RESEARCH DOSSIER
Metacognition and Introspection
Abstract. Metacognition encompasses the evaluation and regulation of cognition; introspection supplies reports about experience, mental processes, and the apparent reasons for behavior. This critical narrative research monograph reconstructs the intellectual history and conceptual structure of both traditions and synthesizes experimental evidence on metamemory, judgments of learning, confidence–accuracy calibration, choice blindness, verbal reports, error monitoring, and cognitive offloading. Signal detection theory, hierarchical monitoring–control frameworks, Bayesian self-evaluation, cue-utilization accounts, evidence accumulation, and models of meta-uncertainty are compared against behavioral and neural findings. The methodological analysis distinguishes objective performance from confidence bias, sensitivity, efficiency, self-report validity, and adaptive use of monitoring information; particular attention is paid to test–retest reliability and model-dependent assumptions. Educational, developmental, social, clinical, and human–artificial implications are examined within their evidentiary limits. The synthesis concludes that self-knowledge is a family of measurable, inferentially organized processes whose accuracy depends on cues, tasks, feedback, and control. This work reports no original experiment or formal systematic-review protocol.
Keywords: metacognition; introspection; self-monitoring; self-knowledge; confidence calibration; metacognitive sensitivity; meta-d-prime; type-two signal detection; judgments of learning; metamemory; error awareness; verbal reports; choice blindness; metacognitive control; cognitive offloading; Artificial Era
The Scientific Problem of Knowing One's Own Mind
Metacognition is the set of representations and operations through which a cognitive system evaluates and regulates its own information processing. Introspection is a family of practices and experiences in which an individual attends to, describes, or reports aspects of mental life. Their intersection is scientifically consequential: confidence in an answer, awareness of uncertainty, recognition of error, and verbal explanation of a choice can all be measured, yet they arise through partly different mechanisms. A theory of self-knowledge must specify the target state, the evidence available to the observer, the form of the self-judgment, and the action that follows. The developmental formulation of metacognitive knowledge and experience in Flavell (1979) remains a useful starting point because it includes the person's conception of tasks and strategies alongside immediate monitoring.
The guiding question of this research dossier concerns the accuracy and limits of access to one's own cognition. People can often identify mistakes without external feedback, adjust their study behavior, and decide when additional evidence is needed. They can also express strong confidence in false responses and invent apparently persuasive reasons for choices influenced by unrecognized factors. These properties must be investigated in their own empirical settings. The argument in Fleming (2024) connects confidence with inferential self-models, while Nisbett & Wilson (1977) demonstrates why fluent causal explanations should be assessed against independent evidence. An introspective report is itself an observable response with a production history, a time of sampling, and potential sources of bias.
This monograph is a critical narrative synthesis of foundational experiments, formal measurement approaches, cognitive and neural models, educational findings, and recent methodological evaluations. It makes no claim to have conducted new participant studies or a protocol-registered systematic review. The evidentiary hierarchy privileges original manipulations and directly inspectable analyses when describing causal effects, and treats reviews as syntheses of a wider literature. Throughout, a distinction is preserved among objective performance, subjective confidence, introspective content, verbal explanation, and adaptive regulation. The evidence does not license a single unitary faculty of perfect access to mental processes; it supports a family of measurable monitoring capacities whose reliability depends upon tasks, signals, reporting demands, and inferential assumptions Fleming & Lau (2014) Rahnev (2025).
Conceptual Distinctions and Levels of Description
A first-order cognitive judgment concerns the world or a task: whether a stimulus was present, whether a remembered statement is accurate, or which response is correct. A second-order judgment concerns the first-order operation: how likely that response is correct, whether retrieval will succeed, or whether more study is necessary. This operational distinction makes metacognition experimentally tractable because first-order outcomes can be compared with second-order estimates. It does not require a claim that every instance of second-order processing is conscious or linguistically articulated. The hierarchical treatment in Nelson (1990) describes monitoring as information flowing from the object level to a metalevel, with control operating in the reverse direction.
Metacognitive knowledge refers to relatively enduring beliefs about capacities, strategies, and task demands; metacognitive experiences are transient feelings of knowing, uncertainty, fluency, or difficulty; and metacognitive control consists of interventions such as checking, changing strategies, or requesting help. Introspection commonly accesses experiential appearances and the material presently available to attention, whereas explanatory introspection attempts to recover causal processing history. A person can accurately report that an answer feels uncertain while misidentifying the precise causes of that feeling. The methodological analysis of verbalization in Ericsson & Simon (1980) provides an essential distinction between reporting attended contents and constructing a theory about inaccessible processes.
These categories also separate reflective self-belief from trial-specific evaluation. Confidence about a particular perceptual discrimination is local; the belief that one is generally capable of accurate visual judgment is global. Learning and social interaction update both, and discrepancies between them are informative rather than logically inconsistent. A person may possess good discrimination of errors within a task and still report a chronically high or low confidence level across tasks. The 2026 integrative neuroscience account describes how local estimates of uncertainty can contribute to broader self-beliefs through multimodal integration Fleming (2026). Explicitly preserving levels of description prevents reduction of self-knowledge to one questionnaire score or one activation pattern.
Historical Origins of Metacognitive Research
Scientific introspection preceded the contemporary term metacognition, but its historical role differed across psychological schools. Early experimental psychology attempted to discipline reports of experience through stimulus control and trained observation; subsequent behaviorist and information-processing traditions favored measures that could be independently timed and manipulated. Contemporary research inherited both ambitions. It asks what a participant reports experiencing and what task conditions predict the report. The decisive change was the development of observable criteria for evaluating self-assessment: predictions of future recall, confidence conditional on correctness, and changes in strategy after perceived errors. The cognitive-developmental program articulated by Flavell (1979) identified knowledge of persons, tasks, and strategies as developing psychological variables.
One influential historical challenge was the review by Nisbett & Wilson (1977) of evidence that people sometimes fail to recognize influences on their judgments and subsequently offer plausible causal explanations. The authors argued that accounts of cognitive processes may derive from general theories of causal plausibility rather than direct observation of the processing itself. This thesis concerned the evidentiary basis of causal process reports and should not be expanded into a claim that all first-person phenomenological descriptions are false. Later experimental work on choice blindness provides a direct way to test discrepancies between selection, reported outcome, and subsequent reasoning Johansson et al. (2005).
A complementary research tradition made verbal reports useful under carefully specified conditions. Ericsson & Simon (1980) analyzed how concurrent thinking aloud may convey information currently attended to in short-term memory, while additional instructions to explain or justify can alter cognition and increase inference. Metamemory experiments then established that people can make informative judgments even about material they cannot immediately retrieve, and that those judgments change with temporal delay and testing cues Nelson (1990) Dunlosky & Nelson (1992). The resulting scientific program integrates critical analysis of self-explanation with psychometric analysis of self-monitoring, rather than treating these two historical traditions as mutually exclusive.
