging evidence: Tolman's rats learned without reward and navigated by cognitive maps, Miller's memory limit counted knowledge-defined chunks, Lashley's skilled sequences demanded advance planning, and Chomsky's review of Skinner showed that language's unbounded novelty cannot be a conditioned repertoire. The digital computer supplied the escape route, proving that inner states and processing could be rigorous vocabulary, and the information-processing view of mind was born. David Marr then gave the field its clearest map of what explanation requires: the computational level (what problem and why), the algorithmic level (what representations and procedures), and the implementational level (what physical machinery). Keep Marr's levels in your pocket. You will use them on brain scans, on mental imagery, on infant looking times, and on large language models before this course is done.

Sources

1. Encyclopaedia Britannica. (2025). Noam Chomsky. Encyclopaedia Britannica. britannica.com
2. Graham, G. (2023). Behaviorism. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. Wikipedia. (2025). Cognitive revolution. Wikimedia Foundation. en.wikipedia.org
4. Wikipedia. (2025). David Marr (neuroscientist). Wikimedia Foundation. en.wikipedia.org
5. Rescorla, M. (2020). The computational theory of mind. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu

The big picture

Whatever the mind turns out to be, it runs on cells. Your brain contains roughly 86 billion neurons, each connected to thousands of others, making perhaps a hundred trillion connections in total. Nothing about any single one of these cells is smart. A neuron is a bag of salty water that does one basic thing: it adds up incoming signals and decides whether to pass a signal along. Yet stack enough of them together, wired the right way, and you get a system that reads, jokes, grieves, and does cognitive science. This lesson is about the bottom of Marr's ladder, the implementational level: what the units are, how they compute, how they change when you learn, and what their sheer slowness tells us, surprisingly, about the algorithms running on them.

Keep one question in front of you throughout: what does knowing about neurons buy the study of the mind? The answer is not everything, as we saw last lesson. But it is far from nothing, because hardware constrains software. By the end you will be able to run that argument yourself, with numbers.

The cell that decides

A neuron has three working parts. Dendrites are branching fibers that receive chemical signals from other neurons. The soma, or cell body, integrates those signals. The axon is a single long output cable, sometimes a meter long in the nerve running to your toe, that carries the neuron's own signal away, often wrapped in myelin, a fatty insulation that speeds transmission the way cladding speeds a signal down a fiber. At rest, a neuron holds an electrical charge across its membrane of about minus 70 millivolts, maintained by pumps that shuttle sodium and potassium ions.

Signals arrive at the dendrites as small pushes. Some are excitatory, nudging the cell toward firing; some are inhibitory, pushing it away. The soma sums these pushes over space and time. If, and only if, the total crosses a threshold, the neuron fires an action potential: a wave of electrical depolarization that races down the axon at up to about 120 meters per second. The action potential is all-or-none. Like a fired gun, it either happens fully or not at all; a stronger stimulus does not make a bigger spike, it makes spikes come faster. Intensity is coded in firing rate and in how many neurons fire, not in spike size.

At the axon's end, the signal must cross the synapse, a gap of about 20 to 40 nanometers, and here electricity becomes chemistry. The arriving spike releases neurotransmitters, molecules such as glutamate (the workhorse excitatory transmitter), GABA (the workhorse inhibitory one), dopamine, and serotonin, which drift across the gap and bind to receptors on the next cell's dendrites, delivering the next round of pushes. Nearly every psychoactive substance you have heard of, from caffeine to antidepressants, works by meddling somewhere in this synaptic conversation.

Key idea: A neuron integrates thousands of excitatory and inhibitory inputs and fires an all-or-none spike when the sum crosses threshold. Information is carried by firing rates and populations, and passed between cells chemically at synapses.

Why that is computation

Summation plus a threshold may sound humble, but in 1943 Warren McCulloch and Walter Pitts proved something remarkable: idealized neuron-like units, wired appropriately, can implement the basic operations of logic (AND, OR, NOT), and networks of such units can, in principle, compute anything a digital computer can. A cell that fires only when two inputs arrive together is computing AND; a cell silenced by an inhibitory input is computing NOT. Real neurons are messier and richer than these cartoons, with dendrites that do local processing of their own, but the proof established the crucial point: the brain's units are the right kind of thing to be computing machinery. The mind does not float above the meat; the meat has the logic inside it.

The threshold matters for a second reason: it makes the neuron nonlinear, meaning its output is not a simple scaled copy of its input. Chains of purely linear elements can only ever compute linear functions, which are hopelessly weak. Nonlinear units stacked in layers can approximate essentially any function, a fact that underwrites both your visual system and the artificial networks of Module 6.

Learning is synaptic change

If knowledge lives anywhere physical, it lives in the strengths of synapses. The founding conjecture came from Donald Hebb in 1949: when one neuron repeatedly helps fire another, the connection between them strengthens. The slogan, coined later, is "cells that fire together wire together." For decades this was theory. Then in 1973 Timothy Bliss and Terje Lomo, stimulating pathways in rabbit hippocampus, discovered long-term potentiation (LTP): brief intense activity leaves synapses strengthened for hours or weeks, exactly the durable, activity-dependent change Hebb predicted. Eric Kandel's work on the sea slug Aplysia, an animal with large identifiable neurons, traced simple learning (habituation and sensitization of a gill-withdrawal reflex) to specific, measurable synaptic changes, work that earned the 2000 Nobel Prize in Physiology or Medicine.

Be careful about the strength of this claim. The evidence that synaptic plasticity is the primary mechanism of learning is very strong; blocking LTP chemically impairs new spatial learning in rats. But "memory is stored at synapses" is a claim about implementation, not a full theory of memory: it does not tell you what is encoded, how retrieval works, or why you forgot your password but remember a 20-year-old jingle. Those are algorithmic questions, and Lesson 8 tackles them with behavioral evidence. Levels, always levels.

Key idea: Learning's physical basis is activity-dependent synaptic change: Hebbian strengthening, demonstrated as LTP and traced cell by cell in Aplysia. That anchors memory in biology without yet explaining memory's algorithms.

From cells to codes

What do neurons actually respond to in a thinking animal? In the late 1950s David Hubel and Torsten Wiesel put microelectrodes into the visual cortex of cats and found cells with startlingly specific tastes: one neuron fired to an edge of light tilted at 45 degrees in one part of the visual field, another to a vertical edge, another to an edge moving left. Cells early in the pathway preferred simple features; cells deeper in preferred combinations. Vision, their work suggested, is built as a hierarchy of feature detectors, simple parts assembled into complex wholes, and the discovery won the 1981 Nobel Prize.

How far up does the hierarchy go? Famously, researchers joked about a "grandmother cell" that fires only for your grandmother. In 2005, recordings from epilepsy patients found neurons responding selectively to particular famous people, one cell to pictures of Jennifer Aniston, another to Halle Berry, even to their written names. Tempting headline: your brain has one cell per person you know. Honest reading: these cells are part of sparse population codes. Each concept is carried by a smallish ensemble of cells, each cell participates in several ensembles, and the experimenters could test only a few hundred images, so "responds to Aniston" really means "responds to Aniston among the pictures we happened to show." Representation in the brain is a team sport, robust because losing any single player barely matters.

The floor plan, briefly

You need only a rough map for this course. The wrinkled outer sheet is the cerebral cortex, where most of the action in later lessons happens: the occipital lobe at the back handles vision; the temporal lobes at the sides handle hearing, object recognition, and language comprehension; the parietal lobes on top handle spatial attention and the body sense; the frontal lobes handle movement, planning, and control, with the prefrontal cortex as the brain's deliberation department. Each hemisphere mostly serves the opposite side of the body and visual world. Beneath the cortex sit older structures you will meet again: the hippocampus (forming new long-term memories), the amygdala (threat and emotional salience), the basal ganglia (habits and action selection), and the cerebellum (coordination and timing, holding more neurons than the rest of the brain combined). Functions localize, but only somewhat: any interesting cognitive act, reading this sentence included, engages networks spanning many regions. The nineteenth-century phrenologists were wrong in detail but onto something real: the brain is not a uniform mush, it is a committee of specialists that never adjourns.

And the committee rebuilds itself. In a famous 2000 study, Eleanor Maguire and colleagues scanned London taxi drivers, who spend years memorizing the city's streets, and found the posterior hippocampus enlarged relative to controls, more so with more years on the job. String players show expanded cortical maps for the fingers of the left hand. Plasticity, the brain's capacity to reorganize with experience, is lifelong, though it is strongest in youth, as the critical periods of Module 4 will show.

Key idea: Cortex divides labor across lobes, subcortical structures handle memory formation, emotion, and habit, and concepts are carried by sparse ensembles rather than single cells. Specialization is real but every cognitive act is a network affair, and experience reshapes the wiring throughout life.

What slow hardware tells you about the software

Here is the payoff argument, and it is a beautiful piece of levels reasoning. A neuron is slow. From input to spike to the next cell takes on the order of a few milliseconds, and firing rates rarely exceed a few hundred spikes per second. A modern computer executes billions of serial steps per second; your brain's components manage perhaps a thousand. Yet you recognize a face, a task no program could touch for decades, in about 150 milliseconds. Divide 150 milliseconds by the few milliseconds a neural step costs and you get the famous 100-step constraint, articulated by Jerome Feldman: whatever algorithm your brain uses for recognition can involve at most about one hundred serial steps. No hand-written serial program of one hundred instructions could recognize a face. Conclusion: the brain's algorithms must be massively parallel, billions of slow units working simultaneously, each doing something tiny.

Notice the direction of the inference. We did not need to know the algorithm to learn something important about it; the implementation constrained the algorithmic level from below. This is why cognitive science keeps neuroscience in the hexagon rather than treating the brain as an interchangeable chip: hardware facts prune the space of possible minds. It is also, historically, one of the arguments that revived neural network models in the 1980s, a thread we pick up in Lesson 5.

Common misconceptions

• "We only use 10 percent of our brains." False in every version. Imaging shows activity throughout the brain over any ordinary day, damage to almost any region costs something, and evolution does not maintain expensive idle tissue. The brain is 2 percent of body weight consuming roughly 20 percent of resting energy.
• "Stronger stimuli make bigger neural spikes." Action potentials are all-or-none. Intensity is coded by firing rate and by how many neurons participate, not by spike amplitude.
• "Each memory or concept is stored in one dedicated cell." The Jennifer Aniston findings show sparse, selective ensembles, not one-cell-per-concept storage. Distributed codes survive cell loss; a literal grandmother cell would be a single point of failure.
• "Logical, analytical people are left-brained; creative people are right-brained." The hemispheres do show specializations (language is usually left-lateralized), but imaging studies find no evidence that individuals have a dominant hemisphere matching a personality type. Every complex task uses both sides.
• "The adult brain is fixed." Taxi drivers, musicians, and rehabilitation patients all show measurable structural change with experience. Plasticity declines with age but never reaches zero.

Recap

Neurons receive chemical signals on dendrites, integrate them in the soma, and fire all-or-none action potentials down the axon when a threshold is crossed, passing the message chemically at synapses. Summation plus threshold makes them nonlinear computing elements, and McCulloch and Pitts showed such units can implement logic. Learning is grounded in Hebbian synaptic plasticity, demonstrated as LTP and dissected in Aplysia. Hubel and Wiesel revealed vision's feature-detector hierarchy, and modern recordings show concepts carried by sparse population codes rather than single cells. The cortex divides labor across occipital, temporal, parietal, and frontal lobes with hippocampus, amygdala, basal ganglia, and cerebellum beneath, all of it plastic with experience. And the hardware's slowness yields the 100-step argument: recognition in 150 milliseconds on millisecond components forces massively parallel algorithms. Next lesson: the instruments that let us watch this system think, and the ways they fool us.

Sources

1. National Institute of Neurological Disorders and Stroke. (2023). Brain basics: The life and death of a neuron. National Institutes of Health. ninds.nih.gov
2. National Institute of Neurological Disorders and Stroke. (2023). Brain basics: Know your brain. National Institutes of Health. ninds.nih.gov
3. Encyclopaedia Britannica. (2024). Neuron. Encyclopaedia Britannica. britannica.com
4. Nobel Prize Outreach. (1981). The Nobel Prize in Physiology or Medicine 1981: Press release. nobelprize.org
5. Wikipedia. (2025). Hebbian theory. Wikimedia Foundation. en.wikipedia.org

The big picture

Suppose you want to know how a factory works, but you may not enter the building. You can measure the hum from outside, minute by minute. You can photograph the heat coming off the roof, but only in blurry three-second exposures. You can study what stops being produced after a fire damages one wing. And occasionally, with permission, you can cut power to a room for a few seconds and see what breaks. None of these gives you the assembly line. All of them together give you a working theory.

That is neuroscience's evidence problem, and this lesson is about the instruments. You will meet the four workhorses: EEG, fMRI, lesion studies, and stimulation. More importantly, you will learn to read their claims critically, because brain imaging produces the most seductive and most frequently overstated figures in all of cognitive science. A colored blob on a brain is not an explanation. By the end of this lesson you should be the person at the table who asks the awkward question about the statistics.

Two currencies: time and space

Every brain method trades in two resolutions. Temporal resolution is how precisely you can say when something happened; spatial resolution is how precisely you can say where. Cognition happens in tens of milliseconds and in millimeters, and no single non-invasive method gives you both.

Method	Measures	Time resolution	Space resolution	Best for
EEG	Electrical fields from synchronized cortical activity, at the scalp	About 1 millisecond	Poor (centimeters, ambiguous source)	Timing and sequence of processing stages
MEG	Magnetic fields from the same currents	About 1 millisecond	Moderate (better source estimates than EEG)	Timing with somewhat better localization
fMRI	Blood oxygenation change (BOLD), an indirect proxy for neural activity	Poor (1 to 2 seconds; the response lags by 4 to 6)	Good (about 1 to 3 mm, whole brain including deep structures)	Where in the brain, including subcortical areas
Lesions	Behavior after damage to a region	Not applicable	Depends on lesion boundaries	Whether a region is necessary for a function
TMS	Behavior during brief magnetic disruption of a region	Good (tens of milliseconds)	Moderate, cortical surface only	Causal tests in healthy volunteers

Read the table once more with an eye on the last two rows, because they carry a special weight. EEG and fMRI are correlational: they tell you what the brain was doing while the person performed a task. Lesions and TMS are causal: they tell you what happens to the task when the region is unavailable. The difference is the difference between watching an orchestra and removing the cellos.

Key idea: No non-invasive method has both good timing and good localization. EEG buys milliseconds at the cost of place, fMRI buys millimeters at the cost of time, and only lesion and stimulation methods speak to necessity rather than correlation.

EEG: the millisecond method

Put electrodes on the scalp and you record the summed electrical activity of millions of cortical neurons whose fields happen to align. Raw EEG is a noisy scribble, so researchers use event-related potentials (ERPs): present the same kind of event dozens of times, average the traces time-locked to onset, and random noise cancels while the stimulus-driven response emerges. What survives are named waves that have become tools of the trade.

The N400 is a negative deflection about 400 milliseconds after a word that does not fit its context, discovered by Marta Kutas and Steven Hillyard in 1980. "He spread his warm bread with socks" produces a large N400 at "socks"; "with butter" produces almost none. It is now a standard index of semantic processing. The P600, by contrast, responds to grammatical violations, evidence that meaning and structure are handled somewhat separately, a claim we return to in Module 4. The P300 tracks surprise and attention allocation to rare events. These waves let you time the mind: they show, for instance, that the brain registers a semantic anomaly in under half a second, before you could possibly report noticing it.

EEG's weakness is the inverse problem: many different arrangements of sources inside a conducting head can produce the same pattern of voltages at the scalp, so the location of the generator cannot be uniquely recovered from the recording. When you see an EEG paper describing a "frontal" effect, that is a statement about electrodes, not a confident anatomical claim.

fMRI: the pictures everyone has seen

Functional magnetic resonance imaging does not measure neural firing. It measures the BOLD signal: blood oxygen level dependent contrast, which tracks the fact that active tissue draws extra oxygenated blood, and that oxygenated and deoxygenated hemoglobin have slightly different magnetic properties. This is a plumbing signal, sluggish and indirect: the hemodynamic response peaks roughly 5 seconds after the neural activity it reflects. Cognitive events lasting 200 milliseconds get smeared across seconds.

Two consequences follow, and both are widely misunderstood. First, fMRI images are almost always subtraction images. The brain is never off; blood flows everywhere all the time. So researchers compare two conditions, faces versus houses, say, and display the difference. The colored blob does not mean "this is where face perception happens." It means "this region responded more to faces than to houses in this contrast, on average, across many trials and usually across many participants." Change the comparison condition and the blob moves.

Second, the statistics are enormous. A brain is divided into tens of thousands of small volumes called voxels, and a naive analysis runs a separate statistical test on each one. At the conventional threshold of p less than 0.05, you expect one false positive in twenty tests by chance alone. With 40,000 voxels, you expect around 2,000 spurious "active" spots even in a brain doing nothing of interest. This is the multiple comparisons problem, and correcting for it properly is the difference between a result and a decoration.

Key idea: fMRI measures blood flow, not spikes, with a lag of seconds, and its images are statistical contrasts between conditions rather than pictures of where a function lives. Its tens of thousands of simultaneous tests demand rigorous correction.

The dead salmon

In 2009, a graduate student named Craig Bennett and colleagues put an Atlantic salmon in an fMRI scanner. The salmon was dead, purchased at a market. They showed it photographs of humans in social situations and asked it, via the standard task instructions, to determine what emotion each person was experiencing. Then they analyzed the data exactly as many published papers did at the time, without correcting for multiple comparisons.

They found a cluster of significant "activity" in the dead salmon's brain cavity.

The study, presented as a poster and later published, won an Ig Nobel Prize and made a serious point unforgettable. The salmon was not evidence that fMRI does not work; it was evidence that uncorrected statistics manufacture findings. Bennett's team showed that with proper correction (controlling the family-wise error rate or the false discovery rate) the salmon's brain went silent, as it should. The lasting effects were methodological: correction is now expected, methods sections must state it, and the salmon became the standard teaching example of why. When you read a neuroimaging claim, ask two questions: was the multiple comparisons correction stated, and how many participants were there? Small samples plus flexible analysis choices are how neuroscience produced its own share of findings that failed to replicate.

Key idea: The dead salmon experiment showed that uncorrected voxel-wise statistics can find "activation" in a dead fish. Proper correction for multiple comparisons is not a technicality; it is the line between signal and noise.

Reverse inference: the seductive backwards step

Here is the most common error in popular science writing about the brain, named and analyzed by Russell Poldrack in 2006. A forward inference is legitimate: "when people feel disgust, the insula activates." A reverse inference runs the other way: "the insula activated, therefore the participant felt disgust." That step is only valid if the region is selectively involved in that one function, and almost no brain region is. The insula also activates during pain, interoception, risk, effort, and hearing your own heartbeat. Poldrack's analysis of a large imaging database showed how weak the backwards inference typically is once you account for how many different tasks light up any given region.

Watch for the pop-science version: "brain scans show that using your phone activates the same reward regions as cocaine, so phones are addictive." Nearly everything pleasant activates parts of the reward system, including looking at photos of your children and hearing music you like. The claim smuggles a strong conclusion out of a weak, backwards inference. A useful mental habit: whenever a headline draws a psychological conclusion from a blob, ask "what else activates that region?"

