Metacognition

Abstract

Metacognition is cognition about cognition: the monitoring and control a person exercises over their own mental processes. This article follows Flavell's coinage and the Nelson-Narens meta-level and object-level framework into metamemory, covering the judgments of learning and feelings of knowing that estimate one's own memory, the cue-utilization mechanism that makes them inferential, and the study decisions they drive. It separates two things a monitoring judgment can be good at, calibration and resolution, reviews the signal-detection measure meta-d-prime, and ties metacognitive accuracy to the anterior prefrontal cortex. Whether a single metacognitive faculty serves memory, perception, and decision alike, or monitoring is domain-specific, remains contested. Three demonstrations allocate study time across a difficulty gradient, compute the gamma correlation that scores monitoring resolution, and drive a calibration curve into the overconfidence that afflicts poor performers.

Keywords: metacognition, metamemory, monitoring and control

Metacognition is the part of cognition that takes the rest of cognition as its object. Where memory stores and retrieval recovers, metacognition asks whether the storage will hold and whether the recovery can be trusted; where a problem is being solved, it judges how close the solution is and whether the current strategy is working. John Flavell, who named the field, defined it as knowledge and cognition about cognitive phenomena, and split it into a stable store of metacognitive knowledge about how minds work and the on-line metacognitive experiences that arise during a task (Flavell, 1979). The scientific traction comes from treating the second component operationally: a metacognitive judgment is a prediction the person makes about their own cognition, and because the predicted event can be measured, the judgment can be scored against it. The sections below build the concept from that core. They lay out the meta-level and object-level architecture, examine the monitoring judgments of metamemory and the inferential machinery behind them, follow those judgments into the control of study, distinguish the two ways a judgment can be accurate, and close on the prefrontal basis of the whole system.

What Metacognition Is

The term metacognition entered psychology through developmental work on memory, where Flavell needed a name for what children come to know about their own remembering. He defined it broadly, as any knowledge or cognition that takes a cognitive object, and organized it into four classes: metacognitive knowledge, the person's beliefs about how cognitive variables of person, task, and strategy interact; metacognitive experiences, the conscious feelings and estimates that accompany an ongoing task; goals; and the strategies invoked to meet them (Flavell, 1979). The distinction that has carried the most weight is the first against the second. Metacognitive knowledge is semantic and stable, the general understanding that a list is harder to learn than a story or that rehearsal aids retention; it can be wrong, and much of cognitive development consists in correcting it. Metacognitive experience is episodic and online, the momentary sense that a name is on the tip of the tongue or that a passage has been understood, and it is these experiences that drive behavior in the moment.

The value of the concept is that it picks out a level of processing that is causally distinct from the cognition it monitors. A person can retrieve an answer and, separately, judge whether the answer is right; the two can dissociate, so that retrieval succeeds while the judgment of it fails, or a confident judgment attaches to an error. Because the monitoring judgment can be measured against the very performance it predicts, metacognition is among the few aspects of subjective experience that admit a rigorous accuracy criterion, and that is what has made metamemory, the metacognition of memory, its most developed province.

Monitoring and Control

The framework that organized the field is Thomas Nelson and Louis Narens's. They proposed that cognitive processes split into two interrelated levels, an object-level that does the first-order work of perceiving, remembering, and problem-solving, and a meta-level that holds a dynamic model of the object-level (Nelson & Narens, 1990). The two levels are joined by two directed relations, and keeping them distinct is the whole point of the scheme. Monitoring is the flow of information upward, from object-level to meta-level: the meta-level is informed of the state of the object-level, and its model is updated accordingly. Control is the flow of command downward, from meta-level to object-level: the meta-level modifies the object-level, initiating, continuing, or terminating a process. Figure 1 renders the two levels and the two flows.

The power of the distinction is that it separates two questions usually run together. Whether a person knows the state of their own cognition is a monitoring question, answered by the accuracy of their judgments; what they do about it is a control question, answered by their allocation of effort. The two are linked but not identical: monitoring can be accurate while control ignores it, or control can be sensible while resting on a poor monitor. Nelson and Narens's insistence that control decisions are driven by the output of monitoring, so that the value of good monitoring lies in the better control it enables, set the research agenda for a generation and gives the interactive demonstrations below their shape, each isolating one flow so it can be measured on its own.

