Human Cognition
Human cognition is the set of mental processes by which people perceive, attend, remember, reason, decide and use language, studied experimentally by psychology and related sciences as an information-handling system with measurable limits. Where artificial intelligence asks how to build a system that does a task, the study of human cognition asks how the one existing example of general intelligence actually works, including where it fails. Its findings supply both the targets that AI has tried to reproduce and the contrasts that show how different engineered systems are.
Perception is not passive recording but active interpretation. Gestalt psychologists in the early twentieth century showed that the mind organizes input into wholes, and David Marr's 1982 book Vision framed the problem as computation: from a two-dimensional image to a description of a three-dimensional scene, through stages of increasing abstraction. Illusions are the evidence, because the visual system relies on assumptions that usually hold and sometimes mislead. The lesson for machine vision was both encouraging and humbling, since tasks that people do effortlessly, like recognizing a face at an odd angle, proved much harder to program than chess, a pattern later summarized as Moravec's paradox.
Attention is a limited resource that selects what is processed deeply. Colin Cherry's 1953 experiments on the cocktail party problem showed that listeners can follow one voice among many, and Donald Broadbent's 1958 book Perception and Communication proposed a filter that passes only part of the input. Later work on inattentional blindness, such as the 1999 study by Daniel Simons and Christopher Chabris in which many viewers failed to notice a person in a gorilla suit while counting basketball passes, showed that unattended events can be missed entirely. Engineering borrowed the word in 2017 with the transformer's attention mechanism, but that is a weighting operation over a sequence, not a model of limited human attention.
Memory is several systems rather than one store. Hermann Ebbinghaus in 1885 measured forgetting with nonsense syllables and found a steep early decline followed by a slower one. Richard Atkinson and Richard Shiffrin's 1968 model described sensory, short-term and long-term stores, and Alan Baddeley and Graham Hitch in 1974 replaced the single short-term store with a working memory that has separate components for verbal and visual material under a central executive. George Miller's 1956 paper on the magical number seven, plus or minus two, popularized the limit on how many items can be held at once, though later work suggests that the limit depends on how items are chunked. Human memory is reconstructive, so recall can be vivid and wrong, a feature that resembles fabrication in generative models only superficially.
Reasoning research found that people are neither ideal logicians nor random guessers. In Peter Wason's 1966 selection task, most participants fail an abstract if-then problem that they solve easily when it is framed in familiar social terms. Amos Tversky and Daniel Kahneman's 1974 Science paper described heuristics such as availability and anchoring that produce systematic biases. Kahneman's 2011 book Thinking, Fast and Slow popularized a two-system picture, using labels from Keith Stanovich and Richard West: fast, automatic, intuitive processes called system 1, and slow, effortful, deliberate ones called system 2. The framework is a useful metaphor, but psychologists treat it as a simplification rather than two literal brain modules, and commentators sometimes map it onto fast pattern-matching versus step-by-step reasoning in language models.
The relation to AI has run in both directions. Newell and Simon built programs as theories of human problem solving, cognitive science grew around the information-processing view, and neural-network models inspired by the brain were used to explain learning and reading. At the same time, AI systems clarified what the mind must be doing by showing what is hard: common sense, commonsense physics, and flexible transfer. Systems that match or exceed people on narrow tests are not thereby models of human cognition, because the route to the answer can be entirely different. Chess engines search millions of positions per second, while strong human players rely on pattern recognition and selective search.
Common misconceptions distort comparisons. The brain-as-computer analogy is a modeling choice, not a claim that brains run stored programs; human thought is embodied, emotional and social, which standard AI benchmarks omit. People are not simply irrational either: many biases look sensible given limited time and information, a view associated with Gerd Gigerenzer's work on fast and frugal heuristics. And the human mind is not a fixed yardstick, since performance varies with expertise, fatigue and framing, which makes claims like human-level performance dependent on which people and which conditions were used.
Read human cognition as the empirical benchmark and the cautionary mirror for AI: it shows what flexible intelligence must achieve and how bounded, biased and context-bound a working mind actually is.