Monitoring and Control in the Metamemory Tradition
The monitoring–control framework treats cognition as an ongoing transaction among task processes, information about those processes, and executive allocations. A learner observing difficulty can monitor expected success, then change study time, request feedback, or select an alternative strategy. The metalevel need not possess a complete internal transcript of memory encoding to exert influence. It can operate with sampled cues that predict success imperfectly. The framework organized feeling-of-knowing experiments, learning judgments, and control of study as related but separable operations Nelson (1990). A rigorous account asks whether monitoring is accurate, whether control is sensitive to that information, and whether the resulting behavior improves the objective outcome.
Judgments of learning illustrate why monitoring can be simultaneously useful and biased. Learners commonly rate whether a studied item will be recalled later, often while the answer remains visible. Koriat (1997) found that such judgments are shaped by intrinsic properties of items, extrinsic study conditions, and mnemonic experiences such as retrieval fluency. The impact of a cue on confidence depends on whether learners believe it diagnoses future recall. An easily processed pair may feel mastered because the target is visible, although the eventual test requires generating the absent target. The relevant error is not simply an excessively high rating; it is a mismatch between the evidence available at judgment and the information required at criterion testing.
Control decisions give these estimates their practical significance. If ease of study becomes the basis for ending practice prematurely, monitoring error produces a learning cost. If uncertainty triggers retrieval practice and corrective feedback, monitoring can improve allocation. A delayed judgment based on the cue alone can provide a better approximation to the later retrieval problem than an immediate judgment made with both cue and target present Dunlosky & Nelson (1992). This result demonstrates how experimental manipulation of the monitoring environment can improve validity without assuming the learner gains direct access to hidden memory traces. It also illustrates the difference between improving calibration and improving retention: each should have its own outcome variable.
Introspective Evidence and the Problem of Causal Access
Introspective evidence has multiple targets. Participants may report a present sensation, classify an emotional state, describe which information they considered, or offer a retrospective explanation of why an outcome occurred. These reports differ in observational opportunity. Perceptual confidence can sometimes be linked to task accuracy; a causal explanation of a complex choice requires reconstructing an earlier sequence of influences that may not have been represented consciously. Nisbett & Wilson (1977) specifically questioned introspective access to higher-order process mechanisms by comparing judgments with experimentally established effects of stimuli. The ensuing research agenda requires a report's validity to be tested for the exact claim it makes.
Choice-blindness experiments provide an instructive causal manipulation. In Johansson et al. (2005), participants selected between face images, but the experimenters sometimes presented an outcome inconsistent with the selected option. Some participants failed to detect the mismatch and gave reasons for preferring the face presented as their choice. The observation shows that elaborated verbal explanation can be generated under experimentally induced misinformation about the person's own action. It does not imply that every explanation is post hoc or that introspective access is entirely absent; detection rates, task context, and what the report purports to explain constrain the inference. Replication and boundary conditions require separating the rate of unnoticed manipulations from the content of subsequent justifications.
Verbal reports can therefore be used as data when their elicitation process is modeled. Ericsson & Simon (1980) cautioned that asking for reasons or analyses recruits operations beyond simply reporting attended information. A distinction between concurrent and retrospective reporting matters because memory reconstruction, task demands, and social expectations change with delay. The experimental test should record behavior, timing, available stimuli, and report content, then compare competing predictions about what could have been attended. The critical standard is methodological specificity: introspection about a perceptual appearance, about attention, and about the causes of behavior calls for different forms of corroboration and different interpretations of disagreement.
Experimental Paradigms for Metacognitive Judgment
The simplest controlled paradigm collects an objective response followed by an estimate of whether it was correct. For a visual discrimination, the experimenter varies signal quality and registers accuracy, response time, and confidence on every trial. The joint distribution allows analysis of two different properties: the ability to discriminate stimuli and the ability to discriminate one's correct judgments from errors. Repeated trials are essential because an isolated confident mistake has little statistical meaning. In metacognitive experiments, matching objective difficulty across conditions reduces one source of ambiguity, but control of trial timing, response mapping, confidence scale, and feedback remains necessary Fleming & Lau (2014) Maniscalco & Lau (2012).
A memory experiment often presents studied materials, collects immediate or delayed judgments of learning, and later tests recall. Here the predicted event lies in the future, so calibration includes forgetting between judgment and test. A feeling-of-knowing procedure may ask whether an unrecalled answer could nevertheless be recognized. Retrospective confidence is obtained after an answer; prospective confidence predicts a later performance; and global self-belief describes a more stable assessment. Each has different informational inputs and outcome conditions Nelson (1990) Dunlosky & Nelson (1992). A good experiment documents these distinctions explicitly, rather than pooling measurements under a broad label such as awareness.
Process-report paradigms combine verbal protocols with gaze, response latency, stimulus interventions, and independent tests of behavior. The choice-blindness method manipulates the apparent outcome after a decision Johansson et al. (2005); thinking-aloud protocols capture what participants attend to while solving tasks Ericsson & Simon (1980). Requiring explanations may become an intervention in its own right. Experimental designs therefore distinguish the validity of a report as a description of current experience from its validity as a statement about hidden causal processes. Preregistration, transparent exclusion rules, and blind coding of open-ended reports improve interpretation, especially when researchers know which participants encountered a manipulated stimulus.
Calibration, Discrimination, Bias, and Efficiency
Metacognitive calibration describes agreement between expressed probability and observed success. If responses assigned approximately seventy percent confidence are correct about seventy percent of the time across an appropriately defined reference set, the ratings are locally calibrated. This criterion concerns reliability of probabilistic forecasting, and it depends on how confidence categories are grouped. Discrimination, often called metacognitive sensitivity, instead asks whether more confident trials tend to be more accurate than less confident trials. A participant can display good discrimination together with systematic overconfidence, or moderate calibration while failing to distinguish particular correct answers from incorrect ones. The conceptual separation is central to Fleming & Lau (2014).
Metacognitive bias refers to a general tendency to assign higher or lower confidence given comparable first-order performance. Efficiency evaluates the quality of confidence-based discrimination relative to the information available in the primary task. In signal-detection formulations, first-order sensitivity is commonly represented by d-prime, while meta-d-prime estimates the sensitivity a hypothetical first-order observer would need to reproduce observed second-order ratings. Their ratio is sometimes used as a normalized efficiency index Maniscalco & Lau (2012). Interpretation requires the assumptions of the underlying model, adequate confidence observations, and consideration of task design. No single scale-invariant statistic eliminates every possible form of confounding.