Metamemory Monitoring

The judgments through which the meta-level reads the object-level are the observable core of metamemory, and they are timed to different phases of learning. Before study, an ease-of-learning judgment predicts how hard an item will be to acquire. During or after study, a judgment of learning predicts whether an item just studied will later be recalled. At test, when recall fails, a feeling-of-knowing judgment predicts whether the unrecalled item would nonetheless be recognized. Josef Hart established the last as a real, measurable competence: people asked whether they would recognize an answer they could not recall predicted their own later recognition well above chance, showing that memory carries information about its own contents that is available even when retrieval fails (Hart, 1965).

The decisive theoretical question is how such judgments are made, and the intuitive answer, that the person inspects the memory trace and reports its strength, is wrong. Asher Koriat's cue-utilization account holds that judgments of learning are inferential: the learner has no direct access to the trace and instead infers its future accessibility from cues available at the moment of judgment, chief among them the fluency with which the item is processed or retrieved (Koriat, 1997). The account explains both the successes and the failures of monitoring in one stroke. When the cues are diagnostic of later memory, the judgments are accurate; when a cue is salient but non-diagnostic, such as the perceptual fluency of easy-to-read text that does not survive to test, the judgment follows the cue and is confidently wrong. Monitoring is only as good as the validity of the cues it reads, which is why its accuracy is an empirical variable rather than a given.

Controlling Study

Monitoring earns its keep by steering control, and the clearest case is the allocation of study time. The commonsense rule is to study the hardest material longest, and learners often do, but the rule is not optimal under a deadline and does not describe skilled self-regulated study. Janet Metcalfe and Nate Kornell proposed instead that learners allocate time to a region of proximal learning, the band of items that are not yet known but are close to being learned, and shift away from an item once its rate of return on further study falls, whether because it is already mastered or because it is currently too hard to move (Metcalfe & Kornell, 2005). Effort follows the gradient of learning rate rather than the gradient of difficulty, and under time pressure the hardest items are rationally abandoned. The first demonstration below implements this policy, letting the reader watch a fixed study budget flow to the region of proximal learning, and shift toward harder items, as competence rises.

For control to be well directed, the monitoring it rests on must be valid, and this is where the timing of a judgment matters. Thomas Nelson and John Dunlosky found that judgments of learning made immediately after study predict later recall poorly, but the same judgments made after a delay predict it far better, the delayed-judgment-of-learning effect (Nelson & Dunlosky, 1991). The delay forces the learner to retrieve the item from a longer-term store rather than reading it out of a still-active short-term buffer, so the cue driving the judgment becomes diagnostic of the very retrieval the judgment predicts. The practical lesson for control is sharp: a learner who tests a delayed judgment allocates study far more effectively than one who judges an item while it is still ringing in mind. Control is only as good as the monitoring that feeds it, and monitoring is improved by arranging for diagnostic cues.

Measuring Metacognition

Scoring a monitoring judgment requires separating two things it can be good at, and conflating them has caused lasting confusion. Calibration is absolute correspondence: across many judgments, does the mean judged probability match the observed proportion correct, so that events called seventy percent likely happen about seventy percent of the time. Resolution is relative discrimination: do higher judgments attach to items more likely to be correct than lower ones, regardless of the absolute level. A learner can be well calibrated but have no resolution, judging every item at the group average, or have high resolution while being badly miscalibrated, ranking items perfectly but shifted uniformly toward overconfidence. Thomas Nelson, comparing the measures then in use, argued that the Goodman-Kruskal gamma correlation between judgment and outcome is the appropriate index of resolution, because it scores the ordinal relationship the judgments are meant to capture without confounding it with the base rate of correct responses (Nelson, 1984). The second demonstration computes gamma directly from a set of judged and tested items.

Confidence-based measurement inherits a problem from signal detection theory: an observer's confidence-accuracy relationship is contaminated by response bias, the general readiness to express high confidence, which varies between people for reasons unrelated to how well they monitor. Brian Maniscalco and Hakwan Lau solved this by adapting the type-2 signal detection model, deriving meta-d-prime, the first-order sensitivity that would be needed to produce the observed confidence-accuracy data if the observer were metacognitively ideal (Maniscalco & Lau, 2012). Because meta-d-prime is expressed in the same units as the first-order sensitivity d-prime, their ratio gives a bias-free measure of metacognitive efficiency, with one denoting an observer whose confidence extracts all the information their performance contains. Stephen Fleming and Hakwan Lau set out the family of such measures and the assumptions each makes, framing the choice among gamma, calibration, and meta-d-prime as a choice about what one wants the number to mean (Fleming & Lau, 2014). Table 1 sets the principal metamemory judgments side by side.