Two related cautions. A 2008 critique by Edward Vul and colleagues, provocatively titled with the phrase "voodoo correlations," pointed out that some social neuroscience papers reported implausibly high brain-behavior correlations because they selected voxels using the same data they then used to compute the correlation, a circular procedure. And a 2016 study by Anders Eklund and colleagues found that common software's cluster-inference assumptions could inflate false positive rates well beyond the nominal 5 percent in certain analyses. The field responded to all three critiques by tightening standards, preregistering analyses, sharing data, and running larger samples. That self-correction is exactly what you should want from a science, and it is worth saying plainly: these were embarrassments that made the field better.

Necessity: lesions and stimulation

If correlational methods cannot establish that a region is required, what can? Historically, damage. When Paul Broca examined a patient in 1861 who could understand speech but produce almost none, and found damage in the left frontal lobe at autopsy, he made the first strong localization claim in the history of the field. Neuropsychology built on this, and its most powerful tool is the double dissociation: patient A can do task 1 but not task 2, patient B can do task 2 but not task 1. That pattern argues that the two tasks rely on at least partly separate machinery, and it cannot be explained away by one task simply being harder.

Lesion evidence has its own limits, and they are real. Natural lesions follow blood vessels, not functional boundaries. A damaged region might be necessary as a relay rather than as the seat of the function, the way cutting a phone line disables a conversation without the line doing the talking. Brains reorganize after injury, so the observed deficit may reflect a rebuilt system rather than the original one. And single famous patients, however illuminating, are single data points. Modern work therefore combines lesion mapping across many patients with transcranial magnetic stimulation (TMS), which uses a magnetic pulse to briefly disrupt a cortical region in a healthy volunteer, producing what is sometimes called a virtual lesion lasting tens of milliseconds. TMS gives causal evidence with timing, but it reaches only the surface of the brain and carries its own confounds, including the loud click and scalp sensation that require careful control conditions.

Key idea: Only interference methods establish necessity. Double dissociations from lesion patients, supported by TMS in healthy volunteers, give causal traction that correlational imaging cannot, at the cost of imprecise damage and small numbers.

What good practice looks like now

Put it together and you get the field's current standard: converge. A strong claim about brain and cognition typically shows a behavioral effect, a timing signature from EEG or MEG, a localization from fMRI, and where possible a causal test from patients or TMS, ideally in preregistered studies with samples large enough to detect the effect. Increasingly, researchers also use multivariate pattern analysis, which asks whether the pattern of activity across many voxels can decode what a person is seeing or thinking, rather than asking whether any single region got more active. This is a genuine advance: it treats representation as distributed, matching what we learned about population codes last lesson.

Finally, keep Marr in view. Even a perfect answer to "where and when" is an implementational answer. It constrains but does not deliver the algorithm. That is why the next lesson turns from instruments to the field's central theoretical fight: what a mental representation actually is.

Common misconceptions

• "fMRI shows the brain in action, in real time." It shows blood oxygenation changes lagging neural activity by several seconds, averaged over many trials, displayed as a contrast between two conditions.
• "A blob means that function lives there." A blob means greater response in one condition than another. Regions participate in many functions, and the same task engages networks, not islands.
• "The dead salmon proved fMRI is junk." It proved that uncorrected statistics are junk. With proper multiple-comparisons correction the salmon showed nothing, which was precisely the authors' point.
• "If the amygdala activates, the person is afraid." That is reverse inference. The amygdala responds to salience and novelty broadly, so activity there does not identify a specific emotion without further evidence.
• "Lesion studies are outdated now that we have scanners." Imaging is correlational. Patients and TMS remain the primary evidence that a region is necessary, which is a different and stronger claim.

Recap

Brain methods trade timing against localization. EEG and MEG resolve milliseconds and give us ERP components like the N400 and P600 but cannot pin down sources; fMRI resolves millimeters throughout the brain but measures a sluggish, indirect blood signal and produces contrast images, not maps of where functions live. Its scale creates a multiple comparisons problem so severe that uncorrected analysis found activation in a dead salmon, a result that permanently changed reporting standards. Reverse inference, reading a psychological state off a region's activity, is invalid when regions serve many functions, and critiques of circular analysis and inflated cluster statistics pushed the field toward preregistration, data sharing, and larger samples. Necessity claims still rest on lesion double dissociations and TMS. Good cognitive neuroscience converges across all of these, and even then it answers the implementational question, not the algorithmic one.

Sources

1. Bennett, C. M., Baird, A. A., Miller, M. B., and Wolford, G. L. (2010). Neural correlates of interspecies perspective taking in the post-mortem Atlantic salmon. Journal of Serendipitous and Unexpected Results, 1(1), 1-5. Summary: en.wikipedia.org
2. Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59-63. pubmed.ncbi.nlm.nih.gov
3. National Institute of Biomedical Imaging and Bioengineering. (2023). Magnetic resonance imaging (MRI). National Institutes of Health. nibib.nih.gov
4. Encyclopaedia Britannica. (2024). Electroencephalography. Encyclopaedia Britannica. britannica.com
5. Wikipedia. (2025). Transcranial magnetic stimulation. Wikimedia Foundation. en.wikipedia.org

The big picture

Think of your kitchen. Now, without moving, count the windows in it.

Most people report doing something odd to answer that: they seem to look around an inner scene. Something in your head stood in for a kitchen you are not currently in, and you used it to extract information you never explicitly memorized. That standing-in-for is representation, and it is the central theoretical concept of cognitive science. Every explanation in this course leans on it: perception builds representations of surfaces, memory stores representations of events, language maps sounds onto representations of meaning.

So what is one, actually? Not where in the brain it lives, but what kind of thing it is. This is Marr's algorithmic level, and it has hosted the field's longest-running fight. Two big answers have been on offer since the beginning, and a famous third debate, about mental images, shows how hard it is to settle such questions with behavior alone. Fair warning: this lesson ends without a winner, because the field has not declared one. Learning to hold an unresolved question precisely is part of your training.

The classical answer: mind as symbol manipulation

The cognitive revolution's first theory was inherited from logic and computing. On this view, thinking is symbol manipulation: the mind contains discrete symbols standing for things (CAT, RED, GRANDMOTHER), and rules that combine and transform those symbols according to their form. Newell and Simon put it as the physical symbol system hypothesis in 1976: a physical symbol system has the necessary and sufficient means for general intelligent action. Jerry Fodor sharpened the psychological version as the language of thought hypothesis: thought has a syntax, an internal combinatorial code often nicknamed Mentalese.

Two arguments made this compelling, and they remain the best case for symbols. The first is productivity: you can entertain an unbounded number of thoughts, including brand-new ones ("a purple hippopotamus reciting the tax code"), which suggests a finite vocabulary plus recursive combination rules, exactly like a language. The second is systematicity, Fodor and Zenon Pylyshyn's 1988 argument: anyone who can think "the dog chased the cat" can also think "the cat chased the dog." These abilities come as a package, never separately, which is precisely what you expect if thoughts have constituent structure and the same parts can be recombined. A system that memorized whole thoughts as unstructured wholes would have no reason to show that pattern.

The symbolic approach delivered. It gave us the first working AI (theorem provers, chess programs, expert systems), Chomsky's generative grammar, production-rule models of problem solving, and cognitive architectures such as ACT-R and SOAR that still make quantitative predictions about reaction times today. Its weaknesses were equally clear: symbolic systems were brittle, degrading catastrophically at the edges of their rules, terrible at perception, and needed everything hand-specified. And there was an awkward physical fact: nobody could say what a symbol was in neural terms.

Key idea: The classical view holds that thought is rule-governed manipulation of structured symbols. Productivity and systematicity are its strongest evidence: thoughts recombine like sentences, which suggests they have parts and a syntax.

The connectionist answer: mind as pattern of activation

In 1986 a two-volume book by David Rumelhart, James McClelland, and the PDP Research Group proposed a different architecture, reviving and extending earlier neural network ideas. In a connectionist model, there are no symbols and no rules. There are simple units, loosely inspired by neurons, connected by weighted links. A representation is not a symbol sitting in a slot; it is a pattern of activation distributed across many units. Knowledge is not a list of propositions; it is the pattern of connection weights, learned gradually from examples by adjusting weights to reduce error.

This buys several things the symbolic view struggled with. Distributed representations degrade gracefully: damage some units and performance declines gradually rather than crashing, which matches both brain damage and human error. Similar inputs get similar activation patterns automatically, so generalization to new cases falls out of the architecture rather than needing to be programmed. Learning is statistical and driven by exposure, matching the way children get better at things gradually.

The flagship demonstration was Rumelhart and McClelland's model of English past tense. Children show a famous U-shaped curve: first correct irregulars ("went"), then a stage of overregularization ("goed"), then correct forms again. The standard explanation was that the child had acquired a rule and was over-applying it. The connectionist network, trained on verb pairs with no rules at all, produced the same U-shaped pattern purely from statistics of exposure. Steven Pinker and Alan Prince fired back in 1988 with a detailed critique of what the model got wrong, and the argument continues, but the demonstration made a permanent point: behavior that looks rule-governed does not prove that rules are in the head. You cannot read the algorithm off the behavior.

Fodor and Pylyshyn's counterattack was the systematicity argument above: nothing in a network's design guarantees that a system able to represent "dog chases cat" can represent "cat chases dog," because the network need not have separable constituents. Connectionists have answered with structured architectures and, more recently, with the observation that large trained networks do show substantial compositional generalization, though how much and how reliably remains actively contested. The honest summary in 2026 is this: neither camp won outright, and most working researchers hold a hybrid view in which some cognition is well described by structured symbolic operations and much of it by learned statistical pattern completion.

Key idea: Connectionism represents with distributed activation patterns and stores knowledge in learned weights, buying graceful degradation, automatic generalization, and gradual learning. The past-tense model showed rule-like behavior can emerge without explicit rules.

A comparison worth memorizing

Feature	Symbolic (classical)	Connectionist
What a representation is	A discrete structured symbol	A distributed pattern of activation
Where knowledge lives	Explicit rules and stored propositions	Connection weights learned from examples
Processing style	Serial, rule application	Massively parallel constraint satisfaction
Effect of damage	Often catastrophic failure	Graceful degradation
Strongest domain	Reasoning, planning, grammar	Perception, motor skill, statistical learning
Chief criticism	Brittle, hand-specified, hard to map to neurons	Systematicity not guaranteed, hard to interpret

Notice how each column's strength is the other's weakness. That asymmetry is the reason hybrid proposals keep reappearing, and it is worth carrying into Module 6, where the same argument returns wearing modern clothes as the debate about what large language models do and do not have.

The imagery debate: a case study in settling questions

Now to the kitchen windows. When you counted them, were you using something picture-like, or were you using descriptions that merely feel picture-like? This is the imagery debate, and it ran for thirty years between Stephen Kosslyn (images are depictive, preserving spatial structure) and Zenon Pylyshyn (images are propositional descriptions, and the picture-feel is an epiphenomenon).

The classic evidence came first. In 1971 Roger Shepard and Jacqueline Metzler showed people pairs of three-dimensional block figures and asked whether they were the same shape or mirror images. The finding was clean and beautiful: response time increased linearly with the angular difference between the two figures. Rotate one figure 40 degrees and people were fast; rotate it 140 degrees and they were proportionally slower, as if turning an object in the head at a roughly constant rate. Kosslyn added image-scanning studies: memorize a map of an island, then imagine moving from the hut to the tree, and the time taken scales with the distance on the actual map. These results feel decisive. Something with spatial properties is being manipulated.

Pylyshyn's reply was sharp, and this is the part worth learning. Perhaps participants are not scanning a picture; perhaps they are simulating what would happen if they were looking at the real object, because they know from experience that far things take longer to reach and bigger rotations take longer to perform. This is the tacit knowledge objection. On this account the timing data reflect the task instructions and world knowledge, not the format of the representation. Supporting his case, Pylyshyn pointed to cognitive penetrability: imagery effects can change when participants are given different beliefs or expectations, which pictures on a screen would not do. He also noted that images are not really like pictures at all; nobody can look at a mental image of a tiger and count its stripes, and ambiguous figures do not usually flip in imagination the way they do in perception.

Later evidence shifted the balance without ending the argument. Neuroimaging found that visual imagery activates early visual cortex in a way that respects spatial layout: imagining a large object activates a different, more peripheral part of the retinotopic map than imagining a small one. Patients with damage to visual cortex can show corresponding deficits in imagery. That is a hard fact for a purely propositional account. But a determined critic can still ask whether that activation is doing the representational work or is a byproduct, and Pylyshyn maintained his position to the end. Meanwhile a genuinely surprising finding entered from the side: aphantasia, described in modern form by Adam Zeman in 2015, in which people report no voluntary visual imagery at all yet perform normally on many spatial tasks, including mental rotation. Whatever imagery is, some people accomplish the same tasks without the experience of it.

Key idea: Shepard and Metzler's linear rotation times and Kosslyn's scanning data suggest depictive, spatially structured images; Pylyshyn's tacit knowledge and penetrability arguments show behavioral timing alone cannot fix representational format. Imaging supports depiction, aphantasia complicates everything, and the debate remains open.

What this fight teaches about the field

Step back and notice the shape of all three disputes. In each case, the disagreement was not about the data. Everyone accepted that people say "goed," that rotation time scales with angle, that networks learn from examples. The disagreement was about what the data licensed you to conclude about the underlying format and procedure. That is the algorithmic level being genuinely hard: many algorithms can produce the same behavior, which is why the field needs the extra leverage of computational modeling, neural evidence, patient dissociations, and cross-linguistic comparison.

It also explains why cognitive science is not simply a settled body of results to memorize. It is a live argument with rules of evidence. When you meet a confident claim in the remaining modules, whether about attention, false memory, or machine understanding, run this lesson's question at it: what would the alternative account predict, and does anything in the data rule it out?

Common misconceptions

• "Connectionism won because deep learning works." Practical success in engineering does not settle a psychological question about human representation. Modern debates about compositional generalization in neural networks are the same systematicity argument, still unresolved.
• "Rule-like behavior proves there are rules in the head." The past-tense model produced the U-shaped overregularization curve with no rules at all. Behavior underdetermines the algorithm, which is the whole lesson.
• "Mental images are pictures in the brain." Even Kosslyn's depictive view does not claim a picture is viewed by an inner eye. Images lack picture properties (you cannot count a remembered tiger's stripes), and positing an inner viewer just relocates the problem.
• "Imaging settled the imagery debate." Early visual cortex involvement is strong evidence for depiction, but critics can still ask whether that activity is doing the work or accompanying it. Aphantasia shows the conscious experience is separable from the performance.
• "Symbols are obsolete." Symbolic architectures still produce quantitative predictions of human reaction times and errors, and language, planning, and mathematics remain their strongest turf. Most researchers now expect a hybrid story.

Recap

A mental representation is an internal state that stands in for something and can be operated on. The classical view says representations are discrete structured symbols manipulated by rules, supported by productivity and systematicity: thoughts recombine like sentences. Connectionism says representations are distributed activation patterns and knowledge is learned connection weights, supported by graceful degradation, automatic generalization, and the past-tense model's demonstration that rule-like behavior needs no rules. Each view's strength is the other's weakness, so hybrids abound. The imagery debate shows why such questions are hard: Shepard and Metzler's linear mental-rotation times and Kosslyn's scanning results argue for depictive images, Pylyshyn's tacit knowledge objection shows timing data cannot fix format by itself, neuroimaging of early visual cortex supports depiction, and aphantasia reminds us that experience and performance can come apart. The general lesson is the durable one: behavior underdetermines algorithm, so cognitive science must converge.

Sources

1. Pitt, D. (2022). Mental representation. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
2. Buckner, C., and Garson, J. (2019). Connectionism. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. Thomas, N. J. T. (2021). Mental imagery. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
4. Aydede, M. (2024). The language of thought hypothesis. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
5. Wikipedia. (2025). Mental rotation. Wikimedia Foundation. en.wikipedia.org

The big picture

Seeing feels like the least mysterious thing you do. Open your eyes and the world is simply there, in color, at a distance, arranged in solid objects. No effort, no thought. That effortlessness is the greatest illusion your brain produces, because vision is arguably the hardest computation you perform, and roughly a third of your cortex is working on it right now.

Here is the plan. First, the problem: why the image on your retina cannot possibly determine what is in front of you, and what follows from that. Then the solution: perception as inference, guessing the most likely cause of an image using assumptions built in by evolution and experience. Then the payoff: illusions stop being party tricks and become experimental evidence about which assumptions your visual system makes. And finally, the argument about how far knowledge and expectation reach down into seeing, which is more contested than popular accounts admit.

The inverse problem

Light strikes your retina and produces a two-dimensional array of intensities and wavelengths. From that, you must recover a three-dimensional world. The trouble is that this recovery is formally impossible, because the mapping runs one way. A large object far away and a small object nearby project exactly the same image. A gray patch under bright light and a white patch in shadow can reflect identical amounts of light to your eye. An ellipse on the retina could be an ellipse facing you or a circle tilted away. Infinitely many worlds are consistent with any given image. This is the inverse problem, and it is why vision cannot work by simply reading off what is there.

So how do you ever see anything correctly? By adding information the image lacks. Hermann von Helmholtz proposed the answer in the 1860s and it has never been improved on in outline: perception is unconscious inference. The visual system takes the ambiguous image as evidence and computes the most probable cause, using assumptions about how the world usually behaves. Modern versions state this in probabilistic terms: the brain combines the incoming evidence with prior expectations about which scenes are likely, and settles on the interpretation with the best overall fit.

What are those assumptions? Some of the well-supported ones: light usually comes from above (we evolved under one sun in the sky); objects are usually rigid and continuous; surfaces are usually opaque; things that move together usually belong together; viewpoints are usually generic rather than freakishly coincidental. None of these are always true. All of them are true often enough to be a good bet.

Key idea: The retinal image underdetermines the world, so seeing must be inference: combining ambiguous evidence with built-in assumptions about how scenes usually are, to compute the most probable cause of the image.

Illusions as experiments

Now the crucial reframe. If perception is inference from assumptions, then an illusion is not a malfunction. It is what happens when a scientist constructs an image in which a normally excellent assumption gives the wrong answer. Illusions are how you find out what the assumptions are, exactly as a well-designed experiment isolates a variable.

Consider Edward Adelson's checker-shadow illusion. A checkerboard sits with a cylinder casting a shadow across it. Two squares, one a light square inside the shadow and one a dark square outside it, look obviously different in shade. They are printed in identical gray. Your visual system is not measuring the light arriving from each patch, which is the same. It is solving a better problem: estimating the surface reflectance, the actual paint on the surface, by discounting the illumination. A patch that sends you medium light from inside a shadow must be a light surface; the same medium light in bright illumination must be a dark surface. The illusion is your visual system doing its job correctly on a scene engineered to punish it. That is why knowing the answer does not dispel the effect: this inference is not under your control.

