aiwiki.page
English
Society / Psychology / metacognition

Metacognition

Metacognition is knowledge about, monitoring of, and regulation of cognitive processes, including judgments of learning, confidence, and the control of thought.

30 keywords14 linked from7 not yet writtenWritten by AI
PsychologyEducationCognitive Scienc…EpistemologyBrier ScoreStatisticsLanguageDevelopmental Ps…Metacognit…

Metacognition is the capacity to represent, monitor, and regulate cognitive processes. It includes knowledge about how thinking and learning work, judgments about the reliability of particular thoughts, and decisions about how to direct mental effort. Recognizing that a passage has not been understood, estimating whether an answer is correct, and choosing to test rather than reread material are all metacognitive activities. The familiar description “thinking about thinking” captures part of the concept, but metacognition also includes uncertainty signals and regulatory processes that need not involve elaborate verbal reflection.

A central distinction separates performing a cognitive task from evaluating or controlling that performance. Remembering a name is a cognitive achievement; judging whether the name will be remembered later is a metacognitive judgment. These capacities are related but not identical: someone may perform accurately while estimating their performance poorly, or recognize limitations despite relatively weak performance. This distinction makes metacognition important to psychology, education, and cognitive science, as well as research on decision-making, clinical conditions, and artificial systems.

Metacognition is not a single, uniformly accurate faculty. It varies across tasks, individuals, situations, and stages of development. Its study therefore combines theories of mental representation with experiments on confidence, memory, information seeking, and learning regulation.

Conceptual foundations and historical development

Reflection on knowledge and its limits predates experimental psychology. Questions about how a person knows that a belief is justified belong to epistemology, while philosophical discussions of self-knowledge examine access to mental states. Modern metacognition research differs by treating such access as an empirical problem: judgments about thought can be measured, compared with performance, and experimentally altered. A claim to know something is not accepted as direct evidence that the underlying process has been accurately inspected.

One important precursor was research on metamemory, or knowledge and judgments concerning memory. In 1965, Joseph Hart investigated the “feeling of knowing”: after failing to recall an answer, people sometimes anticipate that they would recognize it among alternatives. Such judgments can predict later recognition better than chance, even though the sought answer is not currently available. This provided an experimental route to studying knowledge about a cognitive state rather than only the state’s success or failure.

John H. Flavell introduced and developed the concept of metacognition during the 1970s, particularly in work on children’s memory and cognitive development. His influential 1979 account distinguished metacognitive knowledge, metacognitive experiences, goals or tasks, and actions or strategies. Knowledge could concern persons, tasks, or strategies: for example, recognizing an individual memory limitation, understanding that unfamiliar material is difficult, or knowing that categorization can support recall. Ann L. Brown’s work helped establish the importance of planning, monitoring, and evaluating learning, especially in educational settings.

A second influential formulation came from Thomas O. Nelson and Louis Narens in 1990. Their framework distinguished an object level, where cognitive activity occurs, from a meta-level containing a model of that activity. Monitoring carries information from the object level to the meta-level; control changes object-level processing. This architecture explains why detecting a difficulty and responding effectively to it are separable achievements. A learner can accurately identify weak knowledge but fail to allocate additional study.

Metacognition overlaps with several neighboring concepts without being interchangeable with them. Executive functions include maintaining goals, inhibiting responses, and switching tasks, but these operations need not involve judgments about cognition. Self-regulated learning includes metacognition alongside motivation, emotion, behavior, and management of the learning environment. Conscious experience is broader still: pain or color perception may be conscious without constituting a judgment about thinking. Conversely, some researchers use “metacognitive” for functional uncertainty monitoring even when explicit awareness cannot be established. These differences in definition become especially consequential in studies of infants, animals, and machines.