A contemporary methodological reassessment underscores these requirements. Rahnev (2025) evaluated seventeen metacognitive measures and examined validity, precision, response bias, difficulty dependence, split-half reliability, and stability across sessions. The study found generally high split-half reliability alongside poor test–retest reliability for many measures, with substantial dependencies on first-order performance. Rausch et al. (2026) further showed through simulations that the apparent invariance of meta-d-prime divided by d-prime depends on assumptions about how confidence is generated. Thus a reported individual difference in metacognitive efficiency should be accompanied by model checks, sensitivity analyses, trial counts, and a clear account of whether the target is stable ability or context-specific performance.
Signal Detection Theory and Second-Order Analysis
Signal detection theory distinguishes sensitivity to a stimulus from the criterion used to choose a response. In the metacognitive extension, the relevant discrimination concerns correct and incorrect first-order responses, indexed by confidence. One common visualization is a type-two receiver-operating characteristic curve obtained by varying confidence thresholds; a confidence measure that distinguishes correctness will yield a different distribution from an uninformative one. The area under this curve captures ranking performance under particular conventions. It is important to avoid interpreting confidence alone as a complete index of metacognitive competence, because an observer can shift the confidence criterion while retaining similar discrimination Fleming & Lau (2014).
The meta-d-prime model of Maniscalco & Lau (2012) formalizes an observer's second-order performance under an assumed signal-detection structure. It uses the pattern of confidence ratings conditioned on first-order decisions to infer information availability relative to primary sensitivity. This makes comparisons across conditions more principled than simply contrasting raw confidence–accuracy correlations. Yet a normalized score inherits limits from how evidence distributions, response criteria, and confidence generation are specified. The model's practical appeal should be understood as a transparent statistical proposal rather than as an assumption-free measurement of introspective truth. Researchers should inspect whether modeled confidence distributions resemble observed ones and whether trial-level variation is represented appropriately.
A further difficulty is that metacognitive sensitivity can change because confidence uses different evidence than choice. Confidence may incorporate additional samples after a response, learning about task difficulty, or cues from response fluency. Consequently, it is possible to vary primary accuracy and metacognitive readout partly independently without locating a self-monitoring module in a specific brain area. Rausch et al. (2026) tested the dependence of standard efficiency measures on alternative generative models, while Rahnev (2025) compared measures over multiple empirical criteria. These contributions favor explicitly modeling the process that produces ratings, reporting convergent statistics, and treating measurement uncertainty as part of any substantive claim about psychological differences.
Reliability, Validity, and the Unit of Generalization
The reliability of a metacognitive score is the extent to which its variation can be attributed to the construct rather than to sampling noise and fluctuations. Split-half estimates from one long experimental session measure a different property from stability across days. With enough trials, within-session aggregation may look internally consistent, while changes in fatigue, strategy, task learning, or subjective scale use still alter later scores. The large comparison in Rahnev (2025) documents this divergence and recommends treating test–retest stability as a substantive validation question. A study that seeks a stable person-level trait needs repeated measurements and representative tasks rather than one short perceptual block.
Construct validity requires establishing what a confidence measure actually assesses. A confidence rating may reflect probability of correctness, sensory strength, estimated difficulty, self-efficacy, social incentives, or compliance with scale instructions. A numerical confidence scale is not automatically a pure measure of a metacognitive faculty. Convergent evidence includes a prediction that confidence directs behavior under controlled accuracy, a coherent fit to multiple experimental manipulations, and stability across tasks at a level justified by measurement reliability. Desender et al. (2018) provides an instructive dissociation by changing properties of sensory evidence while matching accuracy, showing that subjective confidence can predict information seeking in ways not reducible to objective success.
Generalization across domains remains a live empirical question. Metacognitive efficiency for memory and visual perception can correlate behaviorally, yet the anatomical correlates and mechanisms may differ McCurdy et al. (2013). The strength of a cross-task correlation is affected by the precision of each measure; heterogeneous paradigms can understate shared ability, while common methods and response formats can create apparent agreement. A strong design uses multiple tasks within each domain, latent-variable or hierarchical models, and independent samples. It asks how much variation is attributable to person, domain, task, trial, and reporting procedure. Such decomposition is more informative than labeling the entire construct either globally unified or strictly domain-specific.
Causal Inference and the Interpretation of Reports
A report that one knows, remembers, understands, or chose for a particular reason is a proposition that can be experimentally evaluated only when its referent is specified. Confidence can be checked against future correctness. A stated reason can be tested against manipulated stimulus attributes and subsequent choice patterns. Reports of a private sensation require more indirect convergent measures and careful attention to response criteria. The challenge is not the scientific impossibility of first-person inquiry; it is the risk of conflating evidence for the occurrence of a report with evidence for the precise mental mechanism asserted in the report. Peters (2025) develops a psychophysical approach for studying the structure and limits of such subjective measurements.
Research involving verbal explanation is especially vulnerable to demand effects. Participants may infer that an experimenter expects coherent justifications, and they may construct explanations consistent with a presented outcome. The choice-blindness findings Johansson et al. (2005) demonstrate this concern under controlled substitution. Conversely, concurrent reports of attended information can add valuable temporal information when they are elicited without requiring causal interpretation Ericsson & Simon (1980). An adequate design compares reporting instructions, tests whether reporting itself changes behavior, and uses objective records of what information was available at the time. If the report is collected after extensive delay, a model of memory reconstruction becomes indispensable.
Causal interpretation of metacognitive interventions also demands separation between changed beliefs and changed performance. Feedback can lower expressed confidence without improving accuracy, and practice can improve both performance and calibration for different reasons. Random assignment to reporting conditions, neutral instructions, repeated outcome measures, and clear comparison groups help resolve these possibilities. Combining behavior, confidence, process tracing, and independently validated outcomes can support modest causal conclusions about specific manipulations. Broad claims that individuals directly observe the causes of every thought or that introspective reports universally fail exceed what these experimental designs establish Nisbett & Wilson (1977) Fleming (2024).
Cue-Based and Inferential Accounts of Monitoring
The cue-utilization approach explains many metacognitive judgments as inferences from signals correlated with the target outcome. During study, the ease of processing a familiar-looking pair can serve as a cue to knowledge, even if later free recall requires operations unlike those used while viewing the pair. Koriat (1997) differentiated intrinsic item properties, extrinsic study conditions, and mnemonic experiences, showing why some cues influence objective learning and subjective prediction differently. This perspective identifies the cue environment as part of the causal explanation. Learners are not required to observe memory strength directly in order to produce informative estimates; they learn associations between experiences and outcomes and use them under varying constraints.