The same logic explains a family of size illusions. In the Ponzo illusion two identical bars sit across converging lines, and the upper one looks larger because converging lines signal depth and your system applies size constancy: an object that is farther away but projects the same retinal size must be physically bigger. The Ames room, a distorted chamber viewed through a peephole, makes people appear to grow and shrink as they walk across it, because the room is built to project the retinal image of a normal rectangular room and your assumption that rooms are rectangular is stronger than your belief that people do not change size. The moon illusion, where the moon looks enormous near the horizon and small overhead, is still debated in detail but is generally treated as another constancy phenomenon involving perceived distance.

Key idea: Illusions are not failures of vision; they are the predictable output of normally useful assumptions applied to engineered inputs. The checker-shadow illusion shows the system computing surface color rather than measuring light.

Building blocks and grouping

Underneath the inference sits real machinery, and we met the beginning of it in Lesson 3. Retinal cells feed a hierarchy in which early cortical cells detect oriented edges and later stages respond to increasingly complex combinations. Above these, the visual system organizes fragments into objects using principles catalogued by the Gestalt psychologists in the early twentieth century: elements that are close together group (proximity), elements that look alike group (similarity), lines continue smoothly rather than turning sharply (good continuation), incomplete outlines are perceived as whole (closure), and things that move together group (common fate). These are not arbitrary; each corresponds to a statistical regularity of real scenes, which is why they are useful.

The functional architecture divides further. Two broad cortical streams, described by Leslie Ungerleider and Mortimer Mishkin and later refined by Melvyn Goodale and David Milner, run from visual cortex: a ventral stream toward the temporal lobe supporting recognition (what an object is), and a dorsal stream toward the parietal lobe supporting spatially guided action (how to reach for it). The evidence is a double dissociation of the kind you learned to value in Lesson 4. Patient D.F., after carbon monoxide poisoning damaged ventral regions, could not report the orientation of a slot in front of her, yet posted a card through it accurately, her hand rotating correctly en route. Patients with optic ataxia from parietal damage show the reverse: they describe objects fine but misreach for them. Seeing for knowing and seeing for doing are partly separate systems, and only one of them is the one you are conscious of.

Key idea: Vision builds edges into objects using Gestalt grouping principles that mirror real-world regularities, and splits into a ventral stream for recognition and a dorsal stream for action, demonstrated by the D.F. double dissociation.

How far down does knowledge reach?

Bottom-up processing is driven by the stimulus: edges, motion, color, arriving from the retina upward. Top-down processing is driven by knowledge, context, and expectation, flowing downward to shape interpretation. Both plainly exist. Context effects are easy to demonstrate: the same ambiguous middle character reads as H in TAE CAT and as A in the same position of another word, and you never notice the ambiguity. Degraded images that look like meaningless blotches snap into a clear dalmatian once you are told there is a dog, and afterward you cannot unsee it. Anatomically, feedback connections running from higher to lower visual areas outnumber the feedforward ones, so the wiring supports influence in both directions.

But here is where a good course should slow down, because the popular version of this idea overreaches. In the 1950s the New Look movement reported that poor children overestimated coin sizes and hungry people saw food words more readily, and concluded that motivation directly changes seeing. Many such findings proved fragile or turned out to reflect response bias, that is, what participants were willing to report rather than what they perceived. A widely discussed 2016 review by Chaz Firestone and Brian Scholl argued that essentially all claimed cases of beliefs and desires penetrating perception suffer from a small set of confounds: effects on judgment rather than perception, effects mediated by where attention was directed, or demand characteristics from the task. Their target was not context effects within vision, which are uncontroversial, but the stronger claim that wanting something makes it look bigger.

So the honest position is layered. Vision is thoroughly inferential and richly context-sensitive within its own processing. Whether high-level beliefs and desires reach into early perception is contested, and the strongest popular claims are the weakest empirically. Notice that this is Pylyshyn's cognitive penetrability question from last lesson, arriving in a new domain, which is a good sign you are learning a field rather than a list of facts.

What vision cost AI, and what that taught us

One last piece of evidence for how hard this is. In 1966 Seymour Papert at MIT set a summer project for undergraduates: connect a camera to a computer and have it describe what it sees. The problem was expected to take a summer. Object recognition in unconstrained scenes took roughly half a century, and arrived only with large neural networks trained on millions of labeled images. Meanwhile a toddler does it constantly, in a fraction of a second, from far less data. This gap, sometimes called Moravec's paradox, is one of the most instructive facts in cognitive science: what feels effortless to us is usually the hardest computation, because evolution has already solved it and hidden the work.

Common misconceptions

• "Vision works like a camera recording the world." A camera records light. Vision infers surfaces, objects, and causes from light, discarding the raw measurements. That is why identical gray patches can look different and why size constancy exists at all.
• "Illusions show your senses are unreliable." They show your senses use assumptions, which are usually correct. Illusions are engineered exceptions, and they are evidence rather than embarrassment.
• "Knowing about an illusion should make it go away." Most perceptual inference is impenetrable to belief. You can know two squares are identical and still see them as different, which itself tells you something about where the processing happens.
• "You see everything in your visual field in full detail." High resolution comes only from the fovea, a region about the size of a thumbnail at arm's length. The impression of a rich, detailed world comes partly from the fact that a saccade brings detail wherever you check.
• "Wanting something makes it literally look bigger." Many such claims have failed to replicate or reduce to judgment and attention effects. Context effects within vision are solid; motivational penetration of early perception is not.

Recap

The retinal image is consistent with infinitely many worlds, so vision must add assumptions to the evidence, an idea Helmholtz named unconscious inference and modern work casts in probabilistic terms. Illusions become experiments once you see this: the checker-shadow illusion reveals a system computing reflectance rather than measuring light, and Ponzo and the Ames room reveal size constancy running on engineered depth cues. Underneath, feature detectors feed Gestalt grouping principles that mirror real-world regularities, and processing splits into a ventral what stream and a dorsal how stream, demonstrated by patient D.F.'s ability to post a card through a slot she could not describe. Top-down influence is real within vision, though the strongest claims about beliefs and desires reshaping early perception have not held up well. And the decades AI spent failing at object recognition are a reminder that effortless does not mean simple. Next: why you nevertheless miss a gorilla walking through the middle of it all.

Sources

1. Encyclopaedia Britannica. (2024). Perception. Encyclopaedia Britannica. britannica.com
2. Green, E. J., and Schellenberg, S. (2023). The contents of perception. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. National Eye Institute. (2022). How the eyes work. National Institutes of Health. nei.nih.gov
4. Wikipedia. (2025). Checker shadow illusion. Wikimedia Foundation. en.wikipedia.org
5. Wikipedia. (2025). Two-streams hypothesis. Wikimedia Foundation. en.wikipedia.org

The big picture

In 1999 Daniel Simons and Christopher Chabris asked people to watch a short video of two teams passing basketballs and to count the passes made by the team in white. Partway through, a person in a gorilla suit walks into the middle of the scene, faces the camera, thumps their chest, and strolls off. The gorilla is on screen for about nine seconds. Roughly half of viewers, counting passes, do not see it. Not "did not notice at first," but stare at the screen and insist afterward there was no gorilla, then look genuinely shocked on replay.

The study won an Ig Nobel Prize and became one of psychology's most famous demonstrations, and it is the perfect entry to this lesson because it breaks an assumption you did not know you had: that whatever is in front of your open eyes is in your experience. It is not. Attention decides, and you do not get a receipt for what was thrown away. This lesson covers why the mind must select, how it selects, what happens to the unselected, and what all of this means for the very practical question of whether you can drive and talk on the phone.

Why selection is compulsory

Your senses deliver an enormous stream. The optic nerve alone carries on the order of a million fibers per eye, and comparable floods arrive from hearing and touch. Meanwhile the systems that matter for deliberate action, working memory, decision, speech, are drastically limited: you can hold a handful of items in mind, say one sentence at a time, reach for one thing at a time. Somewhere between the flood and the trickle, most information must be discarded. That funnel is attention: the selection of some information for deeper processing at the expense of the rest.

Attention is not one thing. Researchers distinguish selective attention (focusing on one source among many), divided attention (attempting several tasks at once), sustained attention or vigilance (maintaining focus over time), and executive attention (resolving conflict between competing responses). It also splits by control: endogenous or voluntary attention, which you direct by goals, and exogenous or stimulus-driven attention, which is captured by a sudden movement or a loud noise whether you like it or not. That second kind is why a notification pulls your eyes before you decide anything, and why the design of your phone is not a neutral fact about your willpower.

Key idea: The senses deliver far more than downstream systems can use, so selection is compulsory. Attention comes in voluntary and stimulus-driven forms and in selective, divided, sustained, and executive varieties.

The cocktail party and the filter

The modern study of attention began with a practical problem: air traffic controllers hearing several pilots at once. Colin Cherry in 1953 built the laboratory version, dichotic listening: different messages into each ear, with instructions to shadow (repeat aloud) one of them. People do this reasonably well, and afterward they can report almost nothing about the unattended ear. Not the content, not the language, not whether it was speech played backwards. They do notice physical properties: whether the voice was male or female, whether it stopped, whether a pure tone sounded.

Donald Broadbent's 1958 filter theory explained this with a bottleneck. Information from all channels enters a brief sensory buffer, a filter selects one channel by physical characteristics (which ear, what pitch), and only the selected channel passes for meaning analysis. Elegant, testable, and quickly complicated by an inconvenient finding. About a third of participants notice their own name in the unattended ear, the cocktail party effect Cherry described: across a noisy room you are filtering out a dozen conversations, then someone says your name and it arrives. If the filter blocked by physical features before meaning was analyzed, your name would be as invisible as any other word.

Anne Treisman's attenuation theory (1964) fixed this gracefully: the filter turns unattended channels down rather than off. Signals with low thresholds, your name, a shout of "fire," a word primed by context, can still get through the attenuated stream. Late-selection theorists, notably Deutsch and Deutsch, argued instead that everything is analyzed for meaning and selection happens at the response end. Decades of work produced a compromise that Nilli Lavie formalized as load theory: the locus of selection depends on how demanding the task is. Under high perceptual load, selection is early and unattended material is barely processed; under low load, spare capacity spills over and irrelevant information gets analyzed whether you want it to or not. This is why a boring task makes you more, not less, distractible.

Key idea: Broadbent's filter explained why unattended speech is not understood; the cocktail party effect forced Treisman's attenuation account. Load theory reconciles early and late selection: demanding tasks select early, easy tasks leave capacity for distraction.

Blindness of two kinds

Back to the gorilla. Inattentional blindness is the failure to notice a fully visible but unexpected object when attention is engaged elsewhere. Simons and Chabris found detection rates around 50 percent, and later work showed the effect is not a laboratory curiosity: in a 2013 study, 20 of 24 expert radiologists searching lung scans for nodules missed an image of a gorilla, 48 times the size of a typical nodule, inserted into the scan. Their eye-tracking showed many of them looked directly at it. Trained expertise did not protect them; if anything, focused expert search made it worse.

Its cousin is change blindness: failing to notice large changes in a scene when the change is masked by a flicker, a cut, or an interruption. In a famous 1998 study by Simons and Daniel Levin, an experimenter stopped pedestrians to ask for directions; while they talked, two people carried a door between them, and behind the door the experimenter was swapped for a different person, different height, different clothes, different voice. Roughly half the pedestrians continued the conversation without noticing they were now talking to a stranger.

What do these findings license you to conclude? Be careful, because this is where popular summaries go wrong. They do not show that people are stupid or that vision is broken. They show that visual awareness is far sparser than it feels, and that the impression of a complete, detailed world is a construction supported by the fact that detail is available whenever you look for it. They also do not show that unattended information has no effect at all; unnoticed stimuli can still prime later responses. And there is a healthy critical literature: some early demonstrations were less robust than they seemed, effect sizes vary with the design, and the widely repeated claim that missed gorillas prove eyewitnesses are useless overstates a real but bounded finding. The reliable core, replicated many times, is that unexpected objects are frequently missed when attention is loaded.

Key idea: Inattentional blindness and change blindness show that awareness requires attention: fully visible objects and large scene changes go unnoticed when attention is engaged elsewhere, even in trained experts looking straight at them.

How attention finds things

A quick look at the mechanism. Treisman's feature integration theory proposed two stages. Basic features (color, orientation, motion) are registered in parallel across the whole field, which is why a red dot among green dots "pops out" and search time barely changes as you add distractors. But binding features together into an object requires attention, applied serially, which is why searching for a red vertical bar among red horizontal and green vertical bars gets steadily slower as the display grows. The prediction that misapplied attention should produce illusory conjunctions, seeing a red X when a red O and a blue X were present, was confirmed under brief exposures. Michael Posner's cueing paradigm added the geometry: a valid cue speeds detection, an invalid one slows it, and attention can move without the eyes moving at all.

Multitasking, honestly

Now the applied payoff, and here the evidence is unusually clear and unusually ignored. When two tasks both require central attention, people do not truly perform them simultaneously; they alternate, and the switching costs time and accuracy. Laboratory task-switching studies show reliable switch costs in reaction time and errors. Studies of student learning find that media multitasking during study predicts worse comprehension and lower grades.

The driving research is the sharpest case. David Strayer and colleagues at Utah ran driving simulator studies comparing conversation on a phone with legally intoxicated driving. Drivers on the phone showed slower reactions to braking events and more missed signals, with impairment comparable in some measures to a blood alcohol content of 0.08 percent. Crucially, hands-free devices did not help meaningfully: the impairment comes from attention, not from hands. Phone conversation differs from talking to a passenger, who sees the traffic and pauses when it gets difficult. Strayer's group also documented inattentional blindness at the wheel: drivers on the phone looked at but failed to recall billboards and other objects directly in their field of view. Meanwhile most people rate themselves above-average drivers and above-average multitaskers; research on the small population of so-called supertaskers finds them to be a tiny fraction of the population, and the people most confident that they multitask well tend to perform worst.

Key idea: Multitasking is rapid switching with real costs. Phone conversation while driving produces impairments comparable to legal intoxication in simulator studies, hands-free does not fix it, and self-assessed multitasking ability is inversely related to actual ability.

Common misconceptions

• "If it is in front of my eyes, I see it." Half of viewers miss a chest-thumping gorilla, and expert radiologists miss one in a lung scan while looking at it. Awareness requires attention, not just an open eye.
• "Hands-free calling makes driving safe." The impairment is attentional. Simulator studies find hands-free conversation impairs driving about as much as handheld, and comparably to a 0.08 blood alcohol level on some measures.
• "Some people are natural multitaskers." Genuine supertaskers appear to be a very small minority, and confidence in one's multitasking is a poor predictor of performance, often a negative one.
• "Unattended information is completely blocked." Treisman's attenuation account and the cocktail party effect show that highly relevant items such as your own name get through a turned-down channel.
• "Change blindness proves memory is worthless." It shows that we do not maintain a detailed internal snapshot across interruptions, which is a claim about how vision uses the world as its own memory, not a claim that memory fails in general.

Recap

Sensory input vastly exceeds what downstream systems can use, so attention must select. Dichotic listening showed unattended speech is not understood, motivating Broadbent's early filter; the cocktail party effect forced Treisman's attenuation model, and load theory settled the early-versus-late dispute by making the locus of selection depend on task demands. Inattentional blindness, from the invisible gorilla to the radiologists' scans, and change blindness, from the swapped-stranger study, show that awareness of even large, visible events depends on attention. Feature integration theory explains why some searches pop out while others require serial binding, with illusory conjunctions as evidence. And the applied conclusion is blunt: divided attention is switching, switching is costly, and phone conversation impairs driving at levels comparable to legal intoxication regardless of where your hands are. Next we ask what happens to the small amount of information that does get through: how it is stored, and how badly it can be rebuilt.

Sources

1. Simons, D. J., and Chabris, C. F. (1999). Gorillas in our midst: Sustained inattentional blindness for dynamic events. Perception, 28(9), 1059-1074. pubmed.ncbi.nlm.nih.gov
2. Strayer, D. L., and Johnston, W. A. (2001). Driven to distraction: Dual-task studies of simulated driving and conversing on a cellular telephone. Psychological Science, 12(6), 462-466. pubmed.ncbi.nlm.nih.gov
3. American Psychological Association. (2006). Multitasking: Switching costs. APA. apa.org
4. Encyclopaedia Britannica. (2024). Attention. Encyclopaedia Britannica. britannica.com
5. Wikipedia. (2025). Inattentional blindness. Wikimedia Foundation. en.wikipedia.org

The big picture

In 1953 a 27-year-old man with severe epilepsy underwent surgery that removed much of his medial temporal lobes on both sides, including most of the hippocampus. The seizures improved. But Henry Molaison, known in the literature as H.M. until his death in 2008, could no longer form new lasting memories. He would greet the researcher Brenda Milner warmly every session for decades, always as a stranger. He could hold a conversation, so short-term memory worked. He could recall his childhood, so old memories survived. And in a finding that reshaped the field, he could learn new motor skills, improving day after day at tracing a shape seen only in a mirror, while insisting each morning that he had never done the task before.

One patient, three conclusions: memory is not a single thing; the ability to form new memories is separable from the ability to hold and retrieve old ones; and knowing how to do something is a different system from knowing that something happened. This lesson maps those systems, then turns to the more unsettling half of the story: that remembering is reconstruction, not replay, and that the difference has consequences in courtrooms.

The architecture, briefly

The standard model, from Richard Atkinson and Richard Shiffrin in 1968, has three stages: brief sensory registers, a limited short-term store, and a vast long-term store, with attention controlling entry to the second and rehearsal or deeper processing controlling entry to the third. It is a simplification, and the middle box in particular has been rebuilt.

Alan Baddeley and Graham Hitch replaced passive short-term storage with working memory, a system for holding and manipulating information during a task. Their model has a phonological loop for verbal material (the inner voice you use to keep a phone number alive), a visuospatial sketchpad for visual and spatial material, a central executive that allocates attention between them, and, added later, an episodic buffer that binds information across formats. The evidence for separate subsystems is behavioral and clean: a concurrent verbal task interferes with verbal storage far more than with spatial storage, and vice versa. Capacity is small. Miller's seven plus or minus two was for passive span; estimates for working memory in tasks requiring manipulation cluster nearer four chunks, and individual differences in working memory capacity correlate with reading comprehension and fluid reasoning.

Long-term memory divides too. Declarative (or explicit) memory covers what you can consciously report, and splits into episodic memory for personally experienced events located in time and place (your last birthday) and semantic memory for facts detached from any learning episode (Paris is the capital of France). Nondeclarative (implicit) memory covers what shows up in performance without conscious recollection: procedural skills such as cycling, classical conditioning, and priming, the facilitation of processing by prior exposure. H.M.'s mirror-drawing improvement was procedural learning intact alongside episodic learning destroyed, and later patients with striatal damage show the reverse pattern. A double dissociation again.

Key idea: Memory is several systems, not one. Working memory holds and manipulates a few chunks across separable verbal and spatial subsystems; long-term memory divides into declarative (episodic and semantic) and nondeclarative (procedural, conditioning, priming) forms with distinct neural bases.

Encoding, storage, retrieval

Any memory failure happens at one of three stages, and knowing which matters. Encoding is getting information in. Fergus Craik and Robert Lockhart's levels of processing framework showed that what determines later recall is not how long you rehearse but how deeply you process: judging whether a word is printed in capitals produces poor memory, judging whether it rhymes does better, and judging whether it fits meaningfully into a sentence does best. Elaboration, connecting new material to what you already know, is the practical version, and it is why explaining something to someone else works better than rereading.

