Metacognitive Sensitivity: How Researchers Measure Insight Into Performance

Author: Angela Bogdanova · Published: October 16, 2026 · Editorial Policy

Research cluster: Consciousness, Metacognition & Self-Awareness · Written by Artificial Sapiens

Metacognitive sensitivity is the extent to which a person's confidence judgments distinguish their own correct decisions from their incorrect ones. Researchers typically measure it by asking participants to complete an objectively scored task, rate confidence after each decision, and then analyze how effectively confidence predicts trial-by-trial correctness. The signal detection measure meta-d′ expresses the information about accuracy available in confidence ratings on a scale comparable to first-order task sensitivity, d′. The ratio meta-d′/d′, commonly called M-ratio, is an estimate of metacognitive efficiency relative to primary task performance. These quantities answer related but distinct research questions, and all require careful interpretation. The foundational account is Maniscalco and Lau (2012); contemporary limitations are examined in Rahnev (2025).

The essential scientific distinction is between performing a task well and knowing, on a particular occasion, whether that performance was successful. A highly accurate person may be poor at identifying the rare trials on which they err. A less accurate person may distinguish their own right and wrong answers surprisingly well. Confidence level adds a third dimension: a person's average confidence can be high or low while its discrimination between successes and errors remains strong. Metacognitive sensitivity is a measurable property of judgment under specified experimental conditions, not a diagnosis, a global intelligence score, or a complete measure of self-awareness.

What Metacognitive Sensitivity Means

First-order performance and second-order evaluation

In laboratory language, a type 1 or first-order task asks a participant to decide something about the world: Which image contains more dots? Was the word on the study list? Is the faint grating tilted left or right? The researcher knows the correct answer. Type 1 accuracy is the percentage of correct answers; in a suitable signal detection design, d′ also quantifies the participant's discrimination of the underlying stimulus alternatives. The same participant then gives a type 2 judgment about the decision just made, frequently using a confidence scale. Type 2 sensitivity concerns how well that judgment separates correct from incorrect type 1 decisions. This distinction is introduced in Fleming and Lau (2014).

Metacognitive sensitivity should be interpreted within the family of metacognition concepts. Metacognitive knowledge consists of beliefs about one's capacities and strategies; metacognitive monitoring evaluates current processing; metacognitive control uses that evaluation to change what one does. A confidence rating is one expression of monitoring. An ability to assign appropriate confidence after choices does not, by itself, establish that the person selected a useful study strategy, accurately described their mental mechanisms, or regulated subsequent behavior. The synthesis by Fleming (2024) explains how confidence reports can arise from inferences about one's decisions and internal evidence.

A simple worked example

Imagine two people each answer 100 moderately difficult questions and get 75 correct. Both reach 75% first-order accuracy. Person A reports relatively high confidence on most of the 75 correct answers and relatively low confidence on most of the 25 incorrect answers. Person B gives much the same confidence rating whether an answer is correct or incorrect. Person A exhibits higher metacognitive sensitivity in this example because their confidence distinguishes outcomes that Person B's confidence fails to separate. The example is conceptual: a formal estimate requires trial-level data, the confidence scale and a model or statistic appropriate to the design.

Now imagine Person A reports 60% confidence on average and Person B reports 90%. Their mean confidence difference measures a tendency to express more or less certainty. It does not establish which person detects errors more reliably. Nor does a 75% accuracy score tell us whose confidence is better calibrated. When probability confidence is meaningful, calibration asks whether events assigned, say, 70% confidence are in fact correct about 70% of the time. Sensitivity asks whether confidence can rank individual decisions by probability of correctness. These capacities may coexist, but they are empirically separable. See the account of clinical confidence measurement in Hoven and colleagues (2019).