Cue diagnosticity is task-relative. Reading fluency may be informative about recognizing already seen material, yet misleading about explaining a concept without prompts. Repeated exposure may change objective memory more than it changes subjective feelings, and the presence of answers at study may produce inflated confidence. The experiments described by Koriat & Bjork (2005) demonstrate that visible study information can sustain illusions of competence in paired-associate learning. The same pattern suggests a general principle: a judgment of knowing is accurate only to the degree that the available cues remain predictive under the eventual test conditions. Consequently, intervention design should change diagnostic experience, not merely instruct participants to be less confident.
The cue-based family of models can coexist with formal probabilistic accounts. A cue's predictive validity becomes a likelihood relation between observed information and an unobserved state of performance. Its influence also depends on prior beliefs about the task and the observer. Fleming (2024) emphasizes that propositional confidence is shaped by a model of both the world and one's own cognitive system. This account accommodates confidence even in counterfactual or prospective choices without imagining a literal internal observer inspecting cognition. The scientific question becomes which cues are sampled, how they are weighted, what prior beliefs guide interpretation, and how these components are updated by feedback.
Evidence Accumulation and Dynamic Confidence
Some theories derive confidence from the same evolving evidence that determines a first-order decision. In two-stage dynamic signal-detection theory, evidence accumulates until a choice is made and can continue accumulating after the initial commitment, producing later confidence estimates Pleskac & Busemeyer (2010). This structure predicts that response time, accuracy, and confidence are related but not interchangeable observables. The reported confidence can change after the action because new evidence is sampled, even when the original choice remains fixed. Such models are particularly valuable for perceptual and rapid cognitive decisions where experimental timing allows the sequence of evidence, choice, and self-evaluation to be estimated.
Dynamic accounts highlight a temporal problem in introspection. The subjective appearance of certainty at the moment of report may incorporate information accumulated after the event that is being explained. When a participant is asked what they knew at the time of a choice, the later self-report can be influenced by subsequent recognition, feedback, or rehearsal. Studies must therefore mark the timing of the primary response and of the metacognitive assessment separately. Yeung & Summerfield (2012) connects confidence judgments and error monitoring, emphasizing that evidence about choice accuracy can support online correction before external feedback arrives. Distinguishing simultaneous, immediate postdecisional, and delayed ratings is crucial for locating an effect within the decision sequence.
Evidence accumulation is one component of broader metacognitive computation, not a universal description of all self-knowledge. Complex autobiographical and educational judgments involve multiple cues and semantic beliefs about personal ability. Still, sequential sampling offers a precise explanation for why confidence may differ between two tasks with equal final accuracy: different evidence histories can yield different higher-order summaries. Integrative work Fleming (2026) develops a multiscale perspective in which local confidence computations feed adaptive control and global models of skill. Testing such accounts requires experimental manipulations of information reliability and timing, not a simple correlation between certainty and neural activity.
Bayesian Self-Evaluation and Metacognitive Inference
A Bayesian account formalizes self-evaluation as inference about whether a primary decision was correct, conditional on available evidence and a model of the decision system. In the framework proposed by Fleming & Daw (2017), first-order choice and second-order evaluation are coupled but distinct computations. The observer can combine sensory information, prior beliefs, expected task difficulty, and knowledge of response tendencies to estimate performance. This formulation explains how confidence may remain useful when direct external feedback is absent. It also allows systematic departures from optimal confidence if the internal model of the decision process or environmental statistics is inaccurate. Such departures are empirical phenomena to explain rather than unexplained noise.
A Bayesian representation does not imply that participants calculate explicit numerical posteriors. The framework is a computational hypothesis connecting latent variables to observed choices and reports. Competing architectures can differ over which evidence reaches the second-order estimator, whether information is shared with first-order decision making, and whether confidence precedes or follows response commitment. Fleming & Daw (2017) organizes these possibilities and places error detection within a common self-evaluation problem. Model comparison requires out-of-sample predictions and recovery studies, because flexible mechanisms may fit the same distributions of confidence ratings. The theoretical level must be distinguished from an implementation claim concerning a particular brain circuit.
The self-consistency model of Koriat (2012) offers another route to confidence: judgments can reflect the reproducibility of an answer across internally sampled representations. Confidence may be high when repeated samples agree, even if the samples share a systematic bias. This provides a mechanism for stable yet mistaken certainty and suggests why confidence can arise from evidence supporting the consistency of one's reasoning rather than its objective correctness. The contrast with Bayesian inference becomes experimentally useful when the models make different predictions under manipulations of sample variability, external reliability, or prior expectations. A comprehensive account should specify the relevant evidence distribution and distinguish genuine predictive precision from repetitive agreement among correlated samples.
Meta-Uncertainty and the Limits of Self-Evaluation
Confidence is often treated as an estimate of the reliability of a decision, but the quality of that estimate is itself uncertain. Boundy-Singer et al. (2023) modeled confidence as a noisy estimate of decision reliability and introduced meta-uncertainty to represent uncertainty about the variable controlling decision quality. The model explained datasets from multiple perceptual and cognitive studies and connected a measurable limitation of confidence judgments with variability across individuals. It offers an alternative to describing every metacognitive discrepancy as either deliberate misreporting or a defect of conscious awareness. An accurate first-order response can coexist with noisy estimates of its reliability because the information used to evaluate the decision is imperfect.
The meta-uncertainty proposal changes the meaning of comparisons between apparent metacognitive abilities. Two observers may perform equally well on a discrimination task but have different precision in their estimates of how trustworthy the evidence is. Their confidence–accuracy curves could differ even without changing motivational incentives or global self-beliefs. Conversely, an observer who confidently guesses may be responding to a misestimated local evidence distribution. Boundy-Singer et al. (2023) treats these possibilities as quantitative model predictions, which can be confronted with choices and confidence data. A useful experimental program varies sensory reliability, the availability of feedback, and the distribution of stimulus difficulty, then tests whether the estimated latent precision predicts independent decisions.
Theoretical sophistication does not remove the need for careful validation. Different confidence-generating processes can yield similar average calibration, and normalized metacognitive indices need not remain invariant under all generative models Rausch et al. (2026). Meta-uncertainty therefore belongs to a family of experimentally assessable hypotheses whose explanatory scope must be compared against simpler alternatives. The research question is whether an additional latent estimate improves predictions across tasks, sessions, and interventions. This standard also applies to philosophical language about knowing one's own mind: scientifically credible claims require observations that would discriminate among alternative mechanisms, rather than metaphors of internal transparency or opacity.