Storage involves consolidation, the gradual stabilizing of a memory trace over hours to years, which depends on the hippocampus early and increasingly on cortex later, and which is substantially supported by sleep. Retrieval is getting information out, and here the key finding is that memory is cue-dependent. Endel Tulving's encoding specificity principle says retrieval succeeds to the extent that cues at recall match those present at encoding. Divers who learned word lists underwater recalled them better underwater than on land, and vice versa. The everyday version is the tip-of-the-tongue state: the information is stored, the cue is wrong. Most "bad memory" is a retrieval failure, not an empty shelf.

Remembering is rebuilding

Now the central conceptual point of the lesson. It feels as if remembering is playing back a recording. It is not. Frederic Bartlett showed this in the 1930s by having British participants read a Native American folk tale, "The War of the Ghosts," and recall it repeatedly over weeks. The recalls did not simply fade; they transformed. Unfamiliar elements were dropped, odd details were rationalized, canoes became boats, and the story drifted toward the participants' own cultural expectations. Bartlett proposed that we store gist plus schemas, organized knowledge structures, and rebuild the episode at retrieval, filling gaps with what usually happens.

Reconstruction is efficient and mostly accurate about the gist. But it means every act of remembering is an act of construction, and construction can incorporate material that was never there. This is not a rare pathology. It is how the system works.

Key idea: Memory stores gist and schemas rather than recordings, and rebuilds episodes at retrieval. Reconstruction is normally useful, and it is exactly what makes memories vulnerable to distortion.

Loftus and the misinformation effect

Elizabeth Loftus turned that insight into one of psychology's most consequential research programs. In a 1974 study with John Palmer, participants watched a film of a car accident and were asked how fast the cars were going when they "hit" each other, or "smashed," "collided," "bumped," or "contacted." A single verb changed speed estimates by several miles per hour. A week later, participants asked the "smashed" question were more than twice as likely to report having seen broken glass. There was no broken glass in the film.

This is the misinformation effect: information introduced after an event becomes incorporated into the memory of the event. It has been replicated hundreds of times with leading questions, co-witness discussion, and suggestive interviews. Loftus went further with the "lost in the mall" procedure, in which a plausible false childhood event is described by a trusted family member alongside true ones. A minority of participants, roughly a quarter in the original small study, came to report remembering the false event, sometimes elaborating details never supplied. Later work using more elaborate suggestive interviews produced higher rates for some kinds of events, and a 2016 review by Alan Scoboria and colleagues pooling multiple studies estimated that around 30 percent of participants formed something like a false memory, with about half showing some accepting or believing response.

Read those numbers carefully, because this is where honesty matters. The findings do not show that memory is generally fictional, that most memories are false, or that anyone can be made to remember anything. Rates vary enormously with the plausibility of the suggested event, the strength of the suggestion, and the participant. A 2023 reanalysis debate over the mall study's methods reminds us that even famous findings deserve scrutiny. What the literature does establish robustly is that confident, detailed, sincerely held memories can be substantially wrong, and that the confidence a person feels is a poor guide to accuracy once suggestive procedures have occurred.

The applied stakes are not hypothetical. The Innocence Project reports that mistaken eyewitness identification was a contributing factor in a large majority of DNA exoneration cases in the United States. In response, a 2014 National Research Council report recommended reforms grounded in this science: double-blind lineup administration so the officer cannot cue the witness, unbiased instructions saying the suspect may not be present, recording confidence immediately at first identification (where it is more diagnostic), and video recording of identification procedures. This is applied cognitive science at its best, and a useful counterexample to the idea that the field is only theory.

Key idea: Post-event information reliably contaminates memory, and suggestive procedures can create detailed false memories in a substantial minority of people. Confidence tracks accuracy poorly after suggestion, which is why lineup procedures were reformed.

What about flashbulb memories?

You may object that you vividly remember where you were during some major public event. Those are flashbulb memories, and they have been studied carefully. Ulric Neisser and Nicole Harsch collected accounts the morning after the 1986 Challenger explosion and again two and a half years later. Many accounts had changed substantially, some completely, and participants shown their own handwritten original accounts sometimes insisted the earlier version must be wrong. Later large-scale studies after September 11, 2001 found the same pattern: confidence and vividness stay high while accuracy for details declines much like ordinary memory. Vividness is not a validity check. Nothing in your experience of remembering tells you whether the memory is accurate.

Why forgetting is a feature

All this can sound like a catalog of defects, so end with the reframe. Consider Solomon Shereshevsky, the mnemonist studied by Alexander Luria for thirty years, who could recall lists of dozens of items decades later. His memory was a burden: he struggled to grasp abstract ideas or recognize faces reliably, because every encounter left an unmerged particular. Detail without abstraction is not intelligence.

Forgetting does useful work. It discards outdated information (last year's parking space, your old password), it strips particulars to leave generalizations that transfer, and it prioritizes what recurs and matters. Ebbinghaus's forgetting curve, measured on himself with nonsense syllables in the 1880s, shows steep initial loss then a long slow tail, and modern accounts read this as adaptive: the probability that a memory will be needed again falls in roughly the same way. Retrieval-induced forgetting, in which recalling some items suppresses related ones, keeps competitors from crowding a target.

And two robust findings turn all this into study advice worth more than the course fee. The testing effect: retrieving information from memory strengthens it far more than restudying does. Roediger and Karpicke showed that students who read a passage once and then practiced recalling it outperformed students who read it four times, on tests a week later, even though the rereaders felt more confident. The spacing effect, documented since Ebbinghaus: study sessions distributed across days beat the same total time massed together. Both feel worse while you do them, which is why most students choose the methods that fail. Difficulty during study, what Robert Bjork calls a desirable difficulty, is often the signature of learning rather than its absence.

Key idea: Forgetting is adaptive, discarding the outdated and abstracting the general, as the mnemonist Shereshevsky's difficulties illustrate. Retrieval practice and spaced study reliably outperform rereading and cramming, despite feeling harder and less productive.

Common misconceptions

• "Memory works like video recording." Memory stores gist and rebuilds episodes using schemas. Bartlett's transforming folk tale and the misinformation effect follow directly from that architecture.
• "Vivid, confident memories are accurate memories." Flashbulb memory studies after Challenger and September 11 found high confidence with substantially changed content. Vividness is not evidence of accuracy.
• "Anyone can be made to remember anything." False memory rates depend heavily on plausibility, suggestion strength, and the individual, and a substantial share of participants resist. The robust claim is that sincere, detailed memories can be wrong, not that memory is generally fictional.
• "Amnesia erases who you are, as in films." H.M. retained his identity, childhood memories, vocabulary, and skills. Anterograde amnesia impairs forming new declarative memories while leaving procedural learning largely intact.
• "Rereading and highlighting are efficient study methods." They produce fluency that feels like learning. Testing yourself and spacing sessions produce better retention on delayed tests, despite feeling more effortful.

Recap

Patient H.M. showed that memory is multiple systems: intact short-term and remote memory alongside destroyed new declarative learning and preserved procedural learning. Working memory holds about four chunks across separable verbal and spatial subsystems under a central executive; long-term memory splits into declarative (episodic and semantic) and nondeclarative (procedural, conditioning, priming) forms. Deep, elaborative encoding beats rote rehearsal, consolidation stabilizes traces over time with help from sleep, and retrieval is cue-dependent, so most forgetting is a retrieval failure. Bartlett established that remembering is reconstruction from gist and schemas, and Loftus showed that reconstruction admits post-event misinformation, producing changed details and, in a substantial minority under suggestive procedures, whole false events. Confidence and vividness are poor accuracy signals, as flashbulb studies show, which is why eyewitness identification procedures were reformed. And forgetting is not a bug: it abstracts and prunes, while retrieval practice and spacing exploit the architecture to make learning stick.

Sources

1. Loftus, E. F., and Palmer, J. C. (1974). Reconstruction of automobile destruction. Journal of Verbal Learning and Verbal Behavior, 13(5), 585-589. Overview: en.wikipedia.org
2. American Psychological Association. (2023). Memory. APA Dictionary and topic pages. APA. apa.org
3. Encyclopaedia Britannica. (2024). Memory. Encyclopaedia Britannica. britannica.com
4. National Institute of Neurological Disorders and Stroke. (2023). Brain basics: Know your brain. National Institutes of Health. ninds.nih.gov
5. Wikipedia. (2025). Henry Molaison. Wikimedia Foundation. en.wikipedia.org

The big picture

You are doing something right now that no other species does. Not communicating, plenty of animals communicate. You are decoding an unbounded set of novel structured messages from a small inventory of arbitrary sounds and marks, at a rate of several words per second, and you learned the system before you could tie your shoes, without lessons, from imperfect input, in a few years.

Every intact human child does this. Children raised in poverty do it, children of illiterate parents do it, deaf children do it in sign languages that are full languages with their own grammar. Meanwhile a chimpanzee raised in a human home with intensive training does not. Language is the clearest case in cognitive science of a capacity that appears both biologically special and environmentally shaped, which is why it sits at the center of the hexagon. This lesson covers what the system is, how children get it, and the argument, still live, about how much of it must be built in.

Design features: what "special" means precisely

Charles Hockett proposed a checklist in the 1960s to make the comparison rigorous rather than chauvinistic. Several features appear in animal systems too, so focus on the ones that do the real work.

Arbitrariness: the relation between form and meaning is conventional. Nothing about the sound "dog" resembles a dog, which is why chien, perro, and inu work equally well. Discreteness: language is built from a small inventory of categorical units (English uses roughly 40 phonemes) that combine, rather than from a continuous signal. Duality of patterning: meaningless units (sounds) combine into meaningful ones (words), which combine into larger structures, so a few dozen sounds yield hundreds of thousands of words. Displacement: you can talk about the absent, the past, the future, the hypothetical, and the false. Vervet monkey alarm calls signal a leopard present now; no vervet warns about the leopard that might come tomorrow. Productivity: the system generates and interprets an unlimited number of novel messages, including sentences never before produced in human history, such as this one about a leopard's tax return.

Add one more that Hockett underweighted. Human syntax is recursive: structures can embed inside structures of the same type. "The cat the dog chased ran" embeds a clause in a clause. This is what makes productivity unbounded rather than merely large, and it is why linguists describe grammar as a generative system rather than an inventory. Whether recursion is universal across all human languages became a heated dispute after Daniel Everett's claims about Pirahã, an Amazonian language he argued lacks embedded clauses; the debate remains unresolved and is worth knowing about as a caution against tidy universals.

Key idea: Human language is arbitrary, discrete, doubly patterned, displaced, productive, and recursive. Animal systems have some features but no known natural system combines them, which is why "language" is not simply advanced communication.

What the animal studies actually showed

The apes deserve a fair hearing, because the popular story swings between two exaggerations. Beginning in the 1960s, several projects taught great apes sign language or symbol boards, since their vocal tracts cannot produce speech. Washoe, a chimpanzee raised by Allen and Beatrix Gardner, acquired a substantial sign vocabulary. Kanzi, a bonobo studied by Sue Savage-Rumbaugh, learned symbols apparently by observation rather than training and responded correctly to many novel spoken requests, including instructions like putting one object into another.

These are real achievements: apes clearly learn symbols, use them to request and refer, and comprehend more than they produce. But the systematic reviews are sobering about the grammar. Herbert Terrace's project with a chimpanzee pointedly named Nim Chimpsky analyzed videotape and found that Nim's multi-sign strings were mostly imitations or partial repetitions of what the teacher had just signed, showed little structural regularity, and did not lengthen with development the way children's utterances do. Ape sign vocabularies plateau in the hundreds while a six-year-old child has tens of thousands of words and adds several a day. The reasonable conclusion is neither "apes have language" nor "apes have nothing," but that symbol learning is shared while syntax and unbounded productivity, so far, are not.

How children do it

Now the developmental facts, which any theory has to explain. The timetable is strikingly regular across languages and cultures. Newborns already prefer their mother's voice and the rhythm of the language heard in the womb. In the first year infants perform a remarkable narrowing: at 6 months they discriminate phonetic contrasts from any language, including ones their parents cannot hear, and by about 10 to 12 months they lose sensitivity to contrasts not used around them, work associated with Patricia Kuhl and Janet Werker. Japanese-learning infants stop discriminating English r from l; English-learning infants stop hearing Hindi dental versus retroflex t. Perception is being tuned to the local system.

Babbling starts around 6 to 8 months, with deaf infants exposed to sign babbling manually on the same schedule, which tells you the drive is linguistic rather than vocal. First words come around 12 months, and a vocabulary spurt often follows. Two-word combinations appear around 18 to 24 months, and they are not random: "more milk," "daddy shoe," "no bed" show consistent word order and semantic relations. Then grammar arrives in a rush, and with it the errors that matter most: overregularization. A child who correctly said "went" at two starts saying "goed" and "foots" at three. Nobody taught this; adults never model it, and correction has notoriously little effect. The child has extracted a pattern and is applying it beyond its range, which is direct evidence that acquisition is rule extraction and not imitation.

Two further facts constrain any theory. Critical or sensitive periods: first-language acquisition after early childhood is severely impaired, as tragic isolation cases such as Genie show, and second-language attainment declines with age of first exposure in large-scale data. Deaf children exposed to sign language late show persistent grammatical deficits relative to native signers. And creolization: when children are exposed to a pidgin, a simplified contact system without full grammar, they produce a creole with systematic grammar their input did not contain. The Nicaraguan Sign Language case, documented from the 1980s, is the clearest demonstration: deaf children brought together in new schools converged on a grammatical sign language within a generation, with each cohort of younger children systematizing it further.

Key idea: Acquisition follows a regular timetable, tunes perception to the local language in the first year, and shows rule extraction through overregularization errors nobody models. Critical period effects and creolization show children contribute structure their input lacks.

The poverty of the stimulus argument

Now the argument that has organized fifty years of debate. Chomsky's poverty of the stimulus claim is that the linguistic input available to children is too impoverished, given the speed and uniformity of acquisition, to explain the grammar they end up with by general learning alone. Therefore, the argument runs, children must bring innate constraints, a universal grammar, that narrows the hypothesis space.

The textbook illustration is question formation. English yes-no questions move an auxiliary to the front: "The man is tall" becomes "Is the man tall?" A child hearing thousands of such examples could form a simple linear rule: move the first "is" to the front. Now take "The man who is tall is nice." The linear rule yields "Is the man who tall is nice?" Children do not make this error. They produce "Is the man who is tall nice?", moving the auxiliary from the main clause, which requires knowing that the rule operates over hierarchical structure, not linear order. The claim is that the relevant evidence, complex questions with embedded clauses, is rare in child-directed speech, yet children never go through a linear-rule stage. Add the further observations that children get little or no systematic negative evidence (nobody labels sentences ungrammatical for them), and that they master the system despite hearing false starts and errors.

This is a serious argument, and it is also contested at every joint. Corpus researchers have questioned the empirical claim that the crucial constructions are truly absent from input; some studies find relevant examples occur more often than assumed. Others argue that indirect negative evidence works: consistently not hearing a form is informative. Computational modelers, notably in Bayesian frameworks, have shown that learners with general-purpose biases toward simpler hierarchical grammars can converge on structure dependence from realistic input, suggesting the required innate content may be weaker and less language-specific than classic universal grammar proposed.

The strongest empirical challenge came from statistical learning. In a landmark 1996 study, Jenny Saffran, Richard Aslin, and Elissa Newport played 8-month-old infants two minutes of a continuous nonsense syllable stream with no pauses, in which syllable sequences formed hidden "words" defined only by transitional probabilities: within a word, one syllable predicted the next reliably; across word boundaries, less so. After two minutes, infants listened longer to the statistically illegal sequences, meaning they had extracted the word boundaries from probabilities alone. Infants are powerful statistical learners, and much subsequent work has shown how far this can go for phonology, word segmentation, and some syntax. Usage-based theorists such as Michael Tomasello added that children are also extraordinary social learners, using joint attention and intention reading, and argued that constructions can be built up from concrete item-based patterns without an innate grammar module.

Key idea: Poverty of the stimulus argues that fast, uniform acquisition of structure-dependent rules from limited input requires innate constraints. Statistical learning in infants, corpus reanalyses, and Bayesian models show general learning is more powerful than assumed, so how much must be innate remains open.

Where the debate stands

Most researchers today occupy some middle ground, and it is worth stating what almost nobody disputes: human children have species-specific biological preparation for language (no other animal does this), and experience with a specific language is indispensable (no child raised without input acquires one). The live questions are about content and specificity. Is the innate contribution a rich set of grammatical universals, or general biases toward hierarchical structure, plus formidable statistical and social learning machinery? Chomsky's own later minimalist program narrowed the innate core dramatically, in some formulations to little more than recursive combination. That is a considerable distance from the 1980s picture, and it happened because the opposing evidence was good.

Notice the shape of this, which repeats the pattern from Lesson 5: two camps, shared data, disagreement about what the data license, and gradual convergence under pressure from modeling and new experiments. That is what a healthy young science looks like from the inside.

Common misconceptions

• "Children learn language by imitation and reward." Overregularizations like "goed" are never modeled by adults, correction has little effect, and children produce sentences they have never heard. Imitation cannot generate a productive system.
• "Some languages are primitive or lack real grammar." Every documented human language has a full grammatical system, including creoles that emerged within a generation and sign languages, which are not gestured versions of spoken language but independent languages.
• "Apes have been taught language." Apes learn symbols and comprehend requests, which is genuinely impressive. Systematic analysis of their productions found little syntax and heavy imitation, and vocabulary plateaus far below a young child's.
• "Poverty of the stimulus has been refuted." It has been seriously challenged by statistical learning, corpus work, and computational modeling, and the innate core proposed today is much smaller. Neither total refutation nor unqualified survival is accurate.
• "Bilingual exposure confuses children and delays language." Bilingual children hit milestones on a normal schedule; vocabulary counted in one language may lag, while total conceptual vocabulary across both languages is comparable to monolinguals.

Recap

Human language is distinguished by arbitrariness, discreteness, duality of patterning, displacement, productivity, and recursion, a combination no natural animal system matches, though apes clearly learn symbols and comprehend more than they produce. Children acquire this system on a regular timetable: perceptual narrowing to native phonemes in the first year, babbling, first words, two-word combinations, then rapid grammar with overregularization errors that reveal rule extraction rather than imitation. Critical period effects and the emergence of creoles and Nicaraguan Sign Language show children supply structure their input lacks. Chomsky's poverty of the stimulus argument infers innate constraints from structure-dependent rules acquired from limited evidence, illustrated by question formation. Saffran's statistical learning results, corpus reanalyses, Bayesian models, and usage-based social learning accounts have narrowed what must be innate without eliminating the biological preparation. Next: what happens to all of this in a brain, and whether the language you speak shapes how you think.