Note. Each judgment is scored against the outcome it predicts, by calibration or by resolution (Nelson, 1984). The first three are prospective and drive control; retrospective confidence is the basis of the meta-d-prime measure of metacognitive sensitivity (Fleming & Lau, 2014).

Control does more than allocate time; it also chooses what to report. Asher Koriat and Morris Goldsmith showed that when accuracy matters, people withhold answers they are unsure of and coarsen the grain of the answers they do give, trading completeness for correctness under the control of their own confidence (Koriat & Goldsmith, 1996). Report is thus a second control lever driven by monitoring, and its quality depends on the same thing study allocation does: whether the confidence steering it is well resolved.

Overconfidence

If monitoring is inferential and cue-driven, its failures should be systematic rather than random, and the most consequential systematic failure is overconfidence concentrated among the least able. Justin Kruger and David Dunning demonstrated that people in the bottom quartile of performance on tests of reasoning, grammar, and humor grossly overestimated their rank, placing themselves near the average, and that the same deficits producing their poor performance deprived them of the metacognitive competence needed to recognize it (Kruger & Dunning, 1999). The result is often misread as a claim that the incompetent are more confident than the competent; the data show something narrower and more defensible, that the incompetent are more miscalibrated, their confidence far exceeding their accuracy, because the skills required to do a task and the skills required to judge that doing are frequently the same. The third demonstration drives a calibration curve, showing the bottom performers' judgments detaching from the diagonal into overconfidence while the top performers, if anything, slightly underestimate.

The reading matters for what follows from it. Because the effect is a failure of resolution and calibration rather than mere arrogance, it is improved by the same intervention that improves any monitor: giving the person diagnostic feedback about their performance, which supplies the cue their self-assessment lacked. Overconfidence is not a fixed trait but the predictable output of a monitor reading invalid cues, and the account that explains it is the same cue-utilization account that explains monitoring's ordinary successes.

The Neural Basis

Because metacognitive accuracy is dissociable from first-order performance, it should have a neural substrate separable from the substrates of the tasks it monitors, and the evidence points to the anterior prefrontal cortex. Arthur Shimamura, drawing the early synthesis, argued that the prefrontal cortex implements a dynamic filtering that monitors and controls posterior cognitive processes, casting metacognition as a species of executive control with a frontal signature (Shimamura, 2000). The structural evidence sharpened the localization. Stephen Fleming and colleagues found that people with better metacognitive sensitivity, those whose confidence best discriminated their correct from incorrect perceptual judgments, had more grey matter in the anterior prefrontal cortex and stronger connectivity of that region, with first-order performance held constant so that the correlation was with the monitoring rather than the seeing (Fleming et al., 2010).

Correlation became causation through interference. Elisabeth Rounis and colleagues applied theta-burst transcranial magnetic stimulation to the dorsolateral prefrontal cortex and found that it selectively impaired metacognitive sensitivity, reducing the coupling between confidence and accuracy while leaving discrimination performance itself intact (Rounis et al., 2010). The double result, structure predicting metacognition and stimulation degrading it while sparing the first-order task, is the strongest evidence that metacognition is a distinct function rather than a byproduct of performing well. It also closes the loop with the Nelson-Narens architecture: a meta-level that can be lesioned without touching the object-level is exactly what that framework predicts.

Worked Example

The distinction between calibration and resolution becomes concrete in the computation of the gamma correlation, the standard measure of monitoring resolution. Consider six items a learner has studied and rated with a judgment of learning, then been tested on. The judgments, in points, and the recall outcomes are: item A, 90, recalled; item B, 80, recalled; item C, 60, not recalled; item D, 50, recalled; item E, 30, not recalled; item F, 20, not recalled. Gamma asks a purely ordinal question: taking every pair of items whose outcomes differ, one recalled and one not, does the recalled item carry the higher judgment. It ignores the absolute level of the judgments entirely, which is why it scores resolution and not calibration.

The recalled items are A, B, and D, with judgments 90, 80, and 50; the not-recalled items are C, E, and F, with judgments 60, 30, and 20. That yields nine mixed pairs. A pair is concordant if the recalled item has the higher judgment and discordant if it does not. A over C, E, F: 90 beats all three, three concordant. B over C, E, F: 80 beats all three, three concordant. D over C, E, F: 50 loses to 60 but beats 30 and 20, so one discordant and two concordant. The tally is eight concordant pairs and one discordant. Gamma is their difference over their sum, that is eight minus one over eight plus one, which is seven ninths, about 0.78. The learner's monitoring has high resolution: their judgments order the items by memorability almost perfectly, the single inversion being item D. Crucially, this says nothing about calibration; the same gamma would result if every judgment were inflated by twenty points, because the ordering would be untouched while the absolute correspondence to a fifty-percent recall rate would be destroyed. Resolution and calibration are orthogonal, and gamma sees only the first.