The Measurement Vocabulary: Six Quantities Researchers Keep Separate

Accuracy and d′

Accuracy is the fraction of objectively correct answers, such as 0.75 for 75 of 100 trials. d-prime, written d′, is a signal detection estimate of how clearly an observer distinguishes stimulus classes under a specified model. In a conventional yes/no detection task, d′ is defined using the normal-quantile transformed hit and false-alarm rates. A value of zero represents chance-level discrimination under the model; larger positive values reflect stronger discrimination. The exact analysis of two-alternative forced-choice tasks requires attention to the relevant task formulation. Because task difficulty shapes the number and distribution of correct and incorrect answers, a meaningful analysis of confidence must consider first-order performance (Maniscalco and Lau, 2012).

Confidence level and confidence bias

Confidence level is the magnitude of stated certainty, commonly expressed on a numerical scale or through graded categories. Confidence bias concerns systematic overconfidence or underconfidence relative to a benchmark. A group that often rates its decisions 90% likely to be correct while actually getting 65% right exhibits overconfidence on that comparison. A group may be underconfident on average yet still rank its own correct decisions above incorrect ones. Researchers should report confidence distributions and overall performance before inferring sensitivity from summary averages (Fleming and Lau, 2014).

Calibration

Calibration compares expressed probabilities with empirical success frequencies. If all answers tagged 80% confidence are correct about 80% of the time across an adequately sized dataset, that part of the confidence scale is calibrated. Researchers may display a reliability diagram or calculate calibration error. Calibration can be studied at the group or individual level, but its interpretation depends on confidence elicitation and the number of trials per bin. A perfectly calibrated predictor may still make similar probability assessments across many situations and consequently have limited ability to discriminate among them. Calibration and sensitivity answer different questions (Hoven et al., 2019).

Resolution, discrimination, and metacognitive sensitivity

Resolution in the judgment literature concerns the ordering of occasions: Does greater confidence reliably correspond to greater likelihood of being correct? Confidence discrimination is frequently summarized as average confidence on correct trials minus average confidence on incorrect trials. Sensitivity is the broader capacity to separate successful from unsuccessful decisions with confidence, quantified by several possible statistics. Authors sometimes apply these terms somewhat differently, so the metric and exact target must accompany any reported result. Modern comparisons reveal that different measures can yield different inferences even when they address a common construct (Rahnev, 2025).

Meta-d′

Meta-d′ is the value of first-order discrimination sensitivity that an idealized signal detection observer would need in order to produce the observed pattern of confidence ratings, under the assumptions of the fitted type 2 model. It is estimated from the distribution of confidence reports across correct and incorrect responses, not by simply reading the participant's accuracy percentage. Greater meta-d′ ordinarily corresponds to confidence carrying more information about correctness. Its scale makes it possible to compare that fitted type 2 information with observed d′. Meta-d′ is a model-based estimate, not a direct count of errors noticed consciously, and its results depend on modeling assumptions (Maniscalco and Lau, 2012; Rahnev, 2025).

Metacognitive efficiency: M-ratio and M-difference

M-ratio equals meta-d′ divided by d′. In the standard ideal-observer benchmark, a value near 1 indicates metacognitive discrimination consistent with the information available to first-order decisions. A value below 1 suggests that confidence reflects less discriminative information than the benchmark, while a value above 1 can arise when extra information enters confidence after the first-order choice or when model assumptions and estimation properties matter. M-difference is meta-d′ minus d′ and offers another way to express the relative gap. Both are conditioned on the statistical framework. Ratios become unstable when d′ is very small, and observed M-ratio values should never be translated into a percentage of a person's true self-knowledge (Fleming and Lau, 2014; Rahnev, 2025).

How an Experiment Measures Metacognitive Sensitivity

Step 1: Choose a task with defensible objective answers

An experimental team might present two visual patterns and ask which has higher contrast, or ask whether a studied item appeared earlier. Trials must have a defined correctness criterion. The task should yield enough correct and incorrect responses to support estimating how confidence differs between them. A participant who is correct on virtually every trial provides too few errors for a stable estimate of error discrimination. Conversely, a first-order task at chance can make certain normalizations hard to interpret. Researchers often adjust stimulus difficulty to keep performance in an informative range; that adjustment itself requires controls, as discussed below (Rahnev and Fleming, 2019).

