In September 1956, a small conference on information theory at MIT brought together mathematicians, engineers, and psychologists who sensed they were standing at the edge of a new way of thinking about thinking. Within weeks, a parallel gathering at Dartmouth was proposing that intelligence itself could be mechanized. George Miller had just submitted a paper arguing that human short-term memory is constrained not by raw capacity but by the number of chunks of information it can hold -- a claim that borrowed its vocabulary directly from telecommunications engineering. Something was shifting in the study of the mind, and it was shifting fast.
The field that crystallized from those encounters -- cognitive science -- is now more than six decades old, and it remains one of the most restless and generative enterprises in the modern academy. It has no single founding text, no single methodology, and no settled consensus on its most important questions. What it has instead is a shared conviction that the mind can be studied scientifically, that mental representations and processes are real objects of inquiry, and that progress requires drawing on psychology, neuroscience, computer science, linguistics, philosophy, and anthropology simultaneously rather than sequentially. This breadth is not a weakness; it is the field's defining feature and the source of its most productive tensions.
Understanding cognitive science means understanding both what it has achieved and what it continues to contest. Its models of memory and attention have reshaped clinical practice. Its computational architectures have driven decades of AI research. Its theories of embodiment and language have challenged assumptions that seemed secure. And its engagement with artificial intelligence -- now more urgent than ever in the era of large language models -- keeps returning it to the oldest and hardest questions: what is thinking, and what kind of thing does the thinking?
"The mechanism of our cognitive behavior is to be understood only in terms of representations and processes that operate on representations." -- Ulric Neisser, Cognitive Psychology, 1967
Key Definitions
Cognition: The set of mental processes involved in acquiring, storing, transforming, and using knowledge, including perception, attention, memory, language, reasoning, and problem-solving.
Representation: An internal state that stands for or encodes information about something in the world; cognitive science is largely the study of what representations exist, how they are structured, and how they are processed.
| Cognitive Science Discipline | Core Question | Key Method |
|---|---|---|
| Cognitive Psychology | How do humans perceive, remember, and think? | Experiments, response times, error analysis |
| Neuroscience | What neural structures underlie cognition? | fMRI, EEG, lesion studies |
| Linguistics | How does language structure thought and communication? | Corpus analysis, cross-linguistic comparison |
| Artificial Intelligence | Can cognition be computationally modeled? | Algorithm design, machine learning |
| Philosophy of Mind | What is the nature of mental states and consciousness? | Conceptual analysis, thought experiments |
| Anthropology | How does culture shape cognitive processes? | Ethnography, cross-cultural experiments |
Computational model: A formal specification, often implemented as a running program, of the cognitive processes hypothesized to produce some behavioral phenomenon; models generate testable predictions about human performance.
Cognitive architecture: A theory of the fixed computational and representational structure of the human cognitive system; ACT-R and SOAR are the most influential examples.
4E cognition: A family of related positions -- embodied, embedded, enacted, and extended -- that challenge the classical view by arguing that cognition is not confined to the brain but constituted by the whole brain-body-world system.
The Cognitive Revolution
Behaviorism and Its Limits
Behaviorism dominated American psychology from John Watson's 1913 manifesto through the 1950s. Its core methodological commitment was to study only observable stimuli and responses, treating mental states, images, and intentions as unscientific. B.F. Skinner's radical behaviorism, the most influential version, held that behavior is fully explained by schedules of reinforcement without invoking internal representations at all.
By mid-century, the limits of this program were becoming visible to researchers working on problems that required internal structure to explain. Workers trying to design radar displays and communications systems during World War II had confronted the information-processing limits of human operators. Norbert Wiener at MIT was developing cybernetics, a theory of feedback, control, and communication in animals and machines that required mentalistic vocabulary to be rigorous. Claude Shannon's mathematical theory of communication (1948) provided a precise, quantitative framework for talking about information that could be applied equally to machines and nervous systems.
1956 and the New Vocabulary
The year 1956 is conventionally treated as the birth year of cognitive science, and the concentration of events in that year is striking enough to justify the convention. Allen Newell and Herbert Simon demonstrated the Logic Theorist at the Dartmouth conference -- a program that proved theorems in Whitehead and Russell's Principia Mathematica by searching a space of possible proofs, and that was explicitly designed as a model of how a human mathematician might proceed.