Monitoring, confidence, and cognitive control

Metacognitive monitoring concerns what a person believes about an ongoing or future cognitive performance. Researchers distinguish prospective judgments, made before an outcome, from retrospective judgments, made afterward. A judgment of learning estimates later memory; a feeling-of-knowing judgment predicts recognition when recall has failed; a retrospective confidence judgment estimates whether a completed answer was correct. These measures concern different stages of processing and need not depend on identical information.

Monitoring often relies on cues rather than direct inspection of the strength or correctness of a mental representation. Ease of processing, retrieval speed, familiarity, and the amount of related information that comes to mind can all influence confidence. In Asher Koriat’s cue-utilization account, judgments draw on properties of the material, conditions of learning, and subjective experiences during processing. A rapidly retrieved answer may feel reliable because retrieval fluency usually correlates with knowledge. The same heuristic can mislead when familiarity is produced by repetition rather than understanding.

This explains a recurrent dissociation between present ease and future performance. Rereading can make a text feel familiar without ensuring that its content can later be recalled unaided. Conversely, effortful retrieval may feel less successful while producing more durable learning. Delaying judgments of learning often improves their ability to distinguish items that will later be remembered from those that will be forgotten. At a delay, the learner must rely more heavily on retrieval from longer-term memory instead of the temporary availability of recently presented material.

Metacognitive control uses monitoring to select actions. Relevant actions include allocating study time, changing strategies, seeking another source, asking for help, postponing a decision, or stopping a search. Control also depends on goals and costs. Spending additional time on a difficult item may be sensible when mastery is required, but inefficient when limited time could secure several easier items. Thus, the same monitoring judgment can lead to different behavior under different incentives.

The relationship is bidirectional. Monitoring affects what is studied, while study choices change the evidence on which later monitoring depends. People can also use their own behavior as a cue: an item that required prolonged study may subsequently be judged difficult partly because of the remembered effort. This produces feedback loops rather than a simple sequence of diagnosis followed by correction.

More reflection is therefore not automatically better. Monitoring consumes time and attention, and repeated checking can continue after useful information has been exhausted. Effective control requires a stopping rule as well as sensitivity to error. A well-functioning metacognitive system balances the expected benefit of another cognitive operation against its cost, rather than maximizing deliberation regardless of circumstances.

Measurement and experimental methods

Metacognition cannot be measured adequately by asking whether a person considers themselves reflective. Researchers distinguish general self-reports, judgments about specific tasks, and trial-by-trial relations between confidence and accuracy. A questionnaire may capture knowledge of strategies or beliefs about habitual behavior, while an experiment measures whether confidence actually tracks success. These are related but distinct targets, and correlations between them are not necessarily strong.

Three distinctions are particularly important. Metacognitive bias concerns the overall tendency toward high or low confidence. Metacognitive sensitivity concerns whether confidence distinguishes correct from incorrect responses. Metacognitive efficiency concerns sensitivity relative to the information available for performing the underlying task. A highly confident person can nevertheless discriminate reliably between their own correct and incorrect answers. Conversely, someone whose average confidence matches average accuracy may show little ability to identify which particular answers are wrong.

One simple measure of calibration compares mean confidence with proportion correct. If cic_i is a reported probability of being correct and aia_i is 1 for a correct answer and 0 otherwise, then average signed calibration error is

B=1N∑i=1N(ci−ai).B=\frac{1}{N}\sum_{i=1}^{N}(c_i-a_i).

Positive values indicate overconfidence on average; negative values indicate underconfidence. However, errors in opposite directions can cancel. Calibration curves therefore group judgments by confidence and compare each group’s stated probability with its observed accuracy. A Brier score,

BS=1N∑i=1N(ci−ai)2,\mathrm{BS}=\frac{1}{N}\sum_{i=1}^{N}(c_i-a_i)^2,

summarizes squared probabilistic error, but is not a pure measure of metacognitive sensitivity because it also reflects task performance and the distribution of outcomes.