Step 2: Record a decision and a confidence judgment on every trial

After each choice, the participant reports confidence, perhaps on a four- or six-point scale or a continuous probability scale. The instructions must define whether confidence means certainty in the just-made answer or something else, and the study must specify whether judgments occur before or after feedback. Experiments may also manipulate decision time or allow a confidence report without a forced decision; such designs target different computational aspects and are not automatically interchangeable (Fleming, 2024).

Step 3: Classify trials by correctness and confidence

For each participant, the researcher creates a table linking the objectively correct category, the chosen response, accuracy and confidence. A descriptive first pass compares confidence on correct and incorrect trials and inspects how many trials occupy each rating category. A confidence scale with nearly all observations in one category provides less information than a scale that the participant uses meaningfully. Empty or sparse categories may require prespecified fitting decisions. Comparing raw averages alone is insufficient for strong inference about efficiency (Fleming and Lau, 2014).

Step 4: Estimate type 1 and type 2 performance

Depending on the task, the analysis computes accuracy and d′, draws a type 2 receiver operating characteristic curve, and fits meta-d′. In a type 2 ROC analysis, higher confidence in correct decisions contributes to type 2 hits, while high confidence in incorrect decisions contributes to type 2 false alarms. Varying the confidence criterion generates operating points that summarize confidence discrimination across thresholds. Area under this curve, often abbreviated AUROC2 or AUC2, is one established sensitivity measure. Meta-d′ uses an SDT generative framework to estimate the type 1 sensitivity compatible with the observed type 2 evidence. These are related constructions, but they are not numerically interchangeable (Maniscalco and Lau, 2012).

Step 5: Normalize, quantify uncertainty and examine robustness

Researchers may report meta-d′, M-ratio, another sensitivity index, uncertainty intervals, and sensitivity analyses under alternative assumptions. If the purpose is to compare groups or conditions, investigators should examine whether d′, task format, confidence scale use, response bias and trial counts are comparable. A model fitted to one short session may give a noisy participant-level estimate, even if a group average is informative. Hierarchical Bayesian models pool information across participants and can improve estimation in data-sparse settings, but their prior assumptions and model fit still matter (Fleming, 2017; Rahnev, 2025).

Meta-d′ in Greater Technical Detail

Why signal detection theory matters

Signal detection theory distinguishes sensitivity to evidence from where a person places a decision criterion. Two people can see the same evidence with similar quality yet favor different response thresholds. Confidence introduces further criteria: one person might reserve the word “certain” for exceptional cases while another frequently selects the highest confidence option. Meta-d′ was developed in part to reduce the dependence of a type 2 measure on such response tendencies and on differences in baseline performance. Its major strength is a principled mapping between patterns of confidence and the familiar first-order d′ scale (Maniscalco and Lau, 2012).

In practice, the model estimates first-order evidence distributions and confidence criteria consistent with the available data. A complete analysis should describe the response format, model variant, handling of zero cells, fitting approach, any pooling or regularization, and the uncertainty attached to estimates. The presence of executable software does not itself guarantee that assumptions hold in a particular experiment. Model checking and transparent reporting are part of the scientific result, not optional technical footnotes (Fleming, 2017).

Why meta-d′ and M-ratio can move in different directions

Consider a condition in which participants discriminate stimuli more accurately and, as expected, their confidence more strongly separates correct and incorrect decisions. Meta-d′ may rise simply because more informative first-order evidence is available. If meta-d′ and d′ increase proportionally, M-ratio could remain similar. An increase in sensitivity therefore need not imply a change in metacognitive efficiency. Conversely, a change in M-ratio may reflect changes in the relation between decision evidence and confidence evidence, though alternative explanations must be assessed. This is why scientists need both metrics and the underlying first-order task data rather than a single headline number (Fleming and Lau, 2014).

Why values above one are possible

M-ratio above one is not a logical contradiction. The second-order rating can sometimes use information that was unavailable to, neglected by, or accumulated after the initial choice. It can also be produced by sampling variability or misspecification of assumptions. The conventional ideal-observer reference value of one describes a particular model of how decision evidence becomes confidence evidence. It should not be read as a universal upper bound on observed human confidence discrimination or as proof of exceptional consciousness (Fleming, 2024; Rahnev, 2025).

