Thinking

Abstract

Thinking is a type of mental process, the mental activity that manipulates internal representations to reason, form concepts, solve problems, and choose among options, drawing on stored knowledge rather than on immediate perception. Cognitive psychology studies its representations and its processes: whether some representations are depictive, as the chronometric signatures of mental rotation and image scanning suggest, or uniformly propositional; how concepts are organized around graded prototypes rather than defining features; how human deductive reasoning departs systematically from formal logic; how problems are searched through a space of states and operators; and how judgment and choice diverge from normative standards. A recurring conclusion is that thinking is the joint product of fast, autonomous processes and slow, capacity-limited ones, and that its departures from logical and economic norms are systematic enough to reveal the underlying machinery.

Keywords: thinking, mental representation, categorization, reasoning, dual-process theory

Thinking denotes the internal manipulation of mental representations to reach conclusions, form concepts, and solve problems in the service of a goal, and it was the information-processing analysis of exactly this activity that founded modern cognitive science (Newell & Simon, 1972). Because thinking is not directly observable, its study proceeds by inference from behavior: from the time a mental operation takes, from the errors it reliably produces, and from the way performance shifts when the form of a problem is changed while its logic is held constant. The chapters below trace that inferential method across five domains — mental representation, categorization, reasoning, problem solving, and the dual-process account of judgment — each of which contributed a signature finding to the picture of what thinking is and how it is carried out.

Mental Representation: Images and Propositions

The first question about thinking is what it operates on. A candidate answer that dominated the 1970s is that some thought is carried in depictive representations — internal analogs that preserve the spatial structure of what they represent — as opposed to propositional representations, which encode content in an abstract, language-like format with no intrinsic geometry. The evidence for depiction was chronometric. Shepard and Metzler presented pairs of perspective drawings of three-dimensional block figures and asked whether the two were the same shape or mirror images; the time to decide rose as a strikingly linear function of the angular difference in the figures' orientations (Shepard & Metzler, 1971). The natural interpretation is that participants mentally rotated one figure into congruence with the other at a roughly constant angular rate, so that a larger rotation simply took proportionally longer — a process with the character of a continuous analog operation rather than a discrete symbolic one.

A parallel result came from image scanning. Kosslyn, Ball, and Reiser had participants memorize a map with several landmarks, then form an image of it and scan from one named landmark to another; the time to scan rose linearly with the actual distance between the landmarks on the original map (Kosslyn et al., 1978). Distance in the represented world was preserved as something that took time to traverse in the representation, which is what a depictive medium predicts and a purely propositional one does not obviously require. The demonstration below reproduces the mental-rotation signature: as the angular disparity between two figures is increased, the modeled decision time increases in proportion.

The interpretation did not go unchallenged. Pylyshyn argued that the introspected image is not itself the representation but the product of a more basic propositional code, and that image-like results reflect tacit knowledge — participants unconsciously simulating what would happen in a real perceptual situation — rather than the intrinsic geometry of a mental picture (Pylyshyn, 1973). On this view the linear scanning function appears because people know that farther things take longer to look at and reproduce that expectation, not because anything in the head is literally being traversed. The imagery debate that followed was among the most productive in the field precisely because both sides made the same methodological bet: that the format of a representation, though invisible, leaves measurable traces in the time course of the operations performed on it.

Types of Thinking

Beyond being a subject in its own right, Thinking is a formal category in the National Library of Medicine's Medical Subject Headings, which places it at tree position F02.463.785, beneath Mental Processes, and hangs its recognised narrower kinds beneath it. These subtypes are a classification built to index the literature, not a claim about the mind's natural joints; several are treated at length in their own articles. Table 2 lists the direct children of the descriptor.

Two cautions keep this taxonomy in its place. It is a classification for indexing, built to organise the literature, not a theory asserting these subtypes are mutually exclusive or an exhaustive set of natural kinds. And a MeSH subtype is a narrower topic, not a component process: listing a child under Thinking locates it in an index and says nothing, on its own, about the mechanisms this article describes.

Concepts and Categorization

If thinking manipulates representations, much of what it manipulates are concepts — the mental categories that let an unfamiliar object be treated as an instance of a familiar kind. The classical theory held that a concept is a set of individually necessary and jointly sufficient defining features, so that category membership is all-or-none and every member is an equally good example. Rosch's work overturned this. Across converging measures she showed that natural categories have a graded internal structure: some members are judged more typical than others, and this typicality is not a subjective curiosity but a variable that governs performance (Rosch, 1975). A robin is rated a better example of bird than an ostrich is; the more typical member is named first when people list category members, is verified faster in a category-membership judgment, and is the one to which people generalize most readily.