Sources

1. Saffran, J. R., Aslin, R. N., and Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926-1928. pubmed.ncbi.nlm.nih.gov
2. Encyclopaedia Britannica. (2024). Language acquisition. Encyclopaedia Britannica. britannica.com
3. Laurence, S., and Margolis, E. (2001). The poverty of the stimulus argument. Overview in: Innateness and language. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
4. National Institute on Deafness and Other Communication Disorders. (2022). Speech and language developmental milestones. National Institutes of Health. nidcd.nih.gov
5. Wikipedia. (2025). Nicaraguan Sign Language. Wikimedia Foundation. en.wikipedia.org

The big picture

This lesson has two halves that belong together. The first asks a question about implementation: what does a brain do with language, and what happens when specific parts of it fail? The second asks a question that has escaped the laboratory and become a staple of popular culture: does the language you speak shape the way you think?

Both halves illustrate the same intellectual discipline. In each case there is a famous simple story that turns out to be too simple, a body of careful evidence that says something more limited but more interesting, and a popular version that badly overstates the finding. Learning to hold the accurate middle position, without either dismissing the phenomenon or exaggerating it, is the skill on offer here.

Two patients, two syndromes

In 1861 the French physician Paul Broca examined a patient nicknamed Tan, because "tan" was nearly the only syllable he could produce. He understood what was said to him and was clearly not demented. At autopsy Broca found damage in the left frontal lobe, in what is now called Broca's area. A decade later Carl Wernicke described patients with the opposite profile: fluent, effortless speech that was largely meaningless, together with severe comprehension problems, following damage further back in the left temporal lobe.

The two syndromes are worth knowing precisely. In Broca's aphasia, speech is effortful, halting, and telegraphic, with content words preserved and grammatical function words and endings dropped: a patient asked about a trip might manage "wife... car... Boston... two day." Comprehension is relatively preserved for ordinary conversation, but breaks down for sentences where meaning depends entirely on grammar, as in "The boy was pushed by the girl," where you cannot use plausibility to guess who did what. Patients are typically painfully aware of their deficit, which makes it a frustrating condition. In Wernicke's aphasia, speech is fluent and well-articulated with normal melody, but full of substitutions and invented words, and comprehension is poor. Patients often seem unaware that they are not making sense.

The classical model that grew from this, elaborated by Ludwig Lichtheim and later Norman Geschwind, held that Wernicke's area handles comprehension, Broca's area handles production, and a fiber bundle called the arcuate fasciculus connects them; damage the connection alone and you get conduction aphasia, in which comprehension and speech are decent but repeating what you just heard fails. It was a triumph of nineteenth-century inference: brain function localized by careful behavioral observation plus autopsy, generations before imaging.

Key idea: Broca's aphasia impairs fluent, grammatical production with relatively preserved comprehension except for syntax-dependent sentences; Wernicke's aphasia leaves fluent speech empty of meaning with impaired comprehension. Together they showed language is organized into separable components, usually in the left hemisphere.

Why the textbook picture needed revising

Here is where a good course parts company with a bad one. The Broca-Wernicke model appears in every introductory textbook, and modern language neuroscience regards it as substantially wrong in detail. Several findings forced the revision.

First, lesion-symptom mapping across many patients shows that damage confined to Broca's area does not reliably produce lasting Broca's aphasia; the persistent syndrome typically requires larger damage extending into surrounding cortex and underlying white matter. The historical patients had much bigger lesions than the classic story implies, as later MRI scans of Broca's preserved specimens confirmed. Second, Broca's area is not a speech-output box: it participates in syntactic processing during comprehension, in working memory for sequences, and in non-linguistic tasks including music and action observation. Third, comprehension is not confined to a single posterior area; imaging and lesion evidence support two broad processing streams, a ventral stream mapping sound to meaning and a dorsal stream mapping sound to articulation, in a model developed by Gregory Hickok and David Poeppel. Fourth, the syndromes themselves are clinically messy: real patients rarely fit the textbook categories cleanly, and many clinicians now describe deficits dimensionally rather than by syndrome label.

What survives is the important part. Language depends on a distributed left-lateralized network, in roughly 95 percent of right-handers and a large majority of left-handers, and different components of language, phonology, syntax, lexical access, meaning, can be selectively impaired. That is the finding cognitive science needs: language is not one undifferentiated ability, and the brain respects some of the distinctions linguists drew on purely behavioral grounds.

Key idea: The classical Broca-Wernicke box-and-arrow model has been superseded by distributed dual-stream accounts, but its core insight holds: language is left-lateralized and decomposable, with components that can fail independently.

The other half: does language shape thought?

Now to the second question, and it starts with a cautionary tale about how claims spread. In the 1930s Benjamin Lee Whorf, a fire insurance inspector and amateur linguist working with Edward Sapir, proposed what became known as the Sapir-Whorf hypothesis: that the categories of one's language shape the categories of one's thought. Whorf's most famous example, that Inuit languages have dozens or hundreds of words for snow while English has one, was inflated in retelling until it became a newspaper cliché. The anthropologist Laura Martin and later Geoffrey Pullum documented the inflation: the original claim involved a handful of roots, English itself has snow, sleet, slush, powder, blizzard, and flurry, and Inuit languages are polysynthetic, building elaborate words from roots the way English builds phrases, which makes word-counting nearly meaningless as evidence. The example proved nothing, and its collapse discredited the whole hypothesis for a generation.

Distinguish two versions before looking at evidence. Strong linguistic determinism says language determines thought, so concepts not lexicalized in your language are unthinkable. This is essentially refuted: people routinely think thoughts their language has no single word for, learn new distinctions readily, and can translate, however clumsily, between any two languages. Weak linguistic relativity says language influences habitual thought, attention, and memory, making some distinctions easier or more automatic. That version has real experimental support, and it is what serious researchers defend.

The evidence, case by case

Color. This was the first battleground. Brent Berlin and Paul Kay's 1969 cross-language survey found that basic color terms are not arbitrary: languages draw from a constrained set, and the order in which terms are added across languages is largely predictable. That looked like a decisive win for universals. But later work found relativity effects at the edges. Russian obligatorily distinguishes siniy (dark blue) from goluboy (light blue), and Russian speakers are faster at discriminating blues that cross that boundary than blues within a category, an advantage that disappears under a verbal interference task, suggesting language is involved online. Studies with the Himba of Namibia found category boundaries influencing discrimination performance. Notably, several such effects are stronger in the right visual field, which projects to the left, language-dominant hemisphere, exactly what you would expect if linguistic categories are modulating perceptual judgment.

Space. The most striking evidence. Some languages, including Guugu Yimithirr in Australia and Tzeltal in Mexico, use absolute rather than relative spatial frames: instead of "the cup is to your left," speakers say the equivalent of "the cup is to your north." Stephen Levinson and colleagues found that speakers of such languages maintain astonishingly accurate dead reckoning of cardinal direction, unconsciously, at all times, and that this shows up in nonverbal tasks: asked to reconstruct an array of objects after being rotated 180 degrees, absolute-frame speakers reconstruct it preserving cardinal directions while relative-frame speakers preserve left-right order. The groups solve a nonlinguistic memory task differently, in a way that matches their language.

Number. Peter Gordon and later Everett studied the Pirahã, whose counting vocabulary is extremely limited. On tasks requiring exact matching of quantities above about three, performance degraded with set size, while approximate quantity judgment was fine. This suggests counting words are a cognitive technology for exact large-number representation, not that speakers lack number sense.

Grammatical gender and framing. Lera Boroditsky and colleagues reported that speakers describe objects using adjectives congruent with the grammatical gender of the noun in their language, and that describing crime as a beast versus a virus shifts policy preferences. These are the studies most often cited in popular articles, and they deserve a caution: some effects in this literature have proved fragile or failed to replicate, and small samples were common. Treat the color, space, and number findings as considerably better established than the gender-metaphor ones.

Key idea: Weak relativity has genuine support: Russian blues, Himba color boundaries, absolute-frame spatial reasoning in Guugu Yimithirr and Tzeltal, and Pirahã exact-number performance all show language influencing nonlinguistic task performance. Strong determinism, and the snow-words cliché, do not survive scrutiny.

How to hold this correctly

The mechanism most researchers now favor is not that language rewrites the mind's basic architecture, but that language provides habitual categories and tools that get recruited during thinking. It directs attention (if your language obligatorily marks something, you must attend to it to speak), it supplies labels that support memory and grouping, and it can be used covertly as an inner code during tasks. That explains why relativity effects often vanish under verbal interference: the language was doing online work in the task, not permanently reshaping perception.

Notice too that this is Marr's levels again, plus the penetrability question from Lesson 5 and Lesson 6. When someone claims language shapes perception, ask: is the effect at perception, at attention, at memory, or at the response? The best studies in this literature are careful about exactly that question, and the weakest ones are not.

Common misconceptions

• "Broca's area is the speech center and Wernicke's is the comprehension center." Both regions participate in more than one function, damage confined to Broca's area does not produce the classic lasting syndrome, and modern models describe distributed dual streams.
• "Broca's aphasia spares comprehension entirely." Comprehension is relatively preserved but fails for sentences whose meaning depends on syntax alone, such as reversible passives, which is a theoretically important detail.
• "Eskimo languages have hundreds of words for snow, proving language shapes thought." The claim was inflated through retelling, English has many snow words, and the polysynthetic structure of those languages makes word counts uninformative.
• "If your language lacks a word, you cannot have the concept." Strong determinism fails: people learn new distinctions, describe unlexicalized concepts with phrases, and translate between languages.
• "All linguistic relativity findings are equally solid." Spatial frames, color category effects, and Pirahã number results are well replicated; some grammatical-gender and metaphor-framing effects have proved fragile.

Recap

Broca's and Wernicke's patients revealed that language decomposes into separable components in a left-lateralized network: nonfluent, agrammatic production with syntax-dependent comprehension failures in one case, fluent but empty speech with poor comprehension in the other. The classical box-and-arrow model has since been substantially revised, because lesions confined to Broca's area do not produce the lasting syndrome, both regions serve multiple functions, and dual-stream accounts better fit the data; what survives is decomposability and lateralization. On the relativity question, strong determinism fails and the snow-words example was always bad evidence, but weak relativity is supported by Russian blue discrimination that vanishes under verbal interference, Himba color boundary effects, absolute-frame spatial reasoning that transfers to nonverbal tasks, and Pirahã exact-number performance. The mechanism appears to be habitual attention and covert linguistic tools recruited during thinking, not permanent rewriting of perception. Next module: how that thinking goes right and wrong when you have to judge and decide.

Sources

1. National Institute on Deafness and Other Communication Disorders. (2024). Aphasia. National Institutes of Health. nidcd.nih.gov
2. Encyclopaedia Britannica. (2024). Aphasia. Encyclopaedia Britannica. britannica.com
3. Wolff, P., and Holmes, K. J. Linguistic relativity. Overview in: Stanford Encyclopedia of Philosophy, Relativism entry. Stanford University. plato.stanford.edu
4. Wikipedia. (2025). Linguistic relativity. Wikimedia Foundation. en.wikipedia.org
5. Wikipedia. (2025). Eskimo words for snow. Wikimedia Foundation. en.wikipedia.org

The big picture

Define the word "game."

Try it seriously for a moment before reading on. Whatever you propose, someone can break it. Competition? Solitaire and catch have none. Rules? So do courtrooms. Fun? Ask a professional athlete in the fourth quarter. Winners and losers? Not ring-around-the-rosy. Ludwig Wittgenstein used this example to argue that many concepts have no definition at all, only a network of overlapping similarities he called family resemblance, the way relatives share a nose here and a chin there with no single feature common to all.

This is not a puzzle about words. Concepts are the units of thought: they let you treat a never-before-seen object as a chair, predict that an unfamiliar dog might bite, and combine ideas into new ones. Everything in the previous modules depends on them. So the question of what a concept is, and how categorization works, sits near the foundation of cognitive science. This lesson traces three answers, each better than the last, and ends with the modern argument about whether concepts are abstract symbols or reenactments of experience.

The classical theory and its collapse

From Aristotle until the 1970s, the default view was the classical theory: a concept is a definition, a set of features individually necessary and jointly sufficient for membership. BACHELOR is unmarried plus adult plus male. Membership is all-or-none: something either satisfies the definition or does not, and all members are equally good members.

It fails on three counts. First, most everyday concepts resist definition, as the game exercise showed; decades of philosophical effort produced almost no successful definitions of ordinary terms. Second, category membership is graded in practice. Third, and decisively, people behave as if some members are better than others, which a definitional theory forbids.

That third point became experimental in Eleanor Rosch's work in the 1970s, and her results are the reason the classical theory is now a historical footnote. Ask people to rate how good an example each item is of a category, and they answer readily and agree with each other: a robin is a very good bird, a penguin is a poor one. Those ratings then predict a whole family of behaviors. People verify "A robin is a bird" faster than "A penguin is a bird," the typicality effect. Asked to list category members, they produce typical ones first. Children learn typical members earlier. People generalize new facts more readily from typical members than atypical ones: told robins have some enzyme, they infer other birds do; told penguins have it, they infer less. And typicality effects show up even for categories that do have crisp definitions: people rate 3 as a better odd number than 447, and 4 as a better even number than 106, although oddness is perfectly defined. That last finding matters because it shows the effect is about how concepts are mentally represented, not about vagueness in the world.

Key idea: The classical view that concepts are definitions fails: most concepts cannot be defined, membership behaves as graded, and Rosch's typicality effects predict verification speed, listing order, learning order, and inductive generalization, appearing even in well-defined categories.

Prototypes and exemplars

The prototype theory that grew from Rosch's work says a concept is a summary representation of the category's central tendency: a weighted list of characteristic features, or an abstract average member. You categorize by similarity to the prototype. This explains typicality naturally: robins share more characteristic bird features (flies, sings, small, lays eggs in trees) than penguins do.

Rosch added a second contribution that is easy to overlook and just as important. Categories form hierarchies, and one level is privileged. Above chair sits furniture, below it sits kitchen chair. The middle level, the basic level, is special: it is the level at which category members share the most features and look most alike, the level children learn first, the level with the shortest and most frequent words, and the level people default to when naming something. Point at a Labrador and almost everyone says "dog," not "mammal" and not "Labrador retriever." That is not arbitrary; it is the level that maximizes informativeness relative to effort.

The main rival is exemplar theory, developed by Douglas Medin, Robert Nosofsky, and others. It denies that you store an abstraction at all. Instead you store many individual encountered examples, and you categorize a new item by its similarity to remembered instances. This handles things prototypes struggle with: sensitivity to category variability (you know cheetah sizes vary less than dog sizes), memory for atypical instances, and the ability to use correlated features. Formal exemplar models fit categorization data extremely well. The current consensus is unglamorous but honest: people probably use both, along with rules, depending on the task, the category, and expertise. That may sound like a dodge, but it is what the data support, and the field has largely stopped treating it as a winner-take-all contest.

Key idea: Prototype theory represents categories by central tendency and explains typicality and the privileged basic level; exemplar theory stores instances and better handles variability and atypical cases. Evidence supports a mixture rather than a single mechanism.

What similarity cannot do

Now the deeper problem, and it is the most philosophically interesting part of this lesson. Both prototype and exemplar theories are similarity-based, and similarity is treacherous. A zebra is similar to a horse. It is also similar to a barcode, in a different respect. Any two things share indefinitely many properties if you are allowed to count arbitrary ones. So similarity cannot ground categorization unless something first decides which features count, and that something is knowledge.

Consider Medin and Andrew Ortony's argument via psychological essentialism: people act as though category members share a hidden essence that makes them what they are, even when they cannot say what it is. Frank Keil's transformation studies with children showed this vividly. Tell children that scientists took a raccoon, dyed it black with a white stripe, and implanted a sac of smelly fluid, then ask whether it is now a skunk. Older children say no: it is still a raccoon, because of what is inside. But do the same operation on a coffeepot to make a bird feeder and they accept the change, because artifacts are defined by intended function rather than inner nature. Children treat natural kinds and artifacts differently, using different theories of what makes something what it is.

This motivates the theory-theory: concepts are embedded in intuitive theories of the world, and categorization depends on causal beliefs, not raw feature overlap. The evidence is strong. Gregory Murphy and Medin showed that a man jumping into a swimming pool fully clothed is quickly categorized as drunk at a party, despite matching no prototype, because a causal story explains it. Categories that are causally coherent are learned faster than ones with the same feature statistics but no explanatory glue. Expertise reorganizes categories along theoretical rather than superficial lines: novices sort physics problems by surface features like inclined planes, experts sort by the principle involved.

Key idea: Similarity alone cannot explain categorization because any two things share indefinitely many features. Psychological essentialism and the theory-theory hold that causal knowledge determines which features count, supported by Keil's transformation studies and expert-novice differences.

Are concepts grounded in the body?

The final debate returns us to Module 2's fight about representation. Traditional cognitive science treated concepts as amodal symbols: abstract representations divorced from the sensory and motor systems that produced them, like the word DOG in a database. Embodied or grounded cognition, argued most systematically by Lawrence Barsalou, claims instead that concepts are partial reenactments of perception and action: understanding "kick" involves simulating kicking, and understanding "apple" involves reactivating visual, gustatory, and motor traces.

The supporting evidence is genuinely interesting. Action verbs activate motor and premotor cortex somatotopically: reading "lick," "pick," and "kick" produces activity near the mouth, hand, and foot regions respectively, in work by Friedemann Pulvermuller and others. The action-sentence compatibility effect, from Arthur Glenberg and Michael Kaschak, found that people are faster to respond with a movement away from the body after reading a sentence describing motion away from the body. Comprehending sentences about grasping interferes with simultaneous grasping. Perceptual features are activated during property verification even when unnecessary.

Now the honest caveats, and they matter. First, activation does not establish necessity: motor activity during verb reading might accompany comprehension rather than constitute it. Second, patients with severe motor impairments, including some with motor neuron disease, generally still understand action language, which is hard for strong embodiment. Third, and most seriously for the field's credibility, the action-sentence compatibility effect was the subject of a large multi-laboratory replication attempt reported in 2021 that failed to find the original effect, and several related embodiment findings have had replication trouble. Fourth, abstract concepts such as justice, seven, and inflation are hard to ground in sensorimotor simulation, and proposals to do so via metaphor have their own evidential problems.

The defensible position in 2026: sensory and motor systems participate in conceptual processing more than the old amodal view allowed, and that participation is real and measurable, but strong claims that concepts simply are simulations outrun the evidence, and some flagship results have not replicated. Hybrid accounts, in which concepts have both grounded and abstract components, currently fit the data best.

Common misconceptions

• "Concepts are definitions we learned." Most everyday concepts have no definition, and typicality effects appear even for categories that do, so definitions are not how concepts are represented.
• "Typicality effects just mean some things are rarer." Frequency contributes, but typicality predicts induction, learning order, and verification speed independently, and appears for well-defined categories like odd numbers.
• "Prototype theory won." Exemplar models fit many data sets better, rule use is real, and current evidence supports multiple mechanisms recruited by different tasks.
• "Categorization is similarity computation." Similarity is unconstrained without knowledge about which features matter. Causal theories and essentialist intuitions do much of the work.
• "Neuroimaging proved that understanding words requires simulating actions." Motor activation during action-word reading is well documented, but necessity is not established, patients with motor deficits still comprehend, and key behavioral effects have failed large replication attempts.