Alternative Ways to Quantify Sensitivity

AUROC2, gamma, phi, and a simple confidence difference

AUROC2 summarizes the tradeoff between high-confidence correct decisions and high-confidence errors over different confidence cutoffs. Goodman–Kruskal's gamma represents a rank association between confidence and correctness and is familiar in metamemory research. Phi is a correlation between confidence and accuracy in settings where that coefficient is defined. A simple confidence gap subtracts average confidence on errors from average confidence on correct responses. These methods vary in statistical assumptions, interpretability, precision, dependence on task performance and response conventions. Researchers should not choose among them solely because one yields a larger standardized effect (Rahnev, 2025).

The choice of measure depends on the task and scientific question. Simple summaries can communicate an intuitive relation between confidence and accuracy, whereas a model-based SDT estimate may be preferable when the theoretical aim is to control particular performance and criterion confounds. AUC and correlation measures can be influenced by the mixture of easy and difficult trials, the distribution of response categories, and differences in first-order success. M-ratio brings additional model assumptions. Reporting a complementary set of measures can improve transparency when the measurements do not converge (Fleming and Lau, 2014; Rahnev, 2025).

Hierarchical Bayesian meta-d′

For individual-differences research, sparse trial counts can make independent participant estimates unreliable. Hierarchical Bayesian meta-d′, or HMeta-d, uses a population model to share information across people while estimating metacognitive efficiency. This can be valuable when a study compares groups, conditions or covariates with fewer trials per participant. The method is a statistical extension of the meta-d′ framework rather than a distinct psychological faculty. Shrinkage, prior choices and the assumption that the hierarchical model fits the sample influence interpretation. Investigators should explain whether parameters are estimated at the trial, participant or group level, and how uncertainty is propagated (Fleming, 2017).

Information-theoretic alternatives

Information theory offers ways to quantify how much confidence reduces uncertainty about the correctness of a decision. Dayan (2023) described a measure based on mutual information, often discussed as meta-I, along with normalized forms of efficiency. Later work proposed additional entropy- and divergence-based measures, and an empirical face-matching study evaluated their convergence (Fitousi, 2025). These approaches help address questions about the information contained in confidence scales, but they also require decisions about estimation, sample size, rating granularity and comparability. Their emergence is evidence of active measurement research rather than a settled replacement of meta-d′.

What Newer Research Reveals About Measurement Limits

No metric is assumption-free

A particularly consequential advance is Rahnev's 2025 comparative evaluation, which assessed 17 metacognitive measures. The analysis found a complex pattern: the measures exhibited validity under the study's framework, many had similar precision, and numerous measures were strongly dependent on first-order performance. No single measure was superior under every tested criterion. M-ratio remained a useful option, while alternatives including certain normalized correlation-based measures had similar performance in important dimensions. The empirical result calls for context-sensitive selection and more cautious claims about a universally best measure.

A separate simulation analysis found that M-ratio's apparent independence from task performance and decision criteria can depend on the generative model of confidence. Under several plausible confidence-generation models, values may shift when first-order performance or criteria shift, even without the intended metacognitive change (Rausch and colleagues, 2026). Another analysis demonstrated that deviations from the Gaussian evidence-distribution assumption can bias meta-d′ relative to d′ (Arnold and colleagues, 2023). These studies identify limitations of inference, not reasons to discard the entire framework.

Mixing difficulty levels can inflate results

A common psychophysical design uses an adaptive staircase to adjust stimulus difficulty as the participant progresses. This helps produce enough correct and incorrect decisions. However, a wide distribution of difficulty levels can itself strengthen the apparent relation between confidence and accuracy: very easy items tend to invite both higher confidence and more correct answers, even if the participant has limited insight into variation among items of equal difficulty. Reanalysis by Rahnev and Fleming (2019), followed by a published correction, showed that greater stimulus variability can increase estimates across several metacognition metrics. Therefore, difficulty distribution and staircase behavior should be specified, matched where relevant, and treated as potential contributors to apparent sensitivity.