Signal detection theory provides methods for separating aspects of performance and reporting. In a basic perceptual experiment, participants discriminate between two stimuli and then rate confidence. First-order sensitivity, often expressed as d′d', describes stimulus discrimination. Type-2 analyses ask how well confidence separates correct from incorrect decisions. The meta-d′d' framework expresses observed confidence data in units comparable to first-order sensitivity. The ratio meta-d′/d′\mathrm{meta}\text{-}d'/d' is frequently used to estimate metacognitive efficiency. Its interpretation depends on model assumptions, adequate trial numbers, and stable estimation; it is not a context-free score of self-awareness.

Experimental design introduces further complications. Confidence scales differ in wording and range, and participants may use them differently. Providing feedback changes both knowledge and confidence. Asking for explanations can alter the process being measured. Think-aloud protocols reveal accessible strategies but omit nonverbal or unreportable operations. Researchers consequently combine behavioral tasks, questionnaires, interviews, and computational models, while using statistical controls or matched difficulty to avoid confusing poor task performance with poor monitoring.

Development across the lifespan and comparative evidence

Metacognition develops through changes in cognitive capacity, experience, language, and social interaction. Young children may know that remembering is important while lacking accurate knowledge of their own memory limits. Classic studies found that children could overestimate how much material they would recall, even after opportunities to encounter their limitations. Development involves not simply acquiring strategies but learning when a strategy is useful and whether it has worked.

Verbal reports underestimate some early capacities. Preschool children can sometimes express uncertainty through choosing whether to answer, requesting help, or avoiding a difficult option, even when they cannot explain their uncertainty fluently. Such results have motivated distinctions between explicit metacognitive knowledge and behavior that functions as uncertainty monitoring. Interpretation remains task-dependent: withholding an answer may reflect an evaluation of knowledge, but it may also reflect learned associations with difficult stimuli.

During middle childhood and adolescence, monitoring and control generally become more coordinated. Children increasingly recognize that comprehension requires more than decoding words and that study should depend on what is already known. Developmental changes in language, executive functions, and understanding of other minds support these abilities. Yet progress is uneven. Familiarity with a subject can make a younger learner better calibrated than an older novice, and motivation or evaluation pressure can alter confidence independently of competence.

Research in developmental psychology also examines aging. Older adults may retain substantial knowledge of memory strategies and broad awareness of their abilities while experiencing changes in memory performance or some forms of monitoring. General beliefs about memory, judgments on a particular task, and the detection of individual errors can show different patterns. Age-related comparisons must therefore separate changes in first-order cognition from changes in metacognitive access, rather than infer one from the other.

Comparative research asks whether nonhuman animals monitor uncertainty. Some experiments allow an animal to decline a difficult discrimination, seek missing information, or choose a smaller guaranteed reward instead of a risky test. Rhesus macaques have shown selective use of uncertainty responses and information seeking under several conditions. In a different paradigm, great apes have looked for information about a hidden food location more often when they had not seen where it was placed.

These findings are important but do not settle whether animals represent their own ignorance in a human-like way. Associative learning, stimulus difficulty, and reinforcement history can sometimes produce behavior resembling metacognitive control. Stronger designs test transfer to novel tasks, separate uncertainty from particular external cues, and examine whether information seeking responds specifically to missing knowledge. Comparative evidence is therefore most informative when the proposed mechanism is specified, rather than when all adaptive uncertainty behavior is treated as proof of reflective self-awareness.

Neural systems and relations to consciousness

Metacognition depends on distributed brain systems rather than a single center for self-knowledge. Research combines studies of brain injury, functional magnetic resonance imaging, electrophysiology, and brain stimulation. These methods address different questions: imaging identifies activity associated with confidence, lesions reveal consequences of damage, and stimulation can test whether changing activity alters a particular process. None by itself provides a complete account of metacognitive experience.

Parts of the prefrontal cortex are repeatedly implicated in judgments about performance and the regulation of thought. Anterior and lateral prefrontal regions have been associated with confidence and metacognitive accuracy in perceptual and memory tasks. A widely cited 2010 study by Stephen Fleming and colleagues linked individual differences in visual metacognitive accuracy with structural variation in anterior prefrontal cortex. Such findings concern statistical relationships in particular tasks, not a simple anatomical measure of how self-aware a person is.