Recap

Concepts are the units of thought, and the classical definitional theory of them collapsed under Wittgenstein's family resemblance argument and Rosch's typicality data, which show graded membership predicting speed, listing order, learning, and induction even in well-defined categories. Prototype theory represents categories by central tendency and identifies a privileged basic level; exemplar theory stores instances and handles variability better; the evidence supports a mixture. But similarity alone cannot ground categorization, because any two things share indefinitely many features, so knowledge must select the relevant ones: psychological essentialism and theory-based accounts, supported by Keil's raccoon-to-skunk studies and expert-novice sorting differences, put causal beliefs at the center. Finally, embodied accounts claim concepts are sensorimotor simulations, supported by somatotopic activation for action verbs, but weakened by necessity concerns, preserved comprehension in motor-impaired patients, difficulties with abstract concepts, and a failed multi-laboratory replication of a flagship effect. Hybrid views fit best, which is where this module leaves you: better calibrated, not more certain.

Sources

1. Margolis, E., and Laurence, S. (2023). Concepts. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
2. Shapiro, L., and Spaulding, S. (2024). Embodied cognition. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. Encyclopaedia Britannica. (2024). Concept formation. Encyclopaedia Britannica. britannica.com
4. Wikipedia. (2025). Prototype theory. Wikimedia Foundation. en.wikipedia.org
5. Wikipedia. (2025). Psychological essentialism. Wikimedia Foundation. en.wikipedia.org

The big picture

Here is a description written by a psychologist. Linda is 31, single, outspoken, and very bright. She majored in philosophy. As a student she was deeply concerned with discrimination and social justice, and participated in antinuclear demonstrations. Which is more probable?

(a) Linda is a bank teller. (b) Linda is a bank teller and is active in the feminist movement.

Most people, including most statistically trained people, choose (b). And (b) cannot be more probable than (a), ever, under any circumstances, because every feminist bank teller is a bank teller. The set described in (b) is a subset of the set described in (a). This is the conjunction fallacy, and Amos Tversky and Daniel Kahneman reported rates above 80 percent in some samples, including among doctoral students in decision science.

That single demonstration launched a research program that reshaped psychology, economics, medicine, law, and public policy, and earned Kahneman the 2002 Nobel Memorial Prize in Economic Sciences (Tversky had died in 1996). This lesson works through the classic findings carefully, because they are frequently misdescribed, then turns to the two arguments that keep the field honest: what dual-process theory actually claims, and why a serious camp of researchers thinks the whole "biases" framing is misleading.

Representativeness

Why does Linda seem like a feminist bank teller? Because you are not computing probability at all. You are computing similarity: how well does Linda match your stereotype of each category? Tversky and Kahneman called this the representativeness heuristic: judging probability by resemblance to a prototype. It is fast, often reasonable, and systematically wrong in specific ways.

The most consequential error it produces is base rate neglect. Consider their taxicab problem. A city has two cab companies: 85 percent of cabs are Green, 15 percent are Blue. A cab is involved in a hit-and-run at night. A witness identifies it as Blue, and testing shows the witness correctly identifies colors 80 percent of the time. What is the probability the cab was Blue? Most people answer around 80 percent. The correct answer is about 41 percent. Work it through with 100 accidents: 15 involve Blue cabs, and the witness correctly calls 12 of them Blue; 85 involve Green cabs, and the witness wrongly calls 20 percent, or 17 of them, Blue. So 29 cabs get called Blue and only 12 actually are, giving 12 divided by 29, roughly 41 percent. The witness is reliable; Blue cabs are rare; rarity matters and people ignore it.

This is not a parlor trick. The medical version is the reason it appears on this syllabus. A screening test for a disease affecting 1 in 1,000 people has a 5 percent false positive rate and catches nearly all true cases. A patient tests positive. What is the chance they have the disease? A classic study by Ward Casscells and colleagues found that most Harvard medical staff answered around 95 percent. The answer is about 2 percent: in 1,000 people, 1 true case and about 50 false positives, so 1 out of roughly 51. Gerd Gigerenzer later showed that presenting the same information as natural frequencies, exactly as I just did, dramatically improves performance in doctors and patients alike, which is now a practical recommendation in medical communication.

Representativeness also produces the gambler's fallacy (expecting a short run of coin flips to "correct" itself, because HTHTTH looks more representative of randomness than HHHHHH, though both have identical probability) and insensitivity to sample size (a small hospital has more days with over 60 percent boys born than a large one, but most people say the rates are equal).

Key idea: Representativeness substitutes similarity for probability, producing the conjunction fallacy, base rate neglect, and the gambler's fallacy. Natural frequency formats substantially reduce these errors, which shows the problem is partly in the presentation, not only in the head.

Availability and anchoring

The availability heuristic judges frequency by how easily examples come to mind. Ask people whether more English words start with K or have K in third position; most say the first, because retrieving words by initial letter is easy, while in fact more words have K third. Ask people to estimate causes of death and they overestimate the dramatic and newsworthy (tornadoes, homicide, plane crashes) and underestimate the quiet and common (stroke, diabetes, asthma). Paul Slovic's risk perception work traced this directly to media coverage: what gets reported gets recalled, and what gets recalled feels frequent. This is why the public fear of shark attacks vastly exceeds the fear of swimming pools, and why fear of flying survives statistics showing driving to the airport is the dangerous part.

Anchoring is stranger and more unsettling. Tversky and Kahneman spun a wheel of fortune rigged to stop at 10 or 65, asked participants whether the percentage of African nations in the UN was higher or lower than that number, then asked for their estimate. Median estimates were 25 after the low anchor and 45 after the high one. A number everyone watched being generated at random moved their judgment. Anchoring has been replicated in real estate appraisals by professionals, courtroom sentencing recommendations, and negotiation outcomes, and it is notably resistant to warnings and incentives.

Framing and prospect theory

Now the finding that changed economics. Standard theory says preferences should be invariant: how a choice is described should not change what you choose. Tversky and Kahneman's Asian disease problem tested this. A disease is expected to kill 600 people. Program A saves 200 for certain; Program B has a one-third chance of saving 600 and a two-thirds chance of saving none. Most people choose A. Now the same numbers restated: Program C means 400 people die for certain; Program D has a one-third chance nobody dies and a two-thirds chance 600 die. Most people choose D. A and C are identical. B and D are identical. Simply describing outcomes as lives saved or lives lost flips the majority preference.

Prospect theory explains this with three ideas. First, people evaluate outcomes as gains and losses relative to a reference point, not as final states of wealth. Second, the value function is concave for gains and convex for losses, so people are risk-averse when facing gains and risk-seeking when facing losses, which is exactly the flip you just saw. Third, loss aversion: losses loom larger than equivalent gains, by a factor typically estimated around two. That single asymmetry explains the endowment effect (people demand more to give up a mug than they would pay to buy it), status quo bias, and the reluctance to sell a losing stock.

Loss aversion deserves a caveat, because this course does calibration rather than cheerleading. The exact multiplier varies by domain, some researchers argue the endowment effect has been overinterpreted, and there is an active debate about whether loss aversion is as universal as the textbook version suggests. The framing effect itself, however, is among the most robustly replicated findings in the literature.

Key idea: Framing effects violate the invariance that rational choice theory requires. Prospect theory explains them with reference dependence, risk aversion for gains, risk seeking for losses, and loss aversion, though the size and universality of loss aversion remain debated.

Dual-process theory, stated carefully

The organizing framework popularized in Kahneman's 2011 book divides thinking into System 1, fast, automatic, effortless, associative, and always running, and System 2, slow, deliberate, effortful, and capacity-limited. On this account heuristics are System 1's outputs, and errors occur when System 2 fails to monitor and override them.

The classic evidence is the Cognitive Reflection Test, developed by Shane Frederick. A bat and a ball cost 1.10 dollars in total. The bat costs 1 dollar more than the ball. How much does the ball cost? The number 10 cents arrives unbidden. It is wrong: if the ball were 10 cents, the bat at 1.10 would make 1.20 total. The answer is 5 cents. Most people at elite universities miss at least one of the three items on the test, and performance predicts susceptibility to other biases.

Now the criticisms, which any honest treatment must include. Keith Stanovich, Jonathan Evans, and others who built the theory have themselves warned against the popular version. There are probably not two systems but many processes, differing along several dimensions that do not always line up: automaticity, awareness, effort, and evolutionary age are separable. The labels invite circular explanation: calling an error "System 1" after observing it explains nothing unless the theory predicts in advance which process runs. Intuition is frequently right, and expert intuition in well-structured environments, chess, firefighting, is often better than deliberation, as Gary Klein's naturalistic decision research shows. Kahneman and Klein wrote a notable joint paper trying to specify when intuition can be trusted: when the environment is regular enough to contain learnable patterns and the person has had prolonged practice with rapid, clear feedback. Finally, several dual-process predictions have had mixed replication results.

Key idea: Dual-process theory frames heuristics as fast automatic processing that deliberation sometimes fails to override, with the Cognitive Reflection Test as a workhorse measure. Its own architects caution that "two systems" is a simplification, that the labels can become circular, and that intuition is reliable in learnable, high-feedback environments.

The rationality debate

The deepest challenge comes from a different direction. Gigerenzer and colleagues at the Max Planck Institute argue that the heuristics-and-biases program measures human judgment against the wrong standard. Their objections are worth understanding because they are substantive, not defensive.

First, they argue many "errors" reflect reasonable interpretations of ambiguous questions rather than faulty reasoning. In the Linda problem, conversational norms make "bank teller" naturally read as "bank teller and not active in the feminist movement," since a cooperative speaker would not offer the more specific option otherwise. When the problem is posed in frequency terms ("out of 100 people like Linda, how many are bank tellers? how many are feminist bank tellers?"), conjunction errors drop sharply. Second, they argue that in the real world, where information is scarce and time is short, simple heuristics can outperform complex optimization. Their fast and frugal research shows that the recognition heuristic (if you recognize one option and not the other, pick the recognized one) predicts things like tennis match outcomes surprisingly well, and that take-the-best, which decides on the single best cue, can beat multiple regression on out-of-sample prediction. This is ecological rationality: a strategy is not rational or irrational in the abstract, only well or badly matched to an environment.

Herbert Simon had already supplied the frame with bounded rationality and satisficing: real agents with limited time, memory, and computation do not optimize, they search until they find an option that is good enough. From that angle, heuristics are not sad approximations to a perfect reasoner; they are the design solution for a mind that has to act.

Where does this leave you? With a defensible synthesis rather than a team to join. The empirical demonstrations are real and replicable: framing effects, anchoring, base rate neglect, and the conjunction fallacy happen, in the laboratory and in professionals doing their jobs. The interpretation is genuinely contested: whether these are defects of the mind or artifacts of unnatural formats and mismatched norms depends on the case, and the format effects (natural frequencies, frequency phrasing) show the answer is often "some of both." The practical upshot survives either way: how you present information changes what people conclude, so if you design a form, write a medical brochure, or make a decision that matters, structure the presentation deliberately.

Key idea: Gigerenzer's critique argues that many biases reflect ambiguous framing and mismatched norms, and that fast and frugal heuristics can outperform optimization in real environments. The demonstrations are robust; their interpretation is contested; the practical lesson about presentation holds regardless.

Common misconceptions

• "Heuristics are just bad thinking." They are fast strategies that work well in many environments and fail in specific, predictable ones. Ecological rationality reframes the question as fit between strategy and environment.
• "Only untrained people make these errors." Physicians misread test results, judges anchor on irrelevant numbers, and statistically sophisticated participants commit the conjunction fallacy. Training helps unevenly.
• "System 1 and System 2 are two places in the brain." They are labels for clusters of properties, not anatomical structures, and the theory's own developers warn against treating them as literal homunculi or as a complete taxonomy.
• "Knowing about a bias protects you from it." Anchoring persists under warnings and incentives, and awareness does little for framing. Structural fixes such as natural frequency formats and checklists outperform willpower.
• "Prospect theory says people are irrational." It says people evaluate changes relative to a reference point with asymmetric sensitivity to losses. That is a descriptive model of preferences, not a verdict of stupidity.

Recap

The Linda problem exposes the conjunction fallacy, and behind it the representativeness heuristic, which substitutes similarity for probability and produces base rate neglect in taxicab problems and in real medical test interpretation. Availability judges frequency by ease of recall, distorting risk perception in step with media coverage; anchoring lets an obviously random number shift numerical estimates even among professionals. Framing effects in the Asian disease problem violate the invariance rational choice requires, and prospect theory explains them with reference dependence, risk seeking in the loss domain, and loss aversion. Dual-process theory organizes these as fast automatic processing insufficiently monitored by deliberation, with the Cognitive Reflection Test as its measure, while its own architects caution against the popular two-systems picture. Gigerenzer's ecological rationality program argues that many biases reflect ambiguous phrasing and mismatched norms and that simple heuristics can beat optimization in real environments, with natural frequency formats as the strongest practical evidence that presentation drives performance. Next: how the same mind reads other minds.

Sources

1. Tversky, A., and Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124-1131. pubmed.ncbi.nlm.nih.gov
2. Nobel Prize Outreach. (2002). Daniel Kahneman: Prize in Economic Sciences. nobelprize.org
3. Encyclopaedia Britannica. (2024). Prospect theory. Encyclopaedia Britannica. britannica.com
4. Wheeler, G. (2020). Bounded rationality. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
5. Wikipedia. (2025). Conjunction fallacy. Wikimedia Foundation. en.wikipedia.org

The big picture

Sally puts a marble in her basket and leaves the room. While she is gone, Anne moves the marble to a box. Sally comes back. Where will Sally look for her marble?

You answered "the basket" without effort, and you did something extraordinary to get there. You represented another person's representation of the world, held it separate from your own knowledge of where the marble actually is, and predicted behavior from the false version. Most three-year-olds say "the box." Most five-year-olds say "the basket." Something changes in between, and understanding what changes is one of developmental cognitive science's central projects.

This lesson has two intertwined threads: how children come to understand minds, and how the whole field of cognitive development shifted from Piaget's picture of a child slowly constructing reality to a picture of an infant already equipped with structured expectations. Both threads carry the same methodological lesson, which is really about evidence: what looks like a missing concept is often a missing ability to demonstrate the concept under the demands of a particular task.

Theory of mind and the false belief task

Theory of mind is the capacity to attribute mental states, beliefs, desires, intentions, knowledge, to yourself and others, and to use those attributions to explain and predict behavior. The term entered the literature in a 1978 paper by David Premack and Guy Woodruff asking whether chimpanzees have one. The philosopher Daniel Dennett pointed out in a commentary that the decisive test would involve a false belief, because only then do predictions based on the person's mental state diverge from predictions based on reality. Heinz Wimmer and Josef Perner built the task in 1983, and Simon Baron-Cohen, Alan Leslie, and Uta Frith created the Sally-Anne version in 1985.

The developmental pattern is one of the most replicated findings in the field: a large meta-analysis by Wellman, Cross, and Watson covering hundreds of conditions found children shifting from below-chance to above-chance performance around age 4, with the same trajectory across many countries and languages, though timing shifts somewhat with factors like number of siblings and family talk about mental states. Deaf children of hearing parents with delayed access to fluent language show delayed false belief performance, which suggests conversational access to talk about minds matters.

The 1985 Baron-Cohen study also compared children with autism to children with Down syndrome matched on mental age. The autistic group performed markedly worse on the false belief task specifically, which generated the influential mindblindness hypothesis. That literature deserves careful handling: subsequent work found many autistic individuals pass false belief tasks, especially verbally able ones and older participants; the deficit is better described as differences in social cognitive processing than an absence of theory of mind; and the double empathy problem raised by Damian Milton points out that mutual misunderstanding between autistic and non-autistic people runs in both directions, with non-autistic people also poor at reading autistic mental states. The original finding is real; the popular gloss overreached.

Key idea: False belief tasks isolate the ability to represent someone's mistaken view of the world, and explicit performance shifts robustly around age 4 across cultures. Autistic children as a group show later or lower performance on these tasks, but the mindblindness framing has been substantially qualified.

What changes at four, and does it?

Here is where the field got interesting. If theory of mind arrives at four, what about the toddler who hides from a parent, or the two-year-old who points to show you something? Those behaviors seem mentalistic. The suspicion that the standard task underestimates children was tested with non-verbal methods.

Kristine Onishi and Renee Baillargeon reported in 2005 that 15-month-olds looked longer when an actor reached in the location that contradicted the actor's false belief, suggesting implicit expectations about belief-driven action long before explicit passing. Anticipatory eye-movement studies found 2-year-olds looking to the correct location. This produced a two-systems picture: an early, efficient, implicit system tracking others' perspectives, and a later, flexible, explicit system that supports verbal reporting and depends on language and executive function.

Then came a genuine scientific complication that this course will not hide. Several large multi-laboratory replication efforts, including a 2019 project reported by Kulke and colleagues and further attempts through the early 2020s, failed to reproduce key implicit false belief findings reliably. Some effects held, many did not, and the field is currently divided about whether infant mentalizing is real, fragile, or an artifact of specific designs. The honest statement for a student in 2026 is this: explicit false belief development around age four is rock solid; implicit infant mentalizing is contested and under active reexamination. Being able to say which parts of a literature are secure and which are shaking is exactly the competence this course is for.

Piaget's picture

Jean Piaget, working in Geneva from the 1920s onward, built the first systematic theory of cognitive development and got the big framing right: children are not miniature adults with less information, they think in qualitatively different ways, and they build knowledge actively through interaction with the world. His mechanisms were assimilation (fitting new experience into existing schemes) and accommodation (revising schemes when they fail).

His four stages remain a common vocabulary. In the sensorimotor stage (birth to about 2), infants know the world through action and perception, and, Piaget argued, initially lack object permanence: before roughly 8 months, a hidden object is treated as gone, and even later infants make the A-not-B error, searching where an object was previously found rather than where they just saw it hidden. In the preoperational stage (about 2 to 7), children use symbols and language but fail conservation tasks, judging that pouring water into a taller glass increases the amount, and show egocentrism, failing his three mountains task in which they must report what a doll sees from another side. In concrete operational (about 7 to 11), logical operations arrive but remain tied to concrete situations, and in formal operational (11 and up), abstract and hypothetical reasoning becomes possible.

Piaget's observations were largely accurate; his interpretations were often not. Two systematic problems emerged. First, his tasks confounded competence with performance: they demanded language, memory, motor coordination, and inhibitory control, so failure could reflect any of those rather than a missing concept. Second, development is less stage-like and more domain-specific and gradual than the theory implies; children pass conservation for number before volume, and a child can look concrete-operational in one domain and preoperational in another. Simplify the demands and children look far more competent. Reduce the three mountains task to a naughty-teddy-style hiding game, as Martin Hughes did, and children as young as three take another's visual perspective successfully.

Key idea: Piaget correctly established that children think qualitatively differently and build knowledge actively, and his stage vocabulary endures. But his tasks confounded conceptual competence with task demands, and development is more gradual and domain-specific than stages imply.

Core knowledge and the looking-time revolution

The methodological breakthrough that reopened infancy was simple: stop asking babies to do things and start measuring what surprises them. In the violation of expectation paradigm, infants are habituated to an event until looking time drops, then shown either a possible or an impossible outcome. Longer looking at the impossible outcome is taken as evidence that the infant expected otherwise.