Neural Organization of Confidence and Error Monitoring
Neuroscientific studies connect metacognition to distributed neural systems engaged in decision formation, self-evaluation, attention, and cognitive control. Fleming et al. (2010) reported an association between introspective accuracy in a perceptual task and structural variation in anterior prefrontal regions, after dissociating objective task performance from accuracy of confidence judgments. This finding established a notable individual-differences association, while leaving open whether structural variation causes the behavioral difference, indexes a correlated factor, or reflects experience-dependent adaptation. Brain anatomy is informative at the population level, but it does not provide direct observation of subjective confidence, and reverse inference from regional activation to a single mental faculty requires restraint.
McCurdy et al. (2013) compared metacognitive efficiency in memory and visual perception and reported distinct neuroanatomical correlates alongside behavioral relations between the domains. The results support investigating partially specialized mechanisms while leaving room for shared signals and correlated anatomical development. Event-related neural studies add a different timescale. Boldt & Yeung (2015) identified common neural markers associated with confidence and error detection, suggesting that related evaluative computations can operate around and after response execution. These observations fit a view in which performance monitoring is implemented through multiple interactions rather than a single localized introspection center. Their generalization depends on task demands, statistical controls, and replication across measurement techniques.
The more recent synthesis of Fleming (2026) integrates computational, comparative, neural, and clinical evidence by tracing transformations from local uncertainty estimates to abstract self-directed beliefs. This multiscale picture is compatible with both domain-sensitive readouts and higher-order integration, yet individual empirical studies resolve only parts of the architecture. Neural signals correlated with confidence may also relate to attention, value, arousal, or response preparation; dissociation and causal intervention are necessary to establish computational specificity. The strongest inference emerges when a model jointly predicts behavior, report content, and neural measurements under experimentally varied conditions. The goal is an explanatory bridge across levels, with identifiable assumptions at every connection.
Learning Judgments, Illusions of Competence, and Retrieval
The experience of knowing during study is often a misleading preview of later independent recall. Koriat & Bjork (2005) examined illusions of competence when judgments are made with associative targets available, while the ultimate test requires reconstructing those targets from cues. Learners can rely on familiarity of the visible association, implicitly treating its ease of processing as evidence of retrieval strength. This illustrates a central distinction between accessibility under the current study environment and recoverability under changed conditions. An investigation should measure recall itself, the predicted probability of recall, and changes in study allocation, rather than describing overconfidence as a free-standing cognitive defect.
Manipulating the moment at which a judgment is requested alters the cues the learner can consult. Dunlosky & Nelson (1992) showed that delaying judgments of learning improves prediction of later recall when judgments are based on the cue alone, whereas the analogous delay has little benefit when the cue and target are both presented. The effect is informative because a delayed cue can prompt partial retrieval that resembles the eventual criterion test. This interpretation locates improved monitoring in the match between assessment conditions and target demands. It also means that instructional advice about reflection should specify what the learner tries to retrieve, what assistance is available, and when a judgment is made.
Monitoring quality matters because it helps determine what learners do next. Corrective feedback after difficult retrieval attempts can change the association between experience and objective memory success. Koriat & Bjork (2006) found that experience with testing and delayed judgments can reduce foresight bias, with consequences for allocation of study time. Metcalfe (2017) reviews how errors followed by feedback can promote subsequent learning, including conditions in which initially confident mistakes are particularly amenable to correction. These findings caution against measuring an educational intervention solely by students' immediate sense of understanding. Improvement should be established on independently administered retention or transfer tests and on whether monitoring supports better decisions about subsequent practice.
Introspection, Preferences, and Constructed Reasons
Reason analysis can change the object it purports to describe. Wilson & Schooler (1991) tested how asking participants to explain their preferences affected judgments about strawberry jams and college courses. In those experimental settings, some introspection conditions produced preferences less aligned with external criteria than control conditions. The finding establishes an effect of a particular reporting intervention on decisions; it does not establish a universal rule against reflective reasoning. The relevant mechanism involves the selective accessibility and weight of verbalizable attributes, which may diverge from the information that originally shaped preference. The quality criterion also matters, because expert rankings and personal preferences need not coincide in every domain.
Choice blindness Johansson et al. (2005) adds evidence that explanations can adapt to a falsely presented outcome. The participant's stated intention, observed selection, experimenter-controlled feedback, and later justification are distinct variables that can be experimentally separated. When an outcome is substituted, a participant may rationalize it through available attributes even though those attributes did not produce the originally recorded selection. This is a strong example of why retrospective causal reports need independent behavioral anchoring. It also suggests that confidence in one's explanation should itself be measured, because eloquence, fluency, and willingness to explain may reveal social and linguistic competence rather than accurate process memory.
Taken together, these experiments motivate a constructive account of introspection: first-person descriptions arise from experience, selective attention, memory, and causal inference. Their validity depends on which source contributes the information relevant to the claim. The account avoids both indiscriminate trust in reasons and indiscriminate dismissal of participants' experiences. As argued by Peters (2025), psychophysical approaches can test systematic relationships in subjective reporting even where the report does not provide a transparent window onto every neural or cognitive mechanism. Future experiments should combine report manipulations with choices, timing, and independent criteria, and should distinguish changes in stated reasons from changes in underlying preference.
Error Awareness, Feedback, and Adaptive Regulation
Error awareness connects metacognitive evaluation to immediate changes in behavior. A person may detect that an action was mistaken even before the environment supplies corrective information. Yeung & Summerfield (2012) reviews how confidence and error monitoring can be described through shared decision-evaluation processes, including signals that accumulate around response selection. Boldt & Yeung (2015) observed overlapping neural markers for confidence and error detection, extending the argument that these phenomena should be compared within unified experimental tasks. Yet subjective awareness, electrophysiological error markers, and overt corrective actions can dissociate. An experiment should report each measure separately and determine whether errors were consciously acknowledged, behaviorally corrected, or merely registered by an indirect physiological marker.
Error monitoring has consequences when it changes the allocation of cognitive resources. A detected mistake can trigger slower responding, checking, evidence gathering, or a switch of strategy. Desender et al. (2018) dissociated confidence from objective accuracy under controlled evidence manipulations and found that subjective confidence could guide decisions to seek additional information. This result identifies a functional role for self-evaluation beyond retrospective reporting. However, reduced confidence does not always improve action: under some payoff structures it may produce unnecessary checking, delay, or abandonment of an adequate strategy. Adaptive value must be judged relative to task incentives, cost of additional information, and observed decision outcomes.
Feedback offers another test of the monitoring–control link. Metcalfe (2017) synthesizes evidence for the benefits of correcting errors during learning, while distinguishing the productive use of mistakes from their mere occurrence. A learner who accurately flags uncertainty but never seeks correction has monitoring without effective control. Another who retrieves, checks, and revises may perform well despite imperfect introspective access to the encoding process. These differences justify collecting longitudinal records of judgments and subsequent actions instead of relying on global self-reports of reflectiveness. The scientifically interesting question concerns the circumstances under which a self-evaluation leads to an advantageous update in beliefs, behavior, or strategy.