George Miller published "The Magical Number Seven, Plus or Minus Two" in Psychological Review, using information theory to characterize the capacity of short-term memory as approximately seven chunks of information. The concept of chunking -- grouping lower-level elements into higher-level units through learning -- anticipated later work on expertise, schema formation, and skill acquisition.
Noam Chomsky's review of Skinner's Verbal Behavior (1959) completed the demolition of the behaviorist account of language. Chomsky argued that the productivity and systematicity of human language -- the ability to generate and understand an unlimited number of novel sentences -- cannot be explained by stimulus-response associations but requires positing an internalized, generative grammar. The argument was not merely about language; it was a demonstration that some psychological phenomena demand internal structure to explain.
Ulric Neisser's synthesis, Cognitive Psychology (1967), gave the new approach its name and its program. Drawing on work in perception, pattern recognition, selective attention, and memory, Neisser framed cognition as a set of active constructive processes: the mind does not passively receive information but actively selects, transforms, encodes, stores, and retrieves it.
Computational Models and Cognitive Architectures
The Physical Symbol System Hypothesis
Newell and Simon's most sweeping theoretical claim came in their 1976 Turing Award lecture, "Computer Science as Empirical Inquiry: Symbols and Search." They argued for the physical symbol system hypothesis: that any physical system capable of manipulating symbols according to rules has, in principle, the necessary and sufficient means for general intelligent action. The corollary is that both the human mind and a digital computer are physical symbol systems, differing in implementation but not in computational character.
This hypothesis provided theoretical grounding for cognitive modeling. If the mind is a symbol system, then a running computer program that manipulates the right symbols in the right ways is literally a theory of cognition -- not an analogy or a metaphor but an instantiation of the same computational process.
Production systems -- rule-based programs in which condition-action rules fire when their conditions are matched in a working memory -- became the preferred formalism. Their modularity made them tractable for modeling specific cognitive tasks while remaining commitments to a general architecture.
ACT-R and SOAR
John Anderson at Carnegie Mellon developed ACT-R (Adaptive Control of Thought-Rational), a cognitive architecture that models how the brain stores and retrieves knowledge. ACT-R represents procedural knowledge as production rules and declarative knowledge as chunks stored in a distributed associative memory, where each chunk has an activation level that determines its availability. The architecture has been used to model learning, skill acquisition, problem-solving, language comprehension, and driving behavior, generating quantitative predictions that are tested against human response times and brain imaging data.
SOAR, developed by Newell, John Laird, and Paul Rosenbloom at Carnegie Mellon and then the University of Michigan, is a more comprehensive architecture that attempts to model the full range of human cognition within a single unified system. SOAR uses a problem space formalism: cognition is modeled as search through a space of states using operators, with impasses -- moments when the system cannot proceed -- handled by a chunking mechanism that learns new rules from experience. SOAR has been used in AI applications, robotic control, and cognitive modeling, and it remains an active research platform.
Connectionism and the Neural Network Challenge
Parallel Distributed Processing
The publication of David Rumelhart and James McClelland's Parallel Distributed Processing (1986) in two volumes was a turning point. PDP models represent knowledge not in discrete symbols but in patterns of activation distributed across layers of simple neuron-like units connected by weighted links. Learning occurs through backpropagation: an algorithm that adjusts connection weights by propagating error signals backward through the network from output to input.
Connectionist models offered several advantages over symbolic systems. They learned from examples rather than from hand-coded rules. They generalized in psychologically plausible ways. They degraded gracefully when parts of the network were damaged, unlike symbolic programs that tend to fail catastrophically. They captured frequency effects, typicality effects, and other graded phenomena that symbolic models had struggled with.
The Fodor-Pylyshyn Critique
Jerry Fodor and Zenon Pylyshyn's 1988 paper "Connectionism and Cognitive Architecture" launched the most influential challenge to connectionism from the symbolic side. They argued that human thought is systematic: if a person can think "John loves Mary" they can also think "Mary loves John"; if they can think "the dog is biting the man" they can think "the man is biting the dog." This systematicity -- the fact that the ability to entertain one thought is always accompanied by the ability to entertain related thoughts that share its constituents -- requires that mental representations have constituent structure, that is, symbolic compositionality. Distributed representations in connectionist networks do not obviously have this property.