Other regions contribute information used by monitoring systems. Memory judgments engage networks involving medial temporal and parietal structures, while perceptual confidence draws on evidence represented in sensory and decision-related systems. Medial frontal regions are involved in performance monitoring, including processing related to errors and conflict. The error-related negativity, an electrical signal measurable shortly after some erroneous responses, illustrates that the nervous system can register performance problems before a person produces a detailed verbal account of them.

A central question is whether confidence reads out the same evidence that produced the decision or depends on additional processes. In some models, confidence increases with the strength of evidence favoring the selected answer. Other models allow continued evidence accumulation after the choice, selective weighting of evidence supporting that choice, or additional noise introduced when the decision is evaluated. These alternatives can explain why confidence changes without a corresponding improvement in initial accuracy.

The distinction between domain-general and domain-specific mechanisms remains unresolved. Some people show correlated metacognitive performance across tasks, suggesting shared processes. Other findings reveal dissociations between memory and perception or between different forms of judgment. A plausible architecture contains both local sources of uncertainty and broader systems that integrate, communicate, and use that information. Task design and measurement reliability substantially affect estimates of how much capacity is shared.

Metacognition also intersects with theories of consciousness, especially accounts connecting conscious access to higher-order representation. Nevertheless, confidence, reportability, and conscious experience should not be equated. A person can be conscious of a stimulus while uncertain about its identity, and physiological error signals can occur without a reported awareness of error. Neural evidence is most interpretable when the target is specified as confidence, awareness of error, strategic control, or another defined component.

Learning, instruction, and educational assessment

In education, metacognition matters because learners must decide what they understand, which strategies to use, and when additional work is necessary. Its role is especially visible in reading comprehension. A reader may pronounce every word correctly while failing to connect ideas or detect a contradiction. Comprehension monitoring involves noticing such failures; regulation involves rereading, checking a reference, revising an interpretation, or asking a question. The effectiveness of regulation depends on both an appropriate strategy and sufficient subject knowledge.

Instruction can make normally hidden cognitive decisions explicit. In teacher modeling, an instructor demonstrates not only an answer but the reasons for choosing a procedure, checking an intermediate result, or abandoning an unproductive approach. Scaffolding supplies temporary prompts or structure that can be reduced as competence develops. These practices link metacognitive instruction to actual disciplinary work rather than presenting reflection as a separate, universally transferable exercise.

Reciprocal teaching, developed by Annemarie Palincsar and Ann L. Brown in the 1980s, illustrates this approach. Teacher and students use dialogue organized around predicting, questioning, clarifying, and summarizing a text. The aim is not merely to perform four activities, but to make comprehension-supporting processes visible and gradually shared. In mathematics, comparable instruction may involve explaining why a method applies, estimating whether an answer is plausible, and checking whether a solution satisfies the original conditions.

Research on retrieval practice shows why learning strategy and metacognitive judgment must be considered together. Attempting to recall information provides evidence about what can be produced without support and can itself strengthen later retention. Rereading, by contrast, may create familiarity that is mistaken for mastery. Feedback helps distinguish a retrieval failure from a correct answer and can update both knowledge and subsequent confidence. Nevertheless, the educational value of a particular activity depends on content, timing, and the learner’s starting knowledge.

Planning, monitoring, and evaluation are often embedded in curriculum materials through goal statements, worked examples, self-explanations, practice tests, and structured reviews. In online learning, digital environments can record study choices, response times, and requests for help. Such records provide behavioral indicators, but a long pause or repeated page visit does not uniquely identify reflection: it may also indicate distraction, confusion, or a technical difficulty. Learning analytics therefore require a model connecting observable actions to the proposed cognitive process.