Development, Instruction, and Trainability
Metacognitive competence changes with knowledge, task experience, linguistic development, and opportunities for feedback. Flavell (1979) located metacognitive knowledge and monitoring within cognitive development, noting that children's understanding of mental operations and strategic demands is itself acquired. Developmental comparisons require caution because younger participants may differ in the language used for confidence scales, the strategies available for first-order tasks, and motivation to report uncertainty. Improvements in self-report need not equal improvements in the underlying monitoring mechanism. Age-appropriate nonverbal confidence measures, matched task difficulty, and comparable incentives can help separate these sources. The relevant trajectory is multidimensional: knowledge of strategies, local sensitivity, scale use, and effective control can develop at different rates.
Training research asks whether monitoring capacities can improve through practice and whether gains transfer beyond the trained task. Carpenter et al. (2019) tested adaptive metacognitive training and reported enhancements extending beyond the trained domain, supporting the possibility of some flexible and shared resources. Interpretation nevertheless depends on the chosen control condition, task similarity, and measurement of first-order performance, because improvements in task strategies or use of rating scales can imitate changes in self-evaluation. Transfer should be specified by distance from the training environment, and longitudinal follow-up is needed to determine whether any change persists. A broad educational promise requires replication in real learning contexts with meaningful outcome measures.
Evidence about effective learning methods helps translate monitoring research into an educational design problem. The review by Dunlosky et al. (2013) found broad support for practice testing and distributed practice among the techniques examined, while identifying limitations in evidence for some frequently used methods. These are findings about learning interventions, not proofs that an internal metacognitive score is the direct mediator of every benefit. Instruction can support metacognition by arranging retrieval attempts, transparent feedback, and opportunities to revise plans. Evaluation should simultaneously measure retention, calibration, study allocation, and learners' interpretation of diagnostic cues. Educational claims are strongest when the route from intervention to monitored evidence to strategic control is empirically specified.
Metacognition Across Tasks, Individuals, and Social Contexts
People differ in their habitual confidence, their ability to discriminate correct responses from errors, and their willingness to revise judgments. Yet individual-difference research faces an identification problem: a score observed in one task may be dominated by characteristics of the task and sampling occasion. McCurdy et al. (2013) reported relations between memory and visual metacognitive efficiency alongside partly distinct anatomical correlates. Rahnev (2025) then showed that estimates can have strong within-session reliability while remaining relatively unstable across sessions. These findings motivate designs that repeat measurements over tasks and days, model measurement error, and report confidence intervals around person-level estimates. They also caution against assigning durable psychological labels from a brief online confidence test.
Metacognition is socially consequential because confidence helps others decide whose information to trust, when to collaborate, and when to ask for clarification. A confident expression may function as a communicative signal even when its calibration is imperfect. The 2026 integration by Fleming (2026) places self-evaluation in a framework that includes social coordination and abstract self-beliefs. The same distinction between local confidence and global identity can clarify why expert status, institutional roles, and cultural norms influence how uncertainty is displayed. However, a social difference in expressed confidence should not be interpreted as a difference in metacognitive sensitivity without independent accuracy measures and controls for the incentives surrounding public disclosure.
The study of such differences also requires measurement invariance. Response scales may be understood differently across languages, educational contexts, and cultures; public statements of certainty may carry different interpersonal consequences. A robust cross-population study would translate and validate instructions, establish comparable baseline performance, and test whether latent constructs relate to the same outcomes across groups. The methodological arguments in Fleming & Lau (2014) and Rahnev (2025) apply especially strongly here: variation in bias, sensitivity, and first-order ability must be separated before comparing people. Metacognition has a social organization, but evidence about that organization must derive from sampled populations and actual observations rather than assumptions about groups.
Cognitive Offloading and Human–Artificial Assistance
Metacognitive decisions increasingly concern whether a cognitive task should be performed from unaided memory, supported by an external reminder, or delegated to an information system. This is a problem of control informed by confidence. In an experimental prospective-memory setting, Gilbert (2015) found that both subjective confidence and objective ability predicted participants' strategic use of reminders, even when the two measures were statistically dissociated. A person who mistrusts their memory may offload more often regardless of actual performance; a highly confident person may forgo reliable external support. The adaptive outcome depends on the accuracy of self-assessment, the cost and trustworthiness of the tool, and the consequences of failure.
The same architecture can guide research on assistance from generative artificial systems. A human user estimates the quality of their own understanding, decides when to consult an artificial system, evaluates the returned material, and determines whether independent verification is warranted. External fluency can become an additional cue to confidence, but the perceived clarity of an answer does not establish its factual accuracy. Research should manipulate response reliability and explanation style independently, measure the user's objective decisions, and record how confidence changes before and after assistance. Desender et al. (2018) provides a behavioral foundation for studying confidence-guided information seeking; Fleming (2024) supplies an inferential account of the self-evaluations that may govern that choice.
Within the Artificial Era, the scientific question is how self-evaluation and distributed cognitive control are reorganized when humans work with nonbiological systems. An artificial model can be equipped with uncertainty estimates, self-checking procedures, and confidence-like output variables, but these operational features do not independently establish subjective introspection. Their calibration must be tested against reproducible outcomes and their behavioral claims against actual access to evidence. The benchmark for human–artificial cooperation is therefore a joint performance design: verify individual and collaborative accuracy, detect overreliance and underreliance, estimate when external support improves reasoning, and separate computational self-monitoring from philosophical accounts of Artificial. The distinction follows established measurement principles Fleming & Lau (2014) Rahnev (2025).
Clinical Relevance and the Interpretation of Self-Insight
Metacognition is relevant to clinical research because the accuracy of self-evaluation can influence beliefs about symptoms, learning from feedback, and decisions to seek assistance. Such relations are complex: confidence may concern perceptual evidence, autobiographical memory, decision outcomes, or global self-beliefs, each with different task and clinical correlates. Fleming (2026) reviews how altered self-evaluation may contribute to disruptions of reality monitoring and insight, while a broader neuroscience of uncertainty seeks to connect computational variables with lived experience. This scientific interest does not justify translating a laboratory confidence score directly into a diagnosis. Clinical conclusions require validated instruments, appropriate comparison groups, longitudinal information, and attention to medication, distress, and cognitive performance.