Fodor and Pylyshyn argued that connectionism might be a good model of neural implementation but not of the computational level at which cognition is properly described. The debate was never resolved to everyone's satisfaction and remains live, sharpened now by the need to explain what, if anything, large language models understand.
Deep Learning and the Vindication of Connectionism
From 2012 onward, deep convolutional networks achieved superhuman performance on image recognition, speech recognition, game-playing, and eventually text generation. AlphaGo's defeat of world champion Lee Sedol in 2016 using deep reinforcement learning was widely reported as a landmark. Large language models trained on massive text corpora demonstrated that much of what appears to require symbolic reasoning can be achieved by sufficiently large and deeply trained neural networks.
The vindication is empirical but its theoretical interpretation is disputed. Whether deep learning systems have learned representations that are compositional, systematic, and genuinely conceptual, or whether they have learned powerful statistical approximations that mimic these properties in limited ways, is one of the central open questions in cognitive science and AI.
Embodied, Embedded, and Extended Cognition
The Embodied Mind
The standard computational model treats the body as an input-output device for a central processor in the brain. Researchers in the embodied cognition tradition have argued that this picture fundamentally mischaracterizes what cognition is. Cognition is not computation over amodal symbols but the dynamic regulation of the organism's relationship with its environment through action.
Francisco Varela, Evan Thompson, and Eleanor Rosch's The Embodied Mind (1991) drew on phenomenology -- especially Merleau-Ponty's account of the body as the site of perception and action -- and on biological systems theory to argue that cognition is enacted: it arises through the history of structural coupling between an organism and its environment. There is no pre-given world of objects and properties that cognition represents; the world as experienced is constituted through action.
George Lakoff and Mark Johnson's Philosophy in the Flesh (1999) argued that abstract conceptual thought is structured by bodily metaphors grounded in physical experience. The concept of time is understood through spatial metaphor (time "passes," the future is "ahead"); argument is understood through conflict metaphor; understanding is understood through vision. Conceptual metaphor theory, developed earlier in Lakoff's Women, Fire, and Dangerous Things (1987) and Lakoff and Johnson's Metaphors We Live By (1980), proposed that these mappings from physical experience to abstract domains are not rhetorical ornaments but constitutive of thought itself.
The Extended Mind
Andy Clark and David Chalmers' 1998 paper "The Extended Mind" posed a simple thought experiment. Otto has Alzheimer's disease. He writes information in a notebook and consults it whenever he needs to remember. Inga has a normal memory and retrieves information from her head. Clark and Chalmers argued that if we would describe Inga as using memory, we should describe Otto as using memory too -- the notebook is functionally equivalent to a memory store, and the accident that it is outside Otto's skull gives no principled reason to treat it differently.
This "active externalism" is controversial. Critics argue that the notebook is a tool, not part of Otto's mind, and that confusing the two leads to conceptual inflation that obscures rather than illuminates. Clark responds that the resistance to external cognition reflects a bias toward the skin-and-skull boundary that has no theoretical justification.
The 4E framework -- embodied, embedded, enacted, extended -- now serves as a heading for a large and heterogeneous research program that includes robotics researchers building situated agents, phenomenologists studying skilled action, psychologists studying distributed cognition in teams and workplaces, and neuroscientists studying sensorimotor loops.
Memory and Attention
Working Memory
Alan Baddeley and Graham Hitch's 1974 working memory model replaced the earlier notion of a unitary short-term memory store -- derived from Atkinson and Shiffrin's 1968 modal model -- with a multi-component system. The phonological loop maintains and rehearses auditory-verbal information; it is why you can temporarily remember a phone number by repeating it to yourself. The visuospatial sketchpad holds and manipulates visual and spatial information; it supports tasks like mental rotation and navigation. The central executive is an attentional control system that allocates resources between the other components and interfaces with long-term memory; it is the most theoretically demanding component and the least well specified.