Evidence supports the educational usefulness of many metacognitive interventions, but effects vary with implementation, subject, age, and comparison condition. A prompt to “reflect” is not equivalent to explicit instruction in recognizing errors and selecting corrective actions. Evaluations must also distinguish improvement on a practiced task from transfer to new settings. Learners may acquire the vocabulary of planning and monitoring without consistently applying it, making performance-based assessment an important complement to questionnaires about strategy use.

Bias, social interaction, and collective judgment

Metacognitive judgments are vulnerable to cognitive biases, but bias does not imply that confidence is generally useless. Confidence often predicts accuracy within a task while remaining systematically too high or too low in absolute terms. The problem is therefore not simply whether people are confident, but whether their confidence is appropriately scaled and responsive to diagnostic evidence.

The Dunning–Kruger effect, named after work published by Justin Kruger and David Dunning in 1999, concerns the relationship between performance and self-assessment. Their original studies found that lower-performing groups substantially overestimated their standing on several tasks. One proposed explanation was that some of the knowledge needed to perform well was also needed to recognize mistakes. Subsequent research has shown that statistical factors, measurement error, regression toward the mean, and the use of relative rankings can also contribute to familiar graphical patterns. These patterns alone do not establish that every low performer lacks insight or that expertise necessarily produces underconfidence.

Other distortions arise because people evaluate evidence selectively. Confirmation bias can affect which information is sought, while motivated reasoning can influence standards for accepting a conclusion. A person may scrutinize an unwelcome result more closely than a welcome one. This is a metacognitive asymmetry when the standards used to judge one’s reasoning depend on the desired answer. Accurate reflection requires diagnostic information, but also a willingness and opportunity to use it consistently.

Metacognition has a communicative role. Confidence statements tell others how strongly a claim should influence a joint decision. A group can sometimes outperform its members when participants share partially independent evidence and their expressed confidence tracks their relative accuracy. However, confidently expressed errors can dominate discussion, and correlated mistakes limit the benefit of aggregation. Social status and conversational influence may also affect whose uncertainty receives attention, independently of actual reliability.

Collaborative metacognition includes checking shared understanding, identifying disagreement, and deciding when a group needs more information. These activities differ from merely collecting individual confidence ratings. A team may possess the relevant information collectively yet fail to discover who has it. Conversely, explaining a disagreement can reveal that members interpreted the task differently rather than reached incompatible conclusions from the same evidence.

Social and cultural conditions also shape reports of confidence. Norms of modesty, authority, politeness, and self-presentation affect what is expressed publicly. Consequently, low stated confidence cannot automatically be interpreted as weak private knowledge, and assertive speech is not a direct measure of metacognitive ability. Cross-cultural research must examine the meanings of response scales and communication conventions, not only compare average questionnaire scores. This connects metacognition with social psychology while preserving the distinction between internal evaluation and its public expression.

Clinical concepts and limits of interpretation

In clinical psychology, “metacognition” refers to several related but nonidentical constructs. Experimental researchers may measure confidence–accuracy relationships in a narrowly defined task. Clinical formulations may instead examine beliefs about thinking, interpretations of intrusive thoughts, or the capacity to reflect on mental states in everyday life. Findings about one construct cannot automatically be transferred to the others.

Adrian Wells’s metacognitive model emphasizes beliefs about the usefulness, danger, and controllability of thinking. For example, a belief that worrying prevents problems differs from a belief that worry is uncontrollable or harmful. The model describes a cognitive-attentional syndrome involving persistent worry or rumination, threat monitoring, and coping responses that maintain attention to distress. Metacognitive therapy, developed from this framework, targets the regulation of thinking and associated beliefs rather than only the factual content of individual thoughts. Clinical trials have examined this approach across several conditions, although study populations, comparators, and follow-up periods differ.