A symptom report is not equivalent to a causal analysis of that symptom. A participant may provide important first-person evidence about distress or unusual experiences while remaining uncertain about their psychological and biological origins. Similarly, someone with low confidence may sometimes discriminate correct from incorrect judgments effectively but express a negative global self-assessment. The distinction between bias, sensitivity, and efficiency described by Fleming & Lau (2014) helps prevent simplistic interpretation. The research program should investigate whether specific monitoring mechanisms predict clinically meaningful functioning above and beyond symptom severity, task ability, and known environmental conditions. Treatment implications should follow from trials with appropriate clinical endpoints, rather than from conceptual analogy.
Methodological rigor is especially important where experimental tasks are conducted with small or heterogeneous clinical samples. Differences in attention, fatigue, comprehension of rating instructions, and first-order performance may distort metacognitive comparisons. Multi-session measurement and robustness analyses are valuable in light of the stability concerns raised by Rahnev (2025). A careful clinical synthesis also acknowledges that information about subjective experience may be scientifically indispensable even when causal process descriptions are uncertain Peters (2025). The appropriate conclusion is that clinically significant self-evaluation has several experimentally separable dimensions, each requiring its own empirical validation and ethically designed inquiry.
Critical Scientific Disputes and Alternative Explanations
A central unresolved dispute concerns whether metacognition is a stable, domain-general ability or a collection of context-dependent evaluations. Evidence for shared behavioral variance can arise from genuine common computation, shared response formats, or correlated performance factors. McCurdy et al. (2013) demonstrated that behavioral coupling can coexist with distinct anatomical associations; Rahnev (2025) revealed weaknesses in test–retest stability for multiple indices. The most constructive interpretation asks how hierarchical sources of variance are organized. Researchers should define the level at which generality is claimed: trial, task, domain, person, or population. Without this specification, apparently conflicting reports may estimate different quantities and cannot be reconciled simply by averaging effect sizes.
A second dispute concerns the epistemic status of introspective reports. Nisbett & Wilson (1977) challenged direct causal access to higher-order processes, and Johansson et al. (2005) showed experimentally that reasons can be supplied for outcomes that were covertly substituted. Ericsson & Simon (1980) nonetheless explained how verbal protocols can report attended information under defined conditions, while Peters (2025) treats subjective reports as potentially meaningful psychophysical measurements. The evidence favors operational discrimination among types of introspection and levels of validation. A theory of unreliable causal explanation can coexist with strong evidence for discriminable subjective states. Scientific disagreement is productive when competing theories specify which classes of report should succeed, fail, or change under controlled manipulations.
A third dispute concerns formal measurement itself. Meta-d-prime provides a principled signal-detection comparator Maniscalco & Lau (2012), but model-dependent efficiency can be sensitive to assumptions about confidence generation Rausch et al. (2026). Boundy-Singer et al. (2023) offers meta-uncertainty as a computational limitation, while Bayesian and evidence-accumulation theories propose other structures Fleming & Daw (2017) Pleskac & Busemeyer (2010). None is conclusively selected by a mere association between confidence and accuracy. Future comparisons should predefine discriminating predictions, fit models to trial-level data, evaluate posterior predictive checks, and test external validity. The aim is to explain which mechanism generates a specified pattern of behavior under known conditions, while preserving the possibility that several mechanisms contribute across timescales.
Open Questions and an Evidence-Based Research Agenda
The next generation of studies should integrate sampling of first-order cognition, second-order judgments, and subsequent control across multiple timescales. A compact laboratory study can identify local sensitivity under carefully held difficulty, whereas longitudinal research can investigate the development of global self-beliefs and the stability of control policies. Fleming (2026) proposes bridges from rapid uncertainty transformations to abstract self-evaluation. A compelling test would manipulate the reliability of specific cues while independently varying feedback and incentives, then assess both immediate confidence and later self-beliefs. This design can determine when local errors propagate into enduring beliefs and when contextual learning corrects them. Such research must also quantify measurement uncertainty rather than assuming that an individual difference is an invariant characteristic.
A second priority is investigating interventions whose mechanisms can be identified. Delayed learning judgments, retrieval attempts, explicit feedback, and adaptive metacognitive training can change different elements of the monitoring–control cycle Dunlosky & Nelson (1992) Carpenter et al. (2019). Experimental work should separate improved primary skill, improved sensitivity, reduced bias, and more adaptive control, with independent follow-up tasks. A preregistered multisite design could test whether changes persist over months and transfer across perception, memory, and reasoning. Study protocols should specify the criterion outcomes and report null and contrary findings. When improvement occurs only for a narrowly trained response format, the conclusion must remain correspondingly narrow.
A third priority is designing ecologically realistic tests of assisted cognition. Human–artificial partnerships create repeated opportunities for a person to update confidence using machine-produced evidence. The validity of those updates depends on both the tool's reliability and the user's metacognitive model. Research can combine calibration measures, behavioral traces, and randomized variations in assistance quality to study error detection, trust, and strategic delegation. Gilbert (2015) establishes the relevance of confidence to offloading decisions, and Desender et al. (2018) links confidence with further information seeking. The unresolved issue is how monitoring stays adaptive when sources are heterogeneous, outputs are persuasive, and the costs of independent verification vary across decisions. These questions belong to an empirically accountable psychology of the Artificial Era.
Conclusion
The scientific study of metacognition and introspection reveals an organized family of second-order cognitive capacities through which people estimate, explain, and regulate their own mental activity. The strongest findings concern observable relationships among task outcomes, confidence judgments, reports, and subsequent action. Foundational work established metacognitive knowledge and monitoring as research objects Flavell (1979), created formal monitoring–control models Nelson (1990), and demonstrated that judgments of learning depend on diagnostic cues Koriat (1997). A parallel tradition established that retrospective explanations of cognitive causes can diverge from experimentally observed influences Nisbett & Wilson (1977) Johansson et al. (2005). Together these traditions define the subject: self-evaluation can be useful, systematically biased, causally consequential, and scientifically measurable.
The methodological synthesis has a decisive implication. Accuracy, calibration, bias, efficiency, and introspective explanatory validity answer different questions and must be evaluated separately. Signal-detection and Bayesian models provide explicit accounts of confidence Maniscalco & Lau (2012) Fleming & Daw (2017), while empirical critiques demonstrate that scores are affected by assumptions, task performance, and temporal stability Rausch et al. (2026) Rahnev (2025). Neural investigations reveal relationships among frontal, memory-related, and performance-monitoring systems but require careful separation of correlational and causal evidence Fleming et al. (2010) McCurdy et al. (2013). No single behavioral score, questionnaire, or brain correlate presently exhausts the psychological phenomenon. Scientific progress depends on specifying the sampled evidence and the mechanism producing each self-judgment.