Discussion

Metacognition matters to cognitive psychology because it converts introspection from a discredited method into a measured variable. A monitoring judgment is a claim the mind makes about itself, and because the claim predicts an event that can be observed, its accuracy can be scored, which is what lets the field study self-knowledge without taking the self-report at face value (Nelson & Narens, 1990). The payoff is a set of robust findings that a looser treatment would miss: that monitoring judgments are inferences from cues rather than readouts of traces, and so can be confidently wrong (Koriat, 1997); that the timing of a judgment changes its validity, a delay making it diagnostic (Nelson & Dunlosky, 1991); that accuracy has two orthogonal components, calibration and resolution, which no single naive measure captures (Nelson, 1984); and that the whole monitoring function is separable from first-order performance down to its prefrontal substrate (Fleming et al., 2010; Rounis et al., 2010). The applied stakes are considerable, because study is self-regulated: a learner who monitors well allocates effort well, and the large literature on effective learning techniques is in part a catalogue of ways to make one's own monitoring more diagnostic (Dunlosky et al., 2013). The open questions are about mechanism and generality. Whether a single metacognitive faculty serves memory, perception, and decision alike, or whether monitoring is domain-specific and only looks unified, remains contested; how far the cue-utilization account extends beyond metamemory is unsettled; and the precise division of metacognitive labor within the prefrontal cortex is still being drawn. What is settled is the reframing that Flavell and then Nelson and Narens achieved: self-knowledge is a control system with a measurable output, and the quality of the measurement sets the quality of everything built on it.

Common Misconceptions

A metacognitive judgment reads the strength of the memory directly.

Judgments of learning are inferences from cues such as processing fluency, not readouts of the trace (Koriat, 1997). This is why a fluently read item can be judged well learned and then forgotten: the cue was salient but not diagnostic of later recall. The felt directness of the judgment is itself part of the illusion.

Good monitoring means giving accurate confidence levels.

Accuracy of monitoring has two independent parts. Calibration is whether the absolute level matches performance; resolution is whether higher judgments track more-likely-correct items (Nelson, 1984). A person can be perfectly calibrated with zero resolution by judging every item at the group average, so a single confidence number is not the whole of monitoring quality.

Incompetent people are simply more confident than competent people.

The Dunning-Kruger data show the poorest performers are the most miscalibrated, their confidence far exceeding their accuracy, not that they are absolutely more confident than experts (Kruger & Dunning, 1999). The cause is that judging a task well often requires the very skill the task does, so a deficit harms performance and self-assessment together.

Glossary

Calibration.

The absolute correspondence between judged probability and observed accuracy, so that events called seventy percent likely occur about seventy percent of the time.

Control.

In the Nelson-Narens framework, the downward flow by which the meta-level acts on the object-level, initiating, continuing, or terminating a process.

Cue utilization.

The account on which metacognitive judgments are inferred from cues such as processing fluency rather than read directly from the memory trace.

Ease-of-learning judgment.

A prospective estimate, made before study, of how difficult an item will be to acquire.

Feeling of knowing.

A judgment, made when recall fails, that an unretrieved item would nonetheless be recognized; an early demonstrated form of accurate monitoring.

Gamma correlation.

The Goodman-Kruskal ordinal association between judgments and outcomes, the standard index of monitoring resolution; concordant minus discordant pairs over their sum.

Judgment of learning.

A prediction, made during or after study, that a studied item will later be recalled; the most studied metamemory judgment.

Meta-level.

The processing level that holds a model of the object-level and acts on it, the seat of monitoring and control.

Metacognition.

Cognition about cognition: the knowledge, monitoring, and control a person exercises over their own mental processes.

Metacognitive knowledge.

The stable, semantic beliefs a person holds about how cognition works, including person, task, and strategy variables; distinct from online experience.

Metacognitive sensitivity.

The degree to which confidence discriminates correct from incorrect responses, formalized bias-free as meta-d-prime relative to first-order sensitivity.

Metamemory.

The metacognition of memory: the monitoring and control of one's own remembering, the most developed province of the field.