Baddeley added a fourth component in 2000, the episodic buffer, which integrates information from the phonological loop, the visuospatial sketchpad, and long-term memory into coherent, multimodal episodic representations. The working memory model has been enormously productive, generating hundreds of experiments and clinical applications in understanding reading difficulties, aging, and cognitive impairment.
Episodic Memory and H.M.
Endel Tulving's 1972 distinction between episodic and semantic memory addressed a puzzle in the literature: people could lose autobiographical memory for specific events while retaining general world knowledge, and vice versa. Episodic memory is specifically personal, tied to the experiential context of original encoding; semantic memory is decontextualized and impersonal.
The case of Henry Molaison (H.M.), operated on by surgeon William Beecher Scoville in 1953 to relieve intractable epilepsy, provided the most influential evidence in twentieth-century memory research. The removal of most of the hippocampus bilaterally left H.M. unable to form new long-term declarative memories -- he could not remember meeting researchers he saw every day -- while leaving his procedural skills, working memory, and long-term memories from before the surgery largely intact. The case demonstrated the hippocampus's role in memory consolidation and the separability of memory systems, shaping decades of subsequent research and clinical understanding of amnesia.
Fergus Craik and Robert Lockhart's levels-of-processing framework (1972) challenged the structural approach by arguing that memory depends not on which store information enters but on the depth of processing at encoding. Information processed for its phonological properties is remembered less well than information processed for its meaning; elaborative, meaningful encoding produces stronger, more durable traces. The framework generated productive debates about what "depth" means and how it interacts with distinctiveness and retrieval conditions.
Attention
Donald Broadbent's filter theory (1958) proposed that the information processing system has a limited capacity channel and that a selective filter operates early, before full analysis, admitting attended channels and blocking unattended ones. The dichotic listening paradigm -- in which subjects attend to one ear while a different message plays in the other -- provided the key evidence.
Anne Treisman's attenuation model (1964) modified filter theory in response to evidence that attended information can capture attention even when it is unattended: a subject's name, for instance, is often detected in the unattended channel. Rather than blocking the unattended channel entirely, Treisman proposed that it is attenuated, weakened but not eliminated, and that sufficiently important stimuli can break through.
Treisman and Garry Gelade's feature integration theory (1980) shifted attention research toward visual search. They proposed that visual features such as color, orientation, and size are registered automatically and in parallel across the visual field in a preattentive stage. Binding these features into unified objects requires a second stage of focused, serial attention. The theory makes specific predictions about when searches will be efficient (when a target has a unique feature that "pops out") versus slow and serial (when the target must be distinguished from distractors that share its features), predictions extensively tested in experiments and applied in interface design.
Language, Thought, and Cognitive Linguistics
Prototype Theory and Basic-Level Categories
Eleanor Rosch's work in the 1970s transformed cognitive psychology's account of concepts. The classical view, tracing to Aristotle, held that concepts are defined by necessary and sufficient conditions: a thing is a bird if and only if it satisfies the conditions for birdhood. Rosch's experiments showed that category membership is graded: robins are judged as more typical birds than penguins, and response times in category judgments reflect this typicality structure.
Rosch also identified a basic level of categorization -- between the superordinate (animal) and the subordinate (golden retriever) -- at which categories are psychologically most natural, most easily named, most quickly recognized, and around which cognitive representations cluster most tightly. Basic-level categories (dog, chair, car) are the default level of entry into conceptual knowledge.
Cognitive Linguistics and Embodied Semantics
George Lakoff's Women, Fire, and Dangerous Things (1987) -- the title refers to a category in the Australian language Dyirbal that groups these referents together under a single classifier -- extended prototype theory into a comprehensive alternative to classical formal semantics. Lakoff argued that categories are not defined by binary features but by prototype effects, basic-level structure, image schemas derived from bodily experience, and conceptual metaphors. Meaning is not a formal relation between symbols and external referents but a grounded, embodied construction.
Charles Fillmore's frame semantics proposed that words activate rich conceptual frames -- organized knowledge structures -- that provide the context for their meaning. Understanding the word "buy" requires activating the commercial transaction frame, with slots for buyer, seller, goods, and money. This insight influenced computational linguistics, natural language processing, and the FrameNet lexical database.