The resulting research agenda connects self-knowledge with adaptive learning, error correction, decision making, social communication, and human–artificial cooperation. Metacognitive monitoring becomes consequential when it changes the selection of actions and the allocation of resources; introspection becomes scientifically interpretable when reports are evaluated against their stated referents and modes of production. Evidence from retrieval practice, feedback, confidence-guided information seeking, and cognitive offloading demonstrates these functional relationships Metcalfe (2017) Desender et al. (2018) Gilbert (2015). The enduring conclusion is that knowing one's own mind is an empirically structured achievement shaped by information, inference, and control. The psychology of the Artificial Era inherits this framework while extending it to environments in which cognitive evidence and action are distributed across Homo and Artificial.
References
- Boldt, A., & Yeung, N. (2015). Shared neural markers of decision confidence and error detection. The Journal of Neuroscience, 35(8), 3478–3484. https://doi.org/10.1523/JNEUROSCI.0797-14.2015
- Boundy-Singer, Z. M., Ziemba, C. M., & Goris, R. L. T. (2023). Confidence reflects a noisy decision reliability estimate. Nature Human Behaviour, 7, 142–154. https://doi.org/10.1038/s41562-022-01464-x
- Carpenter, J., Sherman, M. T., Kievit, R. A., Seth, A. K., Lau, H., & Fleming, S. M. (2019). Domain-general enhancements of metacognitive ability through adaptive training. Journal of Experimental Psychology: General, 148(1), 51–64. https://doi.org/10.1037/xge0000505
- Desender, K., Boldt, A., & Yeung, N. (2018). Subjective confidence predicts information seeking in decision making. Psychological Science, 29(5), 761–778. https://doi.org/10.1177/0956797617744771
- Dunlosky, J., & Nelson, T. O. (1992). Importance of the kind of cue for judgments of learning (JOL) and the delayed-JOL effect. Memory & Cognition, 20(4), 374–380. https://doi.org/10.3758/BF03210921
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students’ learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
- Ericsson, K. A., & Simon, H. A. (1980). Verbal reports as data. Psychological Review, 87(3), 215–251. https://doi.org/10.1037/0033-295X.87.3.215
- Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American Psychologist, 34(10), 906–911. https://doi.org/10.1037/0003-066X.34.10.906
- Fleming, S. M., Weil, R. S., Nagy, Z., Dolan, R. J., & Rees, G. (2010). Relating introspective accuracy to individual differences in brain structure. Science, 329(5998), 1541–1543. https://doi.org/10.1126/science.1191883
- Fleming, S. M., & Lau, H. (2014). How to measure metacognition. Frontiers in Human Neuroscience, 8, Article 443. https://doi.org/10.3389/fnhum.2014.00443
- Fleming, S. M., & Daw, N. D. (2017). Self-evaluation of decision-making: A general Bayesian framework for metacognitive computation. Psychological Review, 124(1), 91–114. https://doi.org/10.1037/rev0000045
- Fleming, S. M. (2024). Metacognition and confidence: A review and synthesis. Annual Review of Psychology, 75, 241–268. https://doi.org/10.1146/annurev-psych-022423-032425
- Fleming, S. M. (2026). Towards an integrative neuroscience of metacognition. Nature Reviews Neuroscience, 27, 789–806. https://doi.org/10.1038/s41583-026-01081-x
- Gilbert, S. J. (2015). Strategic use of reminders: Influence of both domain-general and task-specific metacognitive confidence, independent of objective memory ability. Consciousness and Cognition, 33, 245–260. https://doi.org/10.1016/j.concog.2015.01.006
- Johansson, P., Hall, L., Sikström, S., & Olsson, A. (2005). Failure to detect mismatches between intention and outcome in a simple decision task. Science, 310(5745), 116–119. https://doi.org/10.1126/science.1111709
- Koriat, A. (1997). Monitoring one's own knowledge during study: A cue-utilization approach to judgments of learning. Journal of Experimental Psychology: General, 126(4), 349–370. https://doi.org/10.1037/0096-3445.126.4.349
- Koriat, A., & Bjork, R. A. (2005). Illusions of competence in monitoring one's knowledge during study. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(2), 187–194. https://doi.org/10.1037/0278-7393.31.2.187
- Koriat, A., & Bjork, R. A. (2006). Illusions of competence during study can be remedied by manipulations that enhance learners' sensitivity to retrieval conditions at test. Memory & Cognition, 34(5), 959–972. https://doi.org/10.3758/BF03193244
- Koriat, A. (2012). The self-consistency model of subjective confidence. Psychological Review, 119(1), 80–113. https://doi.org/10.1037/a0025648
- Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition, 21(1), 422–430. https://doi.org/10.1016/j.concog.2011.09.021
- McCurdy, L. Y., Maniscalco, B., Metcalfe, J., Liu, K. Y., de Lange, F. P., & Lau, H. (2013). Anatomical coupling between distinct metacognitive systems for memory and visual perception. The Journal of Neuroscience, 33(5), 1897–1906. https://doi.org/10.1523/JNEUROSCI.1890-12.2013
- Metcalfe, J. (2017). Learning from errors. Annual Review of Psychology, 68, 465–489. https://doi.org/10.1146/annurev-psych-010416-044022
- Nelson, T. O. (1990). Metamemory: A theoretical framework and new findings. Psychology of Learning and Motivation, 26, 125–173. https://doi.org/10.1016/S0079-7421(08)60053-5
- Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259. https://doi.org/10.1037/0033-295X.84.3.231
- Peters, M. A. K. (2025). Introspective psychophysics for the study of subjective experience. Cerebral Cortex, 35(1), 49–57. https://doi.org/10.1093/cercor/bhae455
- Pleskac, T. J., & Busemeyer, J. R. (2010). Two-stage dynamic signal detection: A theory of choice, decision time, and confidence. Psychological Review, 117(3), 864–901. https://doi.org/10.1037/a0019737
- Rahnev, D. (2025). A comprehensive assessment of current methods for measuring metacognition. Nature Communications, 16, Article 701. https://doi.org/10.1038/s41467-025-56117-0
- Rausch, M., Hellmann, S., & Zehetleitner, M. (2026). Measures of metacognitive efficiency across cognitive models of decision confidence. Psychological Methods, 31(4), 607–626. https://doi.org/10.1037/met0000634
- Wilson, T. D., & Schooler, J. W. (1991). Thinking too much: Introspection can reduce the quality of preferences and decisions. Journal of Personality and Social Psychology, 60(2), 181–192. https://doi.org/10.1037/0022-3514.60.2.181
- Yeung, N., & Summerfield, C. (2012). Metacognition in human decision-making: Confidence and error monitoring. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1594), 1310–1321. https://doi.org/10.1098/rstb.2011.0416