The organizing construct is the prototype: a summary representation of the category's central tendency, against which candidate members are compared for similarity. Membership is then a matter of degree, set by resemblance to the prototype rather than by satisfaction of a definition, which is why category boundaries are often fuzzy and why the typicality gradient exists at all. Rosch also identified a privileged basic level of abstraction — chair rather than the superordinate furniture or the subordinate armchair — at which categories carry the most information for the least cognitive cost, and which children learn first and adults name fastest. The demonstration lets a category member be selected and reads out its typicality and the verification time a prototype model predicts for it.

Reasoning

Reasoning is thinking that draws conclusions from premises, and the central empirical fact is that human reasoning conforms only loosely to the norms of formal logic. The sharpest demonstration is the selection task. Wason gave participants four cards showing, say, A, K, 4, and 7, told them each card has a letter on one side and a number on the other, and asked which cards must be turned over to test the rule if a card has a vowel on one side, then it has an even number on the other (Wason, 1968). The logically correct choice is the vowel card (A) and the odd-number card (7): the rule is falsified only by a vowel paired with an odd number, so only these two cards can reveal a violation. Most participants select the A alone or the A and the 4 — turning the card named in the rule and the card that matches its consequent — and fail to select the 7. The error is a lawful one: people seek confirming instances and neglect the potentially falsifying case, a bias that formal logic condemns but that dominates performance.

Why reasoning takes this shape is addressed by mental model theory. Johnson-Laird proposed that people reason not by applying formal rules to the syntax of premises but by constructing a mental model of the situation the premises describe, inspecting it for a putative conclusion, and searching for alternative models that would falsify it (Johnson-Laird, 1983). Errors arise because constructing and holding multiple models taxes working memory, so reasoners typically build one model, read a conclusion from it, and stop — omitting the counterexamples that would correct them. The theory explains both the selection-task result and the broader finding that reasoning is easier when the premises describe a concrete, easily modeled state of affairs than when they are abstract, a content effect that a purely syntactic logic cannot accommodate. The clearest case is the thematic selection task: Griggs and Cox recast the rule as a familiar drinking-age regulation — if a person is drinking beer, then the person must be over 19 — with cards showing a beverage on one side and an age on the other, and found that most participants who failed the abstract version now correctly turned the beer and 16-years cards (Griggs & Cox, 1982). The logic is identical to the letters-and-numbers task; only the content differs, which is why the facilitation is evidence for a model-based rather than a rule-based account. The demonstration presents the selection task in both an abstract and a concrete framing so that the same logical structure can be tested under each, exposing the content effect directly.

Problem Solving

Problem solving is goal-directed thinking under conditions where the path to the goal is not immediately known. Newell and Simon formalized it as search through a problem space: a set of states, an initial state, a goal state, and a set of operators that transform one state into another (Newell & Simon, 1972). Because the space of possible states is usually far too large to search exhaustively, solving proceeds by heuristics that keep the search tractable — most notably means-ends analysis, which repeatedly selects the operator that most reduces the difference between the current state and the goal. Figure 1 shows a small problem space of the kind their analysis describes, with the solver moving from an initial state to a goal by applying operators along a path.

What makes many problems hard is not the size of the space but a representation that hides the operators. Duncker documented functional fixedness, the tendency to represent an object only in terms of its conventional function and so to miss a novel use it affords: given a candle, a box of tacks, and matches, and asked to fix the candle to a wall, people are slow to see the tack box as a support rather than a container (Duncker, 1945). The complementary finding concerns analogical transfer. Gick and Holyoak showed that people who had read a story whose solution shared the deep structure of a later problem often failed to apply it, because they encoded the source by its surface content rather than its structure; a transferable solution was retrieved reliably only when the structural parallel was made salient (Gick & Holyoak, 1980). Both results locate the difficulty of problem solving in how the problem is represented, not in the logic of the search once the representation is right.

Dual-Process Theory and the Bounds of Rationality

The findings above have a common shape: an intuitive response arrives quickly and often misleads, while the normative answer requires slower, effortful work. Dual-process theory makes this partition explicit. It distinguishes Type 1 processing — fast, autonomous, high in capacity, and independent of working memory — from Type 2 processing — slow, effortful, and dependent on the limited resource of working memory that also supports hypothetical reasoning (Evans, 2008). Type 1 delivers the immediate impression; Type 2 can, if engaged, monitor and override it. Evans and Stanovich argued that the defining contrast is not speed as such but autonomy and the load on working memory, and that most reasoning errors occur because Type 2 fails to intervene on a Type 1 output rather than because Type 2 reasons badly (Evans & Stanovich, 2013). The same partition was brought to a wide audience under the labels System 1 and System 2 (Kahneman, 2011), though theorists now prefer Type to System to avoid implying two anatomically separate minds. Table 1 sets out the properties conventionally assigned to the two types.