The results transformed the field. Baillargeon's drawbridge studies suggested infants around 4 to 5 months expect a solid object to block a rotating screen, evidence of object permanence far earlier than Piaget claimed. Elizabeth Spelke's work supported a set of core knowledge systems: for objects (cohesion, continuity, contact), for approximate number (5-month-olds discriminate large sets by ratio), for agents (goal-directedness), for space, and for social partners. Karen Wynn's 1992 addition and subtraction study found 5-month-olds looking longer when one doll plus one doll appeared to yield one doll. In the moral domain, Kiley Hamlin, Wynn, and Paul Bloom reported that infants prefer a puppet that helps a climber over one that hinders it.

Now the caveats, and they are not minor. Looking time is a blunt measure: longer looking might mean surprise, but it might mean familiarity preference, perceptual novelty, or a low-level property of the display, and interpreting it as conceptual expectation is an inference, not an observation. Several classic infant findings, including some helper-hinderer results, have had mixed replication outcomes, and the ManyBabies consortium was formed specifically to test key infancy effects across many laboratories with large samples. Its first project confirmed the infant preference for infant-directed speech while finding a smaller effect than the literature suggested, which is roughly the pattern to expect. Core knowledge is a serious and well-supported research program; individual headline studies within it should be held more loosely than textbook summaries suggest.

Key idea: Violation-of-expectation methods revealed structured infant expectations about objects, number, and agents, supporting core knowledge systems rather than a blank slate. Looking time remains an indirect measure, and multi-laboratory replication work is currently recalibrating effect sizes.

The social route and what it explains

Lev Vygotsky, Piaget's contemporary, emphasized what Piaget underplayed: development is social. His zone of proximal development is the gap between what a child can do alone and what they can do with support, and instruction targeted there, with scaffolding that is gradually withdrawn, is more effective than either drilling the mastered or demanding the impossible. He also argued that private speech, the running commentary of preschoolers, becomes internalized as verbal thought, which connects to Module 4's finding that language provides tools recruited in reasoning.

Michael Tomasello's comparative work sharpens the point about what makes human development unusual. On many physical cognition tasks, chimpanzees and 2.5-year-old children perform comparably. On social learning tasks, the children pull far ahead. Children over-imitate, copying even causally irrelevant steps in a demonstration, which looks foolish until you notice it is how cumulative culture is possible: faithful copying preserves practices whose rationale the learner cannot yet evaluate. Human cognition is not just individually powerful; it is designed to inherit.

Common misconceptions

• "Children under four have no understanding of other minds." They track goals, gaze, and desires well before four. What arrives around four is reliable explicit reasoning about false beliefs under verbal task demands.
• "Autistic people lack theory of mind." The original studies found group differences on specific tasks, but many autistic people pass them, the framing has been heavily qualified, and misunderstanding is mutual rather than one-directional.
• "Piaget was simply wrong." His central insight, that children think qualitatively differently and construct knowledge actively, was right and reshaped the field. His timelines were late because his tasks were demanding.
• "Looking-time studies show babies know physics." They show infants look longer at certain outcomes, which supports structured expectations under an inference. The measure is indirect and several headline findings are being re-tested at scale.
• "Development is a fixed ladder of stages." Progress is domain-specific and gradual: a child can conserve number before volume and reason abstractly in a familiar domain while failing in an unfamiliar one.

Recap

Theory of mind is the ability to attribute mental states and predict behavior from them, and the false belief task isolates it by making reality and belief diverge. Explicit performance shifts robustly around age four across cultures, with language exposure and family talk influencing timing; group differences in autism are real but the mindblindness framing has been substantially qualified. Claims of implicit mentalizing in infancy are currently contested after mixed multi-laboratory replications. Piaget established that children think qualitatively differently and build knowledge actively, but his tasks confounded concepts with performance demands, and simplified versions reveal earlier competence. Violation-of-expectation methods support core knowledge systems for objects, number, and agents, with the caveat that looking time is indirect and effect sizes are being recalibrated by consortium replication. Vygotsky and Tomasello add the social route: scaffolded learning in the zone of proximal development and high-fidelity social learning that makes cumulative culture possible. Next module: whether anything we have built out of silicon shares any of this.

Sources

1. Baron-Cohen, S., Leslie, A. M., and Frith, U. (1985). Does the autistic child have a theory of mind? Cognition, 21(1), 37-46. pubmed.ncbi.nlm.nih.gov
2. Encyclopaedia Britannica. (2024). Jean Piaget. Encyclopaedia Britannica. britannica.com
3. Eunice Kennedy Shriver National Institute of Child Health and Human Development. (2023). Child development. National Institutes of Health. nichd.nih.gov
4. Wikipedia. (2025). Theory of mind. Wikimedia Foundation. en.wikipedia.org
5. Wikipedia. (2025). Core knowledge. Wikimedia Foundation. en.wikipedia.org

The big picture

Artificial intelligence sits at a corner of the hexagon for a specific reason: building a working model is a test of whether you understood the process. If your theory of how people recognize faces cannot be turned into something that recognizes faces, the theory is probably vague. AI is cognitive science's engineering wing, and its failures have taught the field as much as its successes.

This lesson tells that history in three acts, and the through-line is a single question that keeps changing form: what does building an intelligent-seeming machine tell you about human minds? The answer in 1960 was "everything, in principle." The answer today is more careful and more interesting. Along the way, expect this lesson to be less impressed than a magazine article and more impressed than a skeptic, because both extremes get the evidence wrong.

Act one: good old-fashioned AI

The 1956 Dartmouth summer workshop proposed that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." The founding approach, later nicknamed GOFAI (good old-fashioned AI) by the philosopher John Haugeland, followed directly from the symbolic view you met in Lesson 5: represent knowledge as symbols and logical rules, then search through possible states to find solutions.

It produced real results. Newell and Simon's Logic Theorist proved theorems in 1956, and their General Problem Solver formalized means-ends analysis: compare the current state to the goal, find an operator that reduces the difference, apply it, recurse. This was a psychological theory as much as a program, and they tested it against human think-aloud protocols. Expert systems followed in the 1970s and 1980s: MYCIN diagnosed blood infections at levels comparable to specialists, and DENDRAL inferred molecular structures. Search algorithms from this era still run inside your route planner.

Then it stalled, twice, in periods called AI winters when funding and confidence collapsed. Understanding why is more valuable than memorizing the timeline. First, the frame problem: representing what changes when an action occurs is manageable, but representing everything that does not change is not. Move a cup and the table stays put, the room stays the same color, the year does not change. Human common sense handles this invisibly; explicit symbolic systems drowned in it. Second, the knowledge acquisition bottleneck: expert systems needed rules hand-extracted from human experts, and experts largely cannot articulate what they know. Third, brittleness: rule systems performed well inside their domain and failed absurdly one step outside it, with no graceful degradation. Fourth, perception and motor control, the things every toddler does, proved hardest of all. Moravec's paradox names this inversion: high-level reasoning requires relatively little computation, while sensorimotor skills require enormous amounts, because evolution spent hundreds of millions of years on the latter and a few hundred thousand on the former.

Key idea: Symbolic AI implemented the classical theory of mind and succeeded in formal, bounded domains. It stalled on the frame problem, the difficulty of extracting expert knowledge, brittleness, and the inversion Moravec noted: what feels easy to humans is computationally hardest.

Act two: learning from data

The alternative was already sketched in Lesson 5. Instead of programming knowledge, train a network on examples. Frank Rosenblatt's perceptron in 1958 learned simple classifications, but Marvin Minsky and Seymour Papert's 1969 book showed single-layer perceptrons cannot compute even the exclusive-or function, and interest collapsed. The revival came with backpropagation, popularized for cognitive modeling in 1986 by Rumelhart, Geoffrey Hinton, and Ronald Williams. The idea is conceptually simple: run an input forward through layers of weighted units, compare the output to the desired answer, then propagate the error backward, nudging each weight in the direction that reduces it. Repeat millions of times.

What turned this from a promising idea into a transformation was not a conceptual breakthrough but scale on three fronts: enormous labeled datasets, graphics processors that perform the required matrix arithmetic in parallel, and architectural refinements. In 2012, AlexNet cut the error rate on the ImageNet object recognition benchmark dramatically, and computer vision changed within a few years. Convolutional networks, inspired loosely by the visual hierarchy Hubel and Wiesel described, now exceed careful human performance on some narrow recognition tasks. In 2016 AlphaGo defeated Lee Sedol at Go, a game whose search space had made it a symbol of AI's limits, using learned evaluation rather than exhaustive search. AlphaFold's protein structure predictions changed a scientific field.

Act three: large language models

The transformer architecture, introduced in 2017, made it practical to train very large models on very large text corpora. A large language model is trained on a deceptively simple objective: given a sequence of text, predict what comes next. Do that over an enormous corpus, with billions of parameters, and the system must implicitly capture syntax, facts, discourse structure, and a great deal about how humans write. Additional training stages then shape the model's behavior toward being helpful and following instructions.

What should a cognitive scientist make of this? Start with what is not in dispute. These systems produce fluent, largely grammatical, often useful text across domains, they perform well on many benchmarks including some professional examinations, and their capabilities were not explicitly programmed. For linguistics, this is a genuinely significant datum: it shows that a great deal of linguistic structure is learnable from distributional statistics at scale, which bears directly on the poverty of the stimulus debate from Lesson 9. Several researchers have argued this weakens strong nativist arguments.

Now the honest qualifications, which the field takes seriously.

The data disanalogy is severe. Leading models are trained on quantities of text a human could not read in thousands of lifetimes, while a child acquires language from something on the order of tens of millions of words plus rich embodied social context. A system that needs orders of magnitude more data to reach comparable linguistic competence is not obviously modeling the child's learning problem, whatever it shows about learnability in principle. Work on models trained with developmentally plausible data budgets exists precisely because this comparison matters.

Second, benchmark performance is a computational-level observation. Recall Marr from Lesson 2: matching human output does not establish a matching algorithm. A model may reach the same answer by a route with no human counterpart, and interpretability research consistently finds internal mechanisms that are not obviously humanlike.

Third, characteristic failure patterns matter as evidence. These systems can produce confident false statements, sometimes called hallucinations or confabulations, because fluency and accuracy are separate properties of the training objective. They can be sensitive to prompt phrasing in ways humans are not. Evaluation is complicated by data contamination: when test items may appear in training data, benchmark scores overstate generalization, which is why researchers increasingly build held-out or freshly constructed tests.

Fourth, the interpretive question is genuinely open, and reasonable researchers disagree. Some argue that next-token prediction at scale requires building internal world models, citing probing studies that recover structured representations from model internals. Others argue the systems are sophisticated pattern completers whose competence is real but different in kind. The dispute has been running since Lesson 5's systematicity argument, and it has not been settled by either camp declaring victory. What you should refuse to accept from any direction is confident metaphysics stated as though it were a measurement.

Key idea: Large language models learn from next-token prediction at enormous scale, demonstrating that much linguistic structure is statistically learnable. Their data requirements, opaque internal algorithms, characteristic confabulations, and contaminated benchmarks all limit what they establish about human cognition.

What AI has actually taught cognitive science

Step back from the arguing and count the contributions, because they are substantial and often overlooked.

Lesson from AI	What it taught about minds
Moravec's paradox	Perception and motor control are computationally harder than logic, so subjective effort is a terrible guide to complexity
The frame problem	Common sense is an enormous, mostly invisible achievement, not a simple background capacity
Backpropagation networks	Rule-like behavior can emerge from statistical learning, as the past-tense model showed
Reinforcement learning	Temporal difference learning signals resemble dopamine neuron firing patterns, linking a learning algorithm to neural data
Deep network vision	Hierarchical feature learning produces representations that predict activity in primate visual cortex better than earlier hand-built models
Language models	A great deal of grammar and world knowledge is extractable from distributional statistics alone

Two of those rows deserve emphasis because they are cases where AI directly advanced neuroscience. The reinforcement learning connection, developed by Peter Dayan, Read Montague, and Wolfram Schultz among others, found that dopamine neurons fire in a pattern closely matching the reward prediction error term in temporal difference learning algorithms. And deep network models of vision, evaluated by Daniel Yamins, James DiCarlo, and colleagues, turned out to predict neural responses in monkey inferotemporal cortex better than the models neuroscientists had built by hand. Those are genuine cases of engineering feeding back into biology.

Key idea: AI's contributions to cognitive science include reframing what is computationally hard, demonstrating emergent rule-like behavior, and supplying algorithms, notably reinforcement learning and deep vision models, that predict neural data better than their predecessors.

How to think about this responsibly

You will spend the rest of your life reading claims about machine intelligence, so here is a checklist drawn from everything above. Ask what the system was trained on, since benchmark performance plus possible contamination is weak evidence of generalization. Ask which Marr level the claim is about, since equivalent output does not imply equivalent algorithm. Ask about the failure cases, because failures are more diagnostic of mechanism than successes. Ask about data efficiency, because a system needing a million times more experience than a child is solving a different problem. And ask who is making the claim and what they are selling, which is not cynicism but ordinary source evaluation.

Next lesson takes the question you have been circling and confronts it directly: could a machine understand, or be conscious, at all?

Common misconceptions

• "Neural networks work like brains." They are loosely inspired by neurons and share distributed representation and learned weights, but differ in learning rules, timing, architecture, and energy budget. Similarity is an analogy, not an identity.
• "Passing professional exams means a system reasons like a professional." That is a computational-level observation, and possible training contamination weakens it further. Failure patterns are the more informative evidence.
• "Symbolic AI was a dead end." It produced planning, search, formal verification, and knowledge representation techniques still in use, and hybrid neuro-symbolic approaches are an active research area.
• "Language models prove language needs no innate structure." They show much structure is statistically learnable, which is a real contribution to the debate, but their data budgets are astronomically larger than a child's, so the learning problems differ.
• "AI progress is smooth and inevitable." The field has had two winters and repeated overconfident predictions, including that machine translation and computer vision were nearly solved in the 1960s.

Recap

Symbolic AI implemented the classical theory of mind, succeeded in bounded formal domains with theorem provers and expert systems, and stalled on the frame problem, the knowledge acquisition bottleneck, brittleness, and Moravec's paradox. Connectionist learning revived with backpropagation and then scaled with large datasets and parallel hardware, producing the vision, game-playing, and protein-structure results of the 2010s. Transformer-based language models trained on next-token prediction demonstrate that substantial linguistic structure and world knowledge are learnable from distributional statistics, which bears on the poverty of the stimulus debate, while their data requirements, opaque internals, confabulations, and contaminated benchmarks limit inferences about human cognition. AI's real gifts to the field include reframing what is computationally hard, showing rule-like behavior emerging without rules, and supplying models, reinforcement learning and deep vision networks, that predict neural data better than their hand-built predecessors. The checklist for any claim is: what data, which Marr level, what failures, how efficient, and who benefits from the claim.

Sources

1. Encyclopaedia Britannica. (2025). Artificial intelligence. Encyclopaedia Britannica. britannica.com
2. Bringsjord, S., and Govindarajulu, N. S. (2024). Artificial intelligence. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. MIT OpenCourseWare. (2010). 6.034 Artificial Intelligence. Massachusetts Institute of Technology. ocw.mit.edu
4. Wikipedia. (2025). History of artificial intelligence. Wikimedia Foundation. en.wikipedia.org
5. Wikipedia. (2025). Moravec's paradox. Wikimedia Foundation. en.wikipedia.org

The big picture

Two questions have shadowed cognitive science since its founding, and this lesson faces them directly. Could a machine understand? And what would it take for anything, machine or animal or person, to be conscious?

These are philosophical questions, which some students take to mean unanswerable or merely verbal. That is a mistake. They are questions about what our concepts commit us to, and getting them wrong has practical consequences: for how we evaluate AI claims, for how we treat animals, for medical decisions about patients who cannot report their experience. Philosophy is in the hexagon because the science needs its concepts clarified, and because bad conceptual reasoning produces bad experiments.

My aim here is not to sell you a position. It is to make you able to argue both sides of each dispute well enough that you could switch teams and still perform. That is the actual test of understanding an argument.

Turing's test, stated correctly

In 1950, Alan Turing opened a paper with "Can machines think?" and immediately declared the question too meaningless to deserve discussion, because it depends on how we choose to use the words. He replaced it with an operational substitute, the imitation game. An interrogator communicates by text with two hidden participants, one human and one machine, and tries to determine which is which. If the machine cannot be reliably distinguished, Turing proposed, we should say it thinks, or rather, we should stop pretending we have a better criterion.

Two things about this are widely misunderstood. First, Turing was not offering a definition of intelligence or a measurement instrument; he was making a philosophical move, shifting the burden from unobservable inner essence to observable performance. Second, he anticipated the objections and answered them in the same paper, which is why it remains worth reading. To the theological objection, he replied that it presumes what it should argue. To the argument from consciousness, that a machine cannot be said to think until it feels, he gave the reply that still bites: taken seriously, that argument forces solipsism, since your only evidence that other people are conscious is also behavioral. To Lady Lovelace's objection, that machines can only do what we tell them, he pointed out that machines regularly surprise their programmers and that learning machines could acquire what nobody specified.

The objections that landed came later. The test measures deception, not intelligence, and rewards evasive conversational strategies; the 1966 program ELIZA, which mostly reflected users' statements back as questions, elicited genuine emotional engagement from people who knew it was a program, a phenomenon named the ELIZA effect. The test is also anthropocentric: it asks whether a system can pass as human, which excludes intelligences that are real but different, and it ignores everything non-linguistic. Most working researchers now treat the Turing test as a historically pivotal thought experiment rather than a live evaluation standard, and evaluate systems with targeted benchmarks instead, with all the caveats Lesson 14 attached to those.

Key idea: Turing replaced an unanswerable question about inner essence with an operational test of indistinguishable performance, and pre-answered several objections, notably that denying machine thought on consciousness grounds threatens to deny it of other people too. The test's weaknesses are that it measures human-imitation and rewards deception.

The Chinese Room

In 1980 John Searle published the most discussed counterargument in the field's history. Imagine yourself locked in a room. You do not know Chinese. Slips of paper with Chinese characters come through a slot. You have an enormous rulebook, written in English, that tells you which characters to write in response to which incoming characters, purely by their shapes. You follow it meticulously and pass the responses back out. To Chinese speakers outside, your answers are indistinguishable from those of a native speaker. But you understand nothing. You are manipulating symbols by their form, with no access to their meaning.

Searle's conclusion: running the right program is not sufficient for understanding. Syntax, formal symbol manipulation, does not by itself yield semantics, meaning. Therefore the strong AI claim, that an appropriately programmed computer would thereby have a mind, is false. Note the precision of the target: Searle does not deny that machines could think, and he explicitly allows that a machine with the right causal powers, a brain being the example he has in hand, does think. His claim is that the program alone is not what does it.

Now the replies, which Searle published alongside the argument and answered.

The systems reply. Of course the person does not understand Chinese. The person is only the central processing unit. The system as a whole, person plus rulebook plus paper plus the whole apparatus, understands. Searle's answer: let the person memorize the entire rulebook and do it all in their head, outdoors, with no room. Now the person is the system, and still understands nothing. Critics reply that this response underestimates what memorizing an entire language-processing program would involve, and that a person who internalized such a system might well have implemented a second, understanding subject.