Lawrence Barsalou's work on grounded cognition extended the program empirically, arguing that conceptual representations are modality-specific simulations: activating the concept "apple" partially reactivates the visual, motor, and olfactory states associated with experiencing apples. Neuroimaging evidence has supported the prediction that conceptual processing recruits sensory and motor areas.
Cognitive Science and Artificial Intelligence
The relationship between cognitive science and AI has moved through several phases. In the founding period, AI and cognitive science were nearly the same enterprise. Newell and Simon wanted to understand human cognition; they built programs and tested them against human data. The expert systems era of the 1970s and 1980s created AI systems with commercial value but limited psychological plausibility, and the two fields began to diverge.
The connectionist revival brought them back into dialogue. Rumelhart's networks were simultaneously engineering systems and psychological models. But the deep learning revolution has again complicated the picture. Systems like GPT-4 and its successors perform impressively on tasks that seemed to require understanding -- answering questions, writing code, translating languages -- but whether they do so by virtue of representations that resemble human conceptual structures is deeply uncertain.
The Turing test, proposed by Alan Turing in "Computing Machinery and Intelligence" (1950), suggests that if a machine performs indistinguishably from a human in conversation, there is no reason to deny it intelligence. Many researchers find this criterion too permissive: performance can be achieved without the underlying structure that produces it in humans. Others argue that it is too restrictive: intelligence might be realized in radically different computational substrates without mimicking human behavior.
The question of consciousness -- whether there is something it is like to be a cognitive system, whether information processing is accompanied by subjective experience -- remains the hardest problem. Francis Crick and Christof Koch's neural correlates of consciousness program sought to identify the brain states that accompany conscious experience. David Chalmers' "hard problem" (1995) argued that explaining why any physical process is accompanied by experience is a different and deeper question than explaining the functional organization of cognition. Whether AI systems of sufficient complexity could be conscious, and what evidence could settle the question, is among the most contested issues at the intersection of cognitive science and AI.
References
Neisser, U. (1967). Cognitive Psychology. Appleton-Century-Crofts.
Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2), 81-97.
Newell, A., & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and search. Communications of the ACM, 19(3), 113-126.
Rumelhart, D. E., & McClelland, J. L. (Eds.). (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition (Vols. 1-2). MIT Press.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1-2), 3-71.
Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.
Clark, A., & Chalmers, D. J. (1998). The extended mind. Analysis, 58(1), 7-19.
Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. H. Bower (Ed.), The Psychology of Learning and Motivation (Vol. 8, pp. 47-89). Academic Press.
Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of Memory (pp. 381-403). Academic Press.
Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psychology, 12(1), 97-136.
Lakoff, G. (1987). Women, Fire, and Dangerous Things: What Categories Reveal about the Mind. University of Chicago Press.
Anderson, J. R. (1983). The Architecture of Cognition. Harvard University Press.
Frequently Asked Questions
What is cognitive science and why is it interdisciplinary?
Cognitive science is the scientific study of the mind and its processes, including perception, attention, memory, language, reasoning, and decision-making. It is inherently interdisciplinary because no single field can account for the full complexity of cognition on its own. Psychology contributes experimental methods and behavioral data. Neuroscience grounds mental phenomena in brain activity. Computer science provides formal models and the computational metaphor for cognition. Linguistics examines how language structures thought and communication. Philosophy raises foundational questions about consciousness, representation, and intentionality. Anthropology and cultural studies show how cognition is shaped by social and cultural context.The interdisciplinary character was built in from the start. The cognitive revolution of the 1950s brought together researchers who were dissatisfied with the limits of behaviorism and saw computation as a new vocabulary for describing mental processes. George Miller bridged psychology and information theory. Noam Chomsky challenged behaviorist accounts of language. Allen Newell and Herbert Simon built computer programs that modeled human problem-solving. Jerome Bruner pushed for cognitive approaches in developmental and educational research. Together they established an enterprise that continues to evolve as each contributing discipline advances and as new fields such as artificial intelligence, network science, and data science are drawn in.