Monitoring.

In the Nelson-Narens framework, the upward flow by which the meta-level is informed of the object-level's state and updates its model.

Object-level.

The first-order processing level that perceives, remembers, and solves, monitored and controlled by the meta-level.

Region of proximal learning.

The band of not-yet-learned items close enough to mastery that study yields the highest rate of return, toward which skilled learners direct effort.

Resolution.

The discriminative accuracy of monitoring: whether higher judgments attach to items more likely correct, independent of absolute calibration.

Key Researchers

John Dunlosky (Kent State University). Cognitive psychologist who established the delayed-judgment-of-learning effect with Thomas Nelson and led the review synthesizing effective, metacognitively grounded learning techniques.
Kent State University - ORCID - Google Scholar

John H. Flavell (1928-2025). Stanford developmental psychologist who coined the term metacognition and framed it as knowledge and cognition about cognitive phenomena, distinguishing stable metacognitive knowledge from online metacognitive experience.
Stanford University - Wikipedia

Stephen M. Fleming (University College London). Cognitive neuroscientist who linked metacognitive sensitivity to anterior prefrontal structure and set out the measurement framework distinguishing calibration, resolution, and meta-d-prime.
University College London - ORCID - Google Scholar

Asher Koriat (University of Haifa). Cognitive psychologist who developed the cue-utilization account of judgments of learning, establishing that metacognitive judgments are inferences from cues such as fluency rather than direct readings of memory.
University of Haifa - ORCID - Google Scholar

Hakwan Lau (RIKEN Center for Brain Science). Cognitive neuroscientist who, with Brian Maniscalco, derived the bias-free meta-d-prime measure of metacognitive sensitivity from the type-2 signal detection model.
Lau Lab - ORCID - Google Scholar

Janet Metcalfe (Columbia University). Cognitive psychologist who proposed the region-of-proximal-learning model of study-time allocation, showing that skilled learners direct effort by rate of return rather than by raw difficulty.
Columbia University

Frequently Asked Questions

What is metacognition?
Metacognition is cognition about cognition, the monitoring and control a person exercises over their own mental processes. It includes stable metacognitive knowledge about how minds work and the online metacognitive experiences, such as a feeling of knowing, that arise during a task. Treating a monitoring judgment as a measurable prediction about one's own cognition is what makes it scientifically tractable (Flavell, 1979).

Who coined the term metacognition?
The developmental psychologist John Flavell introduced the term in the 1970s, defining it as knowledge and cognition about cognitive phenomena and organizing it into metacognitive knowledge, experiences, goals, and strategies. His framework grew out of studies of what children come to understand about their own memory (Flavell, 1979).

What is the difference between metacognitive monitoring and control?
Monitoring is the upward flow of information, by which a person becomes informed of the state of their own cognition; control is the downward flow of command, by which they act on it, allocating study time or choosing whether to answer. Nelson and Narens modeled the two as directed links between a meta-level and an object-level (Nelson & Narens, 1990).

What is metamemory?
Metamemory is the metacognition of memory, the monitoring and control of one's own remembering. It is the most developed area of the field because its judgments, such as predicting whether a studied item will be recalled, can be scored precisely against later memory performance (Nelson & Narens, 1990).

What is a judgment of learning?
A judgment of learning is a prediction, made during or shortly after study, that an item will later be recalled. Such judgments are inferences from cues rather than direct readings of the memory trace, and delaying them makes them far more accurate because retrieval from long-term memory becomes the cue (Nelson & Dunlosky, 1991).

How is metacognitive accuracy measured?
Accuracy has two independent parts. Calibration is whether the absolute confidence level matches performance; resolution, indexed by the Goodman-Kruskal gamma correlation, is whether higher judgments track items more likely to be correct. The bias-free measure meta-d-prime captures resolution within a signal detection framework (Nelson, 1984).

Why are poor performers often overconfident?
Because judging a task well often requires the same skill the task itself demands, a deficit tends to impair performance and self-assessment together. Kruger and Dunning found the lowest performers were the most miscalibrated, their confidence far exceeding their accuracy, an effect corrected by diagnostic feedback (Kruger & Dunning, 1999).

Where in the brain is metacognition?
Metacognitive accuracy is tied to the anterior prefrontal cortex. People with better metacognitive sensitivity have more grey matter there, and disrupting the region with transcranial magnetic stimulation selectively degrades monitoring while leaving first-order performance intact (Fleming et al., 2010).

References