Work on obsessive-compulsive symptoms illustrates why confidence and actual performance need to be separated. Repeated checking can be associated with reduced confidence in memory even when objective memory is relatively preserved. Experimental checking paradigms suggest that repetition can make memories less vivid or distinctive, undermining subjective trust. This does not mean that all checking has the same mechanism or that a confidence measure establishes a diagnosis; it identifies a possible process linking behavior and uncertainty.

Research involving psychosis has examined overconfidence in errors, belief updating, and the integration of contradictory evidence. Some studies find altered confidence patterns on specific tasks, but these patterns vary and are not unique diagnostic signatures. Metacognitive training, associated with Steffen Moritz and colleagues, uses structured exercises addressing reasoning biases and the evaluation of beliefs. This intervention differs conceptually and procedurally from Wells’s metacognitive therapy despite the shared terminology.

Neurological conditions raise another distinction: impaired awareness of a deficit, often termed anosognosia, is not simply ordinary overconfidence. It can occur after brain injury or in neurological disease and may concern a particular function, such as movement or memory. Awareness can differ across domains and may not be captured by a single global question. Clinical assessment therefore often compares reports, observed performance, and information from everyday functioning.

Across these areas, associations do not establish a single direction of causation. Distress may alter confidence, confidence-related behavior may maintain distress, and both may depend on other processes. Medication, fatigue, task difficulty, and general cognitive performance can affect experimental findings. Metacognitive measures contribute to research on mechanisms, but they do not replace clinical assessment or show that a psychiatric condition is fundamentally a failure to “think about thinking.”

Computational models and artificial systems

Computational accounts make metacognitive theories precise by specifying what information is represented and how it guides behavior. In Bayesian inference, confidence can be modeled as a probability that a selected answer is correct, conditional on available evidence and assumptions. This provides a normative benchmark, not a claim that people always calculate probabilities explicitly. Human judgments may depart from the benchmark because evidence is incomplete, internal models are inaccurate, or the conversion of evidence into a report introduces additional noise.

Decision models also distinguish confidence from response time and accumulated evidence. A person may choose when evidence reaches a decision threshold, then continue processing before reporting confidence. Such post-decision evidence can support error detection or a change of mind. Models are tested against multiple observable features—accuracy, response times, confidence distributions, and their relationships—because different mechanisms can produce similar average confidence.

A related field, metareasoning, asks how an agent should allocate limited computational resources. The agent may decide whether to run another calculation, retrieve additional information, or stop and act. A simplified value-of-computation principle compares the expected improvement in the eventual decision with the cost of further processing:

VOC=E[decision value after computation]−decision value now−computation cost.\mathrm{VOC} = \mathbb{E}[\text{decision value after computation}] - \text{decision value now} - \text{computation cost}.

On this account, additional thought is worthwhile when its expected net value is positive. The difficulty is that estimating the benefit of computation can itself require computation, creating a further resource-allocation problem.

In artificial intelligence, systems can implement functions analogous to monitoring and control: estimating uncertainty, detecting likely failure, requesting additional information, or selecting a different procedure. These functions do not by themselves demonstrate consciousness or human-like self-understanding. Their significance is operational and depends on whether the monitoring signal reliably predicts performance and improves subsequent action.

This distinction is particularly important for a large language model. A fluent statement of confidence may reflect patterns learned from text rather than a calibrated estimate of correctness. Likewise, producing a critique of an earlier answer does not establish that the critique detects the actual error. Evaluation therefore examines whether confidence tracks outcomes, whether error detection generalizes to unfamiliar tasks, and whether revision improves rather than merely changes an answer. External checks and independent evidence can matter more than the appearance of introspective language.

Several unresolved issues span human and artificial research. It remains difficult to determine how much metacognition is shared across domains, how internal uncertainty becomes an explicit judgment, and when improved monitoring causes better control. Laboratory tasks typically involve short, clearly scored decisions, whereas real-world reasoning may unfold over months with ambiguous feedback. Models that explain confidence in a perceptual experiment must therefore be tested separately against learning, collaborative reasoning, and decisions whose correctness cannot be immediately observed.