What was the cognitive revolution and when did it happen?
The cognitive revolution is the name given to the intellectual shift in the mid-1950s that replaced behaviorism with a mentalist, information-processing approach to the study of mind and behavior. Behaviorism, dominant in American psychology since the early twentieth century, had insisted that science should study only observable stimuli and responses, treating mental states as unscientific fictions. The cognitive revolution rejected this constraint and asserted that internal mental representations and processes could and should be studied rigorously.Several events in 1956 are often cited as landmark moments. The Dartmouth Summer Research Project on Artificial Intelligence, organized by John McCarthy and Marvin Minsky, proposed that intelligence could be simulated in machines. George Miller published 'The Magical Number Seven, Plus or Minus Two' in Psychological Review, using information-theoretic concepts to explain limits on short-term memory. Allen Newell and Herbert Simon demonstrated the Logic Theorist, a program that could prove mathematical theorems, showing that symbol manipulation could model reasoning. Noam Chomsky delivered a critique of B.F. Skinner's behaviorist account of language, arguing that syntactic structure could not be explained by stimulus-response associations.Ulric Neisser's 1967 book 'Cognitive Psychology' gave the new field its name and synthesized work on perception, attention, memory, and language under the information-processing framework, cementing the revolution's gains.
What is the physical symbol system hypothesis and why does it matter?
The physical symbol system hypothesis (PSSH), formulated by Allen Newell and Herbert Simon in their 1976 Turing Award lecture 'Computer Science as Empirical Inquiry: Symbols and Search,' holds that a physical symbol system has the necessary and sufficient means for general intelligent action. A physical symbol system is any device that can manipulate symbols -- discrete, physically instantiated tokens -- according to rules, and that can represent and process information about the world through those symbols.The hypothesis matters because it provides a theoretical foundation for the view that both human minds and computers are, at some level of abstraction, the same kind of thing: systems that process symbols. This justified the program of cognitive science in which computational models -- production systems, logic programs, semantic networks -- are taken to be genuine theories of human cognition, not merely useful metaphors.The hypothesis is also controversial. John Searle's Chinese Room argument (1980) challenged whether symbol manipulation is sufficient for understanding or intentionality. Connectionist critics argued that cognition is better modeled by distributed, subsymbolic neural networks than by explicit rule-following symbol systems. Embodied cognition researchers questioned whether cognition is best described at the symbol level at all, pointing out that the brain is not a general-purpose serial processor but a massively parallel biological system deeply entangled with a body and an environment. The PSSH remains one of the central contested claims in cognitive science.
What is the difference between symbolic AI and connectionism?
Symbolic AI, the dominant paradigm from the 1950s through the 1980s, represents knowledge as discrete symbols and processes it through explicit logical or rule-based operations. Programs like the General Problem Solver (Newell and Simon), LISP-based expert systems, and natural language parsers all manipulate structured symbolic representations. Cognitive architectures such as ACT-R (John Anderson, Carnegie Mellon) and SOAR (Newell, Laird, and Rosenbloom) implement this approach as models of human cognition, specifying the computational mechanisms underlying memory, problem-solving, and skill acquisition.Connectionism, revived with force by David Rumelhart and James McClelland's 'Parallel Distributed Processing' (1986), models cognition with artificial neural networks: webs of simple units connected by weighted links, where knowledge is stored not in discrete symbols but in the pattern of connection weights distributed across the entire network. Learning occurs by adjusting weights through experience, not by programming explicit rules. Connectionist models captured important phenomena such as learning by gradient descent, generalization from examples, and graceful degradation under damage.Jerry Fodor and Zenon Pylyshyn launched an influential critique in 1988, arguing that connectionist networks lack the systematicity and compositionality that characterize human thought: if you can think 'John loves Mary' you can also think 'Mary loves John,' and this productivity requires symbolic structure, not distributed weights. The tension between symbolic and connectionist approaches has never been fully resolved. Deep learning's spectacular success since 2012 has vindicated connectionism empirically, while the debate about whether neural networks capture the right kind of cognitive structure continues.
What is embodied and extended cognition?