The robot reply. The problem is that the system has no causal contact with the world. Put the program in a robot with cameras, effectors, and a body, so its symbols connect to things. Searle answers that adding perceptual inputs just adds more uninterpreted symbols; nothing about being wired to a camera makes symbol shuffling meaningful. This is where the argument connects to the symbol grounding problem, named by Stevan Harnad: how do internal symbols get their meaning without being interpreted by another mind? Many find the robot reply the most promising, and it links directly to embodied cognition, which the next lesson takes up.

The brain simulator reply. Suppose the program simulates, neuron by neuron, exactly what happens in a Chinese speaker's brain. Searle: a simulation of a thing is not the thing. A perfect simulation of a rainstorm leaves nobody wet, and a perfect simulation of digestion digests no food. Opponents answer that this analogy may fail for information processing, because unlike rain, cognition might be constituted by the processing itself rather than by the substrate.

Where does this leave you? The argument has been debated for over forty years without resolution, which itself tells you something: it targets an intuition (understanding is not the same as behaving) that many people share strongly and others reject outright. Two things are worth taking away regardless of your verdict. First, the argument sharpened the distinction between syntax and semantics, and the grounding problem it raises is a live technical question, not just a philosophical one. Second, notice that Searle's argument is not empirical: no experiment would settle it, which is exactly why it belongs to philosophy's corner of the hexagon rather than psychology's.

Key idea: The Chinese Room argues that formal symbol manipulation is insufficient for understanding, because syntax does not yield semantics. The systems, robot, and brain simulator replies each locate the understanding somewhere Searle's thought experiment does not look, and the dispute remains unresolved.

The hard problem

Now to consciousness, and the first task is to say what the problem actually is, because the word covers several different things. David Chalmers drew the crucial distinction in 1995. The easy problems are the ones about function: how does the brain discriminate stimuli, integrate information, report mental states, focus attention, control behavior, produce the difference between waking and dreamless sleep? These are called easy only by comparison; they are ordinary hard science, and cognitive science is making steady progress on them.

The hard problem is different in kind. Why is any of that accompanied by subjective experience at all? Why is there something it is like, in Thomas Nagel's phrase from his 1974 paper asking what it is like to be a bat, to see red rather than merely to process wavelength information and act accordingly? You could specify every functional detail of color vision and it would seem, to many people, that a further question remains unanswered. Frank Jackson's knowledge argument dramatizes it: Mary is a brilliant scientist who knows every physical fact about color vision but has lived her whole life in a black and white room. When she leaves and sees red for the first time, does she learn something new? If yes, then physical facts did not include everything.

Not everyone accepts the hard problem as legitimate. Daniel Dennett argued for decades that it dissolves once the easy problems are solved, that the residue is an illusion generated by our own cognitive machinery, and that the intuition of an unexplained remainder is exactly what a well-designed self-monitoring system would produce. Illusionists such as Keith Frankish make this explicit: the target of explanation is not phenomenal experience but why we are so convinced we have it. This is not a fringe position, and a student should be able to state it fairly.

Key idea: The easy problems concern functions such as discrimination, integration, attention, and report; the hard problem asks why any of it is accompanied by subjective experience. Illusionists deny the hard problem is a genuine explanandum, arguing instead that we must explain the conviction.

Two scientific theories, sketched

Meanwhile empirical work proceeds on the easy problems and on the search for the neural correlates of consciousness, and two theories currently dominate the discussion.

Global workspace theory, proposed by Bernard Baars and developed neurally by Stanislas Dehaene and Jean-Pierre Changeux, treats consciousness as broadcasting. Most processing runs in parallel, specialized, and unconscious. A stimulus becomes conscious when it wins a competition and is broadcast widely across a fronto-parietal network, making it available to memory, report, planning, and voluntary action. The theory is functional and testable, and it makes predictions: it accounts for the capacity limit of consciousness, for why attention and consciousness are related but separable, and for the finding that stimuli near threshold produce an all-or-none late brain response, sometimes called ignition, when reported as seen. Critics point out that it is a theory of access, of what makes information globally available, and therefore may not address the hard problem at all.

Integrated information theory, developed by Giulio Tononi, starts from the opposite end: from the properties experience seems to have, and asks what a physical system must be like to have them. It proposes that consciousness is identical to integrated information, quantified as a measure called phi, which captures how much a system's causal structure is both differentiated and unified, more than the sum of its parts. Strong claims follow: consciousness comes in degrees, could exist in systems very unlike brains, and, in principle, is absent in feed-forward systems no matter how intelligent their behavior. That last implication is why the theory says a digital simulation of a brain could behave identically while being far less conscious, which many find either its most profound feature or its clearest problem. Critics note phi is essentially incomputable for real systems, and a widely publicized 2023 open letter signed by many researchers called the theory pseudoscience for making unfalsifiable claims, a charge its proponents contest vigorously.

Notably, the two theories have been pitted against each other in adversarial collaborations, in which proponents of both agree in advance on experiments and on what results would count against their own view. The first major results, reported in 2025, supported some predictions of each and challenged others, with posterior cortical involvement fitting integrated information theory better and some frontal findings fitting global workspace better, and neither theory emerging cleanly victorious. Whatever you conclude about consciousness, that methodology is a model for how a field should handle theoretical disputes, and it is worth knowing about as a piece of scientific practice.

Key idea: Global workspace theory explains consciousness as broadcast of information to a wide fronto-parietal network, addressing access; integrated information theory identifies consciousness with integrated causal structure measured by phi, addressing experience but facing charges of unfalsifiability. Adversarial collaborations are now testing both.

Common misconceptions

• "The Turing test defines intelligence." Turing offered an operational replacement for an unanswerable question, not a definition, and its weakness is that it measures successful imitation of a human rather than intelligence as such.
• "Searle argued machines can never think." He explicitly allows that machines with the right causal powers think, brains being his example. His target is the narrower claim that running the right program suffices.
• "The Chinese Room has been refuted." It has been answered many times in many ways, none of which commanded consensus. Reporting it as settled in either direction misrepresents the literature.
• "The hard problem is just the easy problems in disguise." That is one serious position, illusionism, not a neutral statement. Others hold it is a distinct explanatory gap. Both camps contain careful thinkers.
• "Neuroscience has located the seat of consciousness." Research identifies neural correlates and tests specific theories. Correlates are not explanations, and current theories disagree about which brain regions are even the right place to look.

Recap

Turing replaced "can machines think" with an operational test of indistinguishable text conversation, pre-answering objections including the argument from consciousness, which he noted would force solipsism if applied consistently; the test's real weaknesses are that it measures human imitation and rewards deception, as the ELIZA effect showed. Searle's Chinese Room argues that syntax is insufficient for semantics, so running a program cannot by itself produce understanding, and the systems, robot, and brain simulator replies each relocate the understanding in ways Searle rejects and critics defend, leaving the debate open and the symbol grounding problem live. Chalmers separated the easy problems of function from the hard problem of why function is accompanied by experience, with illusionists denying the hard problem is a genuine target. Global workspace theory explains conscious access through broadcast across a fronto-parietal network; integrated information theory identifies consciousness with integrated causal structure and faces falsifiability objections. Adversarial collaborations testing both are the healthiest development in the area. One lesson remains: where cognition lives, what animals think, and where this field is heading.

Sources

1. Oppy, G., and Dowe, D. (2021). The Turing test. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
2. Cole, D. (2024). The Chinese Room argument. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. Van Gulick, R. (2022). Consciousness. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
4. Encyclopaedia Britannica. (2024). Alan Turing. Encyclopaedia Britannica. britannica.com
5. Wikipedia. (2025). Integrated information theory. Wikimedia Foundation. en.wikipedia.org

The big picture

We began with a hexagon and a bet: that the mind is an information-processing system that six disciplines can study together. Fifteen lessons later, you have watched that bet pay off repeatedly and also strain at the edges. This final lesson gathers the strain into three questions and then, honestly, tells you what to do with all of it.

First: is cognition confined to the brain, or does it include the body, the tools, and the environment? Second: how much of what we have studied is uniquely human, and what does animal research reveal? Third: how healthy is this science, what did the replication crisis do to it, and where can it take you? Consider this the lesson where the course stops teaching content and starts handing you the keys.

Four E's beyond the skull

The classical picture treats the brain as a central processor: input comes in, computation happens, output goes out, and body and world are just the periphery. A cluster of research programs, often called the four E's, argues that this misplaces the boundary.

Embodied cognition, which you met in Lesson 11, holds that the specific body shapes the concepts and processes available. Embedded or situated cognition emphasizes that intelligent behavior exploits environmental structure rather than modeling it internally. Enacted cognition, from Francisco Varela and colleagues, holds that cognition arises through an organism's active engagement with its world rather than through internal representation of a pre-given reality. And extended cognition claims that cognitive processes can literally include things outside the body.

The situated case is easier to accept than it first sounds, and the examples are concrete. Bartenders arrange differently shaped glasses in the order drinks were requested, offloading memory onto the physical layout. Experienced Tetris players rotate pieces on screen rather than mentally, because the physical rotation is faster than the mental one, a case David Kirsh and Paul Maglio called epistemic action: acting to make thinking easier rather than to accomplish a goal directly. Air traffic controllers manipulate physical or virtual strips that hold the state of the airspace so it does not have to be held in working memory. Even counting on your fingers is a person recruiting an external structure into an internal process. Once you look, cognition is constantly leaning on the world.

Key idea: Embodied, embedded, enacted, and extended approaches argue that the brain is not a self-sufficient processor. People routinely offload cognitive work onto bodies and environments, as bartenders, Tetris players, and finger-counters all show.

The extended mind, argued both ways

The strongest version deserves its own treatment. In a 1998 paper, Andy Clark and David Chalmers introduced Otto and Inga. Inga hears about an exhibition, recalls from memory that the museum is on 53rd Street, and goes. Otto has Alzheimer's disease and writes everything in a notebook he carries constantly. He consults the notebook, reads that the museum is on 53rd Street, and goes.

Clark and Chalmers argue that if we count Inga's belief as a standing belief before she consults her memory, we should count Otto's notebook entry the same way. Their parity principle: if a process in the world functions as a process in the head would, and we would call the internal one cognitive, then the external one is cognitive too. The notebook is not a tool Otto uses to think; on this view, it is part of his cognitive system. This is the extended mind thesis, and if you carry a phone with your calendar, contacts, and notes, it is a claim about you.

The main objections are serious. First, coupling is not constitution: your thermostat is coupled to your comfort without being part of your physiology, and the fact that a resource is reliably used does not make it a part of the system that uses it. Second, Fred Adams and Ken Aizawa's mark of the cognitive objection: internal memory and notebook entries differ in ways that matter, since biological memory is subject to interference, reconstruction, priming, and the fallibility patterns you studied in Lesson 8, while a notebook is not. If the two have different laws, calling them the same kind of thing may obscure more than it reveals. Third, the thesis may prove too much: with generous parity, almost everything becomes cognitive, and a concept that includes everything explains nothing.

Defenders answer that the objection about differing mechanisms mistakes implementation for function, which is Marr's point in reverse, and that requiring internal and external processes to work identically stacks the deck. This dispute has not been settled either, and you should now find that unsurprising: like the imagery debate and the Chinese Room, it turns on what our concepts should carve, not on a missing measurement. What is not disputed is the empirical core: people really do distribute cognitive work across brain, body, and world, and studying only the brain misses much of how tasks get done.

Key idea: The extended mind thesis uses the parity principle to argue that Otto's notebook is part of his memory. Critics answer that coupling is not constitution and that internal and external memory obey different laws. The empirical claim about distributed cognitive work survives whichever way the conceptual question falls.

What other animals think

Comparative cognition is the other frontier that keeps the field honest, and it comes with a methodological warning attached to a horse. In the early 1900s, Clever Hans appeared to do arithmetic, tapping out answers with his hoof. A careful investigation by Oskar Pfungst showed that Hans was reading subtle, unintentional postural cues from questioners, and failed whenever the questioner did not know the answer. Hans was doing something remarkable, just not arithmetic. The Clever Hans effect is why comparative research now uses blind procedures as standard, and why extraordinary claims about animal minds get examined for simpler explanations first.

The complementary error is the opposite one. Anthropomorphism attributes human mental states without evidence; Frans de Waal coined anthropodenial for the reflexive refusal to attribute any mental states to animals whose behavior and neural machinery closely resemble ours. Both are failures of evidence discipline. Morgan's canon, the rule that we should not attribute a higher psychological process when a lower one suffices, is the traditional corrective, though it too can be applied dogmatically.

With those cautions in place, the findings are striking. Corvids plan: New Caledonian crows make and modify hook tools, and one, Betty, famously bent wire to retrieve food. Western scrub jays cache food and recover perishable items sooner, and re-cache when they were watched by another bird, but only if they themselves have stolen caches before, which some interpret as experience-based social prediction. Chimpanzees make tool kits, transmit local traditions, and pass mirror self-recognition tests, as do elephants, dolphins, and magpies. Dolphins and elephants show cooperative problem solving. Alex the grey parrot, studied by Irene Pepperberg for thirty years, labeled objects, colors, and materials, answered questions about which of two objects differed, and appeared to grasp a concept of none.

And the honest interpretive frame: these are not partial credit toward being human. Each species solves the problems its niche poses, so cognition is a bush rather than a ladder. Bats navigate by echolocation, honeybees encode distance and direction in a dance, and octopuses, with a nervous system organized so differently that most neurons sit in the arms, solve manipulation problems in ways that make the very question "how smart are they compared to us" look badly posed. The comparative lesson for cognitive science is that human cognition is one solution among many, and the features we think define mind (language, tools, planning, self-recognition) are distributed unevenly across the tree.

Key idea: Comparative research documents planning, tool manufacture, culture, and self-recognition across corvids, primates, cetaceans, and parrots, guarded by blind procedures because of the Clever Hans effect. Cognition is a branching bush of niche-specific solutions, not a ladder with humans on top.

The replication crisis and what the field did

You have met this thread in almost every module: infant mentalizing, embodiment effects, some priming and ego-depletion findings, some imaging analyses. It is time to name it. Beginning around 2011, psychology confronted evidence that a substantial fraction of published findings did not replicate. The Open Science Collaboration's 2015 project repeated 100 published studies and successfully replicated a minority, with effect sizes about half the originals on average. Social priming effects that had been textbook staples repeatedly failed.

The causes were structural rather than villainous: small samples with low statistical power, flexible analysis choices that let researchers find something publishable (p-hacking), reporting hypotheses after results were known, publication bias against null results, and a career system rewarding novelty over verification.

What matters for you is what happened next, because the response is the best argument for taking this science seriously. Preregistration of hypotheses and analysis plans became common. Registered reports, where journals accept a study on its design before results exist, removed the incentive to find something. Data and code sharing became normal. Sample sizes grew and power analysis became routine. Multi-laboratory consortia, ManyBabies and ManyLabs among them, now test key effects at scale. Replication studies became publishable rather than career suicide.

The right attitude is neither dismissal nor denial. A field that audits itself, publishes the failures, and changes its methods is behaving as a science should. The findings in this course that survived scrutiny, the ones repeatedly replicated across laboratories and methods, are worth your confidence precisely because the field checked. And you now have the habits to ask, of any new claim: how large was the sample, was it preregistered, has it replicated independently, and what would count as evidence against it?

Key idea: The replication crisis exposed low power, analytic flexibility, and publication bias. The field responded with preregistration, registered reports, data sharing, larger samples, and multi-laboratory consortia, which is what self-correction looks like in practice.

Where this takes you

Cognitive science is unusual in that its graduates rarely have the job title. What they have is a combination that is hard to assemble any other way: experimental design, statistics, computational modeling, and conceptual clarity about mind and behavior.

Path	What you would do	Typical preparation
Research science	Run studies in cognitive psychology, neuroscience, linguistics, or development	PhD, usually preceded by undergraduate lab experience
Human-computer interaction and UX research	Study how people actually use systems and redesign them accordingly	Bachelor's or master's plus a portfolio of real studies
Machine learning and AI research	Build and evaluate learning systems, often on the cognitive side of evaluation	Strong programming, mathematics, and often graduate study
Clinical and speech-language paths	Assess and treat language, memory, and attention disorders	Accredited graduate degree plus licensure
Education and learning science	Apply memory and instruction research to how material is taught	Graduate work in learning sciences or education research
Applied policy, safety, and forensics	Eyewitness procedure, aviation and medical human factors, behavioral policy	Domain graduate training plus methods background

The practical advice is the same regardless of destination: learn to program, learn statistics properly rather than by recipe, get into a research laboratory early, and write. The methodological training is the transferable part.

As for where the field is going, four directions look live. Naturalistic and large-scale methods are replacing artificial laboratory tasks, studying cognition in movies, conversations, and daily life with sensors and big datasets. Computational modeling is increasingly the common language across the hexagon. Cross-cultural work is confronting the fact, documented forcefully by Joseph Henrich and colleagues, that most published findings come from WEIRD samples, western, educated, industrialized, rich, and democratic, which are unrepresentative of humanity on many measures. And artificial systems have become both tools and objects of study, forcing old questions about representation, learning, and understanding into unavoidably concrete form.

Common misconceptions

• "Extended cognition means your phone is conscious." The thesis is about cognitive processes such as belief storage and problem solving, not about experience. Even its defenders do not claim the notebook feels anything.
• "Animal cognition research shows animals are like little humans." It shows species-specific solutions to species-specific problems. Cognition is a branching bush, and treating humans as the endpoint of a ladder misreads the evidence.
• "Clever Hans proved animals cannot count." It proved that unblinded procedures leak cues. Later blind studies do find real numerical abilities in several species; the lesson is about method, not about animal limits.
• "The replication crisis means psychology is worthless." It means a field measured its own error rate, published the result, and changed its methods. Findings that survived rigorous replication deserve more confidence, not less.
• "Cognitive science findings apply to everyone equally." Most published work uses WEIRD samples that differ from most of humanity on measures including perception, reasoning, and social behavior, which is why cross-cultural replication matters.

Recap

Embodied, embedded, enacted, and extended approaches challenge the brain-as-central-processor picture, and the empirical core is not in doubt: people offload cognitive work onto bodies and environments constantly. The extended mind thesis pushes further with Otto's notebook and the parity principle, and meets the objections that coupling is not constitution and that internal and external memory obey different laws. Comparative research, disciplined by the Clever Hans effect and by the twin errors of anthropomorphism and anthropodenial, documents tool manufacture, planning, culture, and self-recognition across many lineages and reframes cognition as a bush of niche solutions rather than a ladder. The replication crisis revealed low power, flexible analysis, and publication bias, and the field answered with preregistration, registered reports, data sharing, and multi-laboratory consortia. Cognitive science trains a rare combination of experimental design, statistics, modeling, and conceptual clarity, feeding research, HCI, AI, clinical, educational, and policy careers. The hexagon still stands, the arguments are still open, and you are now equipped to follow them and to ask, of any claim about the mind, exactly what the evidence shows.

Sources

1. Clark, A., and Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7-19. Overview: plato.stanford.edu
2. Andrews, K. (2024). Animal cognition. Stanford Encyclopedia of Philosophy. Stanford University. plato.stanford.edu
3. Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251). pubmed.ncbi.nlm.nih.gov
4. American Psychological Association. (2024). Careers in psychology and cognitive science. APA. apa.org
5. Wikipedia. (2025). Clever Hans. Wikimedia Foundation. en.wikipedia.org