The content of the intuitive channel was charted by the heuristics-and-biases program. Tversky and Kahneman showed that people judge probability by a handful of heuristics — representativeness, judging likelihood by resemblance to a prototype, and availability, judging frequency by the ease with which instances come to mind — which are efficient but generate predictable errors such as base-rate neglect and the conjunction fallacy (Tversky & Kahneman, 1974). The same systematic departure appears in choice. Prospect theory replaced the expected-utility assumption that people weight outcomes by their objective probability with a descriptive account in which value is defined over gains and losses from a reference point, losses loom larger than equivalent gains, and probabilities are transformed nonlinearly, so that logically equivalent problems yield different choices when framed as gains rather than losses (Kahneman & Tversky, 1979).

These departures do not show that people think badly; they show that thinking is bounded. Simon had already argued that rationality is limited by the information available, the cognitive resources of the mind, and the time to decide, so that people do not optimize but satisfice — search until an option is found that meets an acceptable threshold, then stop (Simon, 1955). The heuristics of the intuitive channel are precisely such resource-conserving strategies, well adapted to the ordinary environment even where they fail a laboratory norm. The binding resource is itself narrow: Miller's review placed the capacity of immediate memory at about seven items, a limit circumvented only by chunking information into larger meaningful units (Miller, 1956). Deliberate Type 2 thinking is carried out within that small workspace, which is one reason it is effortful and easily displaced by the faster intuitive response.

Worked Example

The mental-rotation finding can be stated as a quantitative model, and doing so shows why the linear response-time function was read as evidence for an analog process. Let the decision time for a same-different judgment be RT(θ) = a + bθ, where θ is the angular disparity between the two figures in degrees, a is a baseline that collects the time to encode the figures, compare them, and execute the response, and b is the time added per degree of rotation. Taking a = 500 ms and a rotation rate of 50° per second — that is, b = 20 ms per degree, a value within the range Shepard and Metzler reported (Shepard & Metzler, 1971) — the model predicts, for a disparity of 60°:

RT(60) = 500 + 20 × 60 = 500 + 1200 = 1700 ms.

Doubling the disparity to 120° does not double the time, because only the rotation component scales with angle while the baseline is fixed:

RT(120) = 500 + 20 × 120 = 500 + 2400 = 2900 ms.

The difference between them, 2900 − 1700 = 1200 ms, is exactly b × 60, the cost of the extra 60° of rotation, and it equals the cost of the first 60° above baseline. That constancy — equal increments of angle adding equal increments of time — is the signature of a process running at a fixed rate, and the slope b is a direct estimate of the rate of mental rotation: 1000 ⁄ b = 50° per second. A propositional lookup with no analog traversal has no reason to produce a slope at all; it is the linearity, reproduced in the demonstration above, that made the depictive interpretation compelling.

Key Researchers

Jonathan St. B. T. Evans (b. 1948). Emeritus Professor of Cognitive Psychology at the University of Plymouth; he originated the heuristic-analytic theory of reasoning and, with Stanovich, the modern Type 1 / Type 2 formulation of dual-process theory. Faculty Page - Google Scholar - Wikipedia

Philip N. Johnson-Laird (b. 1936). Stuart Professor of Psychology Emeritus at Princeton University; he developed the mental-model theory of reasoning, on which people reason by constructing and inspecting models of the situations premises describe. Faculty Page - ORCID - Google Scholar - Wikipedia

Daniel Kahneman (1934-2024). Eugene Higgins Professor of Psychology Emeritus at Princeton University and 2002 Nobel laureate in economic sciences; with Amos Tversky he founded the heuristics-and-biases program and prospect theory. Faculty Page - Nobel Prize - Wikipedia

Stephen M. Kosslyn (b. 1948). Former John Lindsley Professor of Psychology at Harvard University; his image-scanning experiments and depictive theory anchored the case that some mental representations preserve metric spatial structure. Faculty Page - ORCID - Google Scholar - Wikipedia

Eleanor Rosch (b. 1938). Professor of the Graduate School in psychology at the University of California, Berkeley; she established prototype theory and the graded, basic-level structure of natural concepts. Faculty Page - Wikipedia

Keith E. Stanovich (b. 1950). Emeritus Professor of Applied Psychology and Human Development at the University of Toronto; with Evans he advanced the dual-process synthesis and the study of rational thinking and its dissociation from intelligence. Google Scholar - Wikipedia

Discussion

The five domains surveyed here were, for much of the twentieth century, studied as separate topics, but the modern account joins them. Each supplied a case in which thinking departs from a normative standard in a lawful way, and in each the departure proved more informative than conformity would have been. The linearity of mental rotation revealed the format of a representation; the typicality gradient revealed that concepts are graded; the selection-task error revealed that reasoning is model-based rather than rule-based; functional fixedness revealed that problem difficulty lives in the representation; and the heuristics-and-biases results revealed the intuitive channel by the fingerprints it leaves on judgment. The dual-process framework is the connective tissue: it treats the fast, autonomous process as the source of the systematic first impression and the slow, capacity-limited process as the sometimes-absent corrective.