Reliability differs from validity

A measure may detect meaningful confidence–accuracy relationships while producing unstable person-level rankings across days. In Rahnev's 2025 assessment, split-half reliability within a session was often high, while test–retest reliability was poor for most measures; normalized measures did not exceed 0.5 test–retest reliability even with 400 trials in the analyzed dataset. These results came from a particular dataset and require replication across protocols, yet they caution against treating a single test as a stable personal trait score. The article also notes that 50 trials yielded poor reliability and that roughly 100 trials could serve as a very approximate lower bound in some individual-differences settings, not a universal sample-size rule (Rahnev, 2025).

Confidence reports can change the task

Confidence judgments are behaviors influenced by instructions, response scales, time pressure, incentives, feedback, fatigue and the availability of further evidence. Confidence may be computed by partly shared mechanisms with the initial choice, by post-decision accumulation, or through broader self-models and expectations. The 2024 review by Fleming places propositional confidence within inferential models of one's own performance. There is consequently no single guarantee that two tasks using the word “confidence” measure exactly the same psychological operation.

Is Metacognitive Sensitivity a General Ability?

Perception, memory, reasoning, and executive tasks

Researchers measure confidence in visual perception, episodic memory, semantic memory and several kinds of executive decision. Correlations in sensitivity across tasks are relevant to whether metacognitive capacity includes a domain-general component or is primarily specific to particular cognitive processes. An investigation of 181 participants across four forced-choice tasks found significant associations in metacognitive efficiency for most task pairs when using a hierarchical Bayesian model, suggesting some shared influence across domains (Mazancieux and colleagues (2020)). This finding supports a domain-general component under those conditions, not an unrestricted, invariant personal “metacognition quotient.”

A later confirmatory study explicitly examined domain generality across visual perception and episodic and semantic memory, illustrating the continued importance of replicated cross-task designs (Lund and colleagues, 2025). Domain claims are sensitive to trial counts, performance matching, the reliability of the estimated metrics and whether all tasks use analogous confidence scales. A participant may show strong sensitivity on one task and weaker sensitivity on another without the difference representing a stable personal deficit.

Metamemory and prospective judgments

Metamemory refers to monitoring and evaluating one's own memory. Retrospective confidence following a recognition choice can be analyzed with meta-d′ when the response design supports it. Prospective judgments of learning ask how likely future recall will be; they compare predictions with later recall and often use other resolution and calibration statistics. These are related forms of metacognition, but they have different temporal targets and data structures. A good prospective judgment of whether material will be remembered should not automatically be interpreted as the same parameter measured by retrospective perceptual meta-d′ (Fleming and Lau, 2014; Fleming, 2024).

Brain Mechanisms and the Limits of Neural Interpretation

Studies of metacognitive judgments implicate distributed neural systems supporting evidence integration, decision confidence and monitoring of performance. Prefrontal and other networks have been investigated using structural MRI, fMRI, lesion designs and computational models. The review by Fleming and Dolan (2012) synthesizes early work connecting metacognitive ability to neural correlates, while Fleming (2024) integrates confidence computations with modern accounts of uncertainty and self-performance beliefs. Neural correlation with a score does not mean that the score reads a single brain region, isolates conscious experience or measures the presence of a subjective self.

Researchers distinguish neural processes carrying task evidence, updating confidence, detecting a potential error, and adjusting subsequent choices. Those processes can overlap in real behavior. A pattern of activity associated with revised confidence may support an explanation of how monitoring works, but assigning a particular psychological cause still requires converging experiments, appropriate controls and theory-specific predictions. This scientific separation prevents a computational estimator from becoming an unsupported claim about introspection or consciousness.

Applications: What a Sensitivity Measure Can and Cannot Tell Us

Learning and decision making

In principle, recognizing uncertain decisions can guide information seeking, reviewing an