Embodied cognition is the view that the nature of the mind is shaped by the form and dynamics of the body, not just the brain. The standard computational model treats the brain as a central processor receiving inputs from passive sensors and issuing commands to peripheral effectors. Embodied cognition researchers argue this picture is wrong: perception is active and exploratory, reasoning is grounded in bodily simulation and metaphor, and cognition cannot be understood in isolation from the sensorimotor loops that connect an organism to its environment.Francisco Varela, Evan Thompson, and Eleanor Rosch's 'The Embodied Mind' (1991) brought together phenomenology, enactivism, and cognitive science to argue that cognition is enacted through bodily action in the world. George Lakoff and Mark Johnson's 'Philosophy in the Flesh' (1999) extended conceptual metaphor theory -- the idea that abstract thought is structured by bodily metaphors rooted in physical experience -- into a wholesale critique of disembodied AI-style approaches to mind.Andy Clark and David Chalmers extended the critique further in 'The Extended Mind' (1998). They argued that cognitive processes need not be confined to the brain and body; they can extend into the environment. Their thought experiment -- Otto, an Alzheimer's patient who uses a notebook as a memory prosthetic -- is meant to show that external cognitive aids can be genuine components of cognitive processes, not just inputs to them. This 'active externalism' gave rise to the framework of 4E cognition: embodied, embedded, enacted, and extended, a cluster of related positions that challenge the classical computational view.
How do cognitive scientists study memory and attention?
Memory research was transformed by Alan Baddeley and Graham Hitch's 1974 working memory model, which replaced the earlier notion of a single short-term memory store with a multi-component system. The phonological loop holds and rehearses verbal information; the visuospatial sketchpad handles visual and spatial material; the central executive controls attention and coordinates the subsystems; a fourth component, the episodic buffer, was added by Baddeley in 2000 to integrate information across modalities into coherent episodes.Endel Tulving's 1972 distinction between episodic memory (autobiographical memory for personally experienced events, bound to specific times and places) and semantic memory (general world knowledge, detached from personal context) became foundational. The case of Henry Molaison (H.M.), who lost the ability to form new long-term memories after bilateral hippocampectomy in 1953, provided decisive evidence that the hippocampus is essential for converting short-term memories into long-term ones, and that different memory systems can be dissociated. Fergus Craik and Robert Lockhart's levels-of-processing framework (1972) argued that deeper, more elaborative encoding produces better memory, challenging the structural model.Attention research developed through several theoretical frameworks. Donald Broadbent's filter theory (1958) proposed that attention acts as a bottleneck, filtering out unattended information before full processing. Anne Treisman's attenuation model modified this, arguing that unattended channels are weakened rather than fully blocked. Treisman and Garry Gelade's feature integration theory (1980) proposed that simple features such as color and orientation are registered automatically and in parallel, but that binding features into unified objects requires focused attention -- a proposal supported by extensive visual search experiments.
What is the relationship between cognitive science and artificial intelligence?
Cognitive science and artificial intelligence have been intertwined from the beginning, though their relationship has evolved considerably. In the 1950s and 1960s, AI researchers and cognitive scientists often saw their goals as identical: to understand intelligence by building systems that exhibit it. Newell and Simon explicitly built their programs -- the Logic Theorist, the General Problem Solver -- as theories of human problem-solving, generating predictions about human response times and error patterns that could be tested experimentally.The expert systems era of the 1970s and 1980s widened the gap: commercial AI became more interested in practical performance than psychological plausibility. Connectionism brought the fields back together by proposing neural networks as both engineering tools and cognitive models.Modern deep learning, which has produced remarkable achievements including superhuman performance in chess and Go (AlphaGo, 2016) and large language models capable of sophisticated text generation, has reopened fundamental questions. Do large language models constitute genuine cognition or sophisticated statistical pattern matching? The Turing test, proposed by Alan Turing in 1950, suggested that indistinguishable performance from a human should count as intelligence, but critics including John Searle and Gary Marcus have argued that behavioral equivalence does not entail understanding, intentionality, or consciousness. Cognitive scientists continue to debate whether AI models are useful theories of mind, useful engineering tools, or both -- and whether the deep learning revolution requires a fundamental revision of cognitive theory.