The framework is not without critics. The strong version of the imagery debate was never fully resolved, and the mapping of behavior onto two discrete systems has been challenged as a convenient dichotomy imposed on a continuum of processing. Even so, the reframing has been productive: it recast heuristics not as failures of a single reasoning faculty but as the ordinary output of an autonomous system that a second system may or may not check, and it connected laboratory reasoning to Simon's older argument that rationality is bounded by the mind's real resources. The durable lesson is methodological. Thinking cannot be observed, but it can be measured by its regularities — the times its operations take and the errors they reliably commit — and it is those regularities, not introspective report, that have carried the field.

Glossary

Analogical transfer.

Applying the structure of a previously solved problem to a new one; it succeeds only when the deep structure, not the surface content, is encoded and retrieved.

Availability heuristic.

Judging the frequency or probability of an event by the ease with which instances of it come to mind.

Basic-level category.

The level of abstraction (e.g., chair) that carries the most information for the least cognitive effort, learned first and named fastest.

Bounded rationality.

Simon's principle that real choice is limited by available information, cognitive resources, and time, so people satisfice rather than optimize.

Chunk.

A unit of meaningful information in immediate memory; grouping items into chunks raises the amount that can be held within the capacity limit.

Deductive reasoning.

Drawing a conclusion that follows necessarily from given premises; the norm against which the selection-task error is measured.

Dual-process theory.

The account that thinking arises from fast, autonomous Type 1 processes and slow, working-memory-dependent Type 2 processes.

Functional fixedness.

The tendency to represent an object only by its conventional function, obstructing a novel use it could serve.

Heuristic.

A quick, resource-conserving strategy that usually yields a workable answer but can produce systematic error.

Mental model.

An internal representation of the situation premises describe, which reasoners inspect for a conclusion and search for counterexamples.

Mental rotation.

The imagined turning of a represented object, whose time cost rises linearly with the angle of rotation.

Problem space.

The set of states, an initial and a goal state, and the operators linking them, through which problem solving searches.

Propositional representation.

An abstract, language-like encoding of content with no intrinsic spatial geometry, contrasted with a depictive image.

Prototype.

A summary representation of a category's central tendency; membership is judged by similarity to it, yielding graded typicality.

Representativeness heuristic.

Judging probability by resemblance to a prototype, which neglects base rates and can produce the conjunction fallacy.

Type 1 processing.

Fast, autonomous mental processing that does not draw on working memory and delivers the immediate intuitive response.

Type 2 processing.

Slow, effortful mental processing that depends on limited working memory and can monitor or override a Type 1 output.

Typicality effect.

The faster and more accurate verification of typical category members than atypical ones, evidence of prototype structure.

Frequently Asked Questions

What is thinking in cognitive psychology?
It is the internal manipulation of mental representations to reason, form concepts, solve problems, and choose, analyzed since the 1970s as information processing through a space of states and operations (Newell & Simon, 1972).

What is the mental rotation effect?
It is the finding that the time to judge whether two figures are the same shape rises linearly with the angular difference between their orientations, as though one figure is turned into alignment at a constant rate (Shepard & Metzler, 1971).

Are mental images like pictures in the head?
The depictive view holds that some images preserve metric spatial structure, supported by scanning times that rise with represented distance, though a propositional critique attributes such results to tacit knowledge rather than to a literal picture (Kosslyn et al., 1978).

What is a prototype?
A prototype is a summary representation of a category's central tendency; membership is judged by similarity to it, so categories are graded and typical members are recognized faster than atypical ones (Rosch, 1975).

Why do people fail the Wason selection task?
Most people turn the cards that match the rule and neglect the card that could falsify it, showing a confirmation bias that departs from the requirements of deductive logic (Wason, 1968).

What is the difference between Type 1 and Type 2 thinking?
Type 1 processing is fast and autonomous and does not use working memory, whereas Type 2 processing is slow, effortful, and dependent on limited working memory, and can override a Type 1 response (Evans & Stanovich, 2013).

What is bounded rationality?
It is Simon's principle that choice is constrained by available information, cognitive capacity, and time, so people satisfice by accepting a good-enough option rather than optimizing over all alternatives (Simon, 1955).

What is functional fixedness?
It is the tendency to represent an object only by its usual function, which blocks solutions that require using the object in an unconventional way (Duncker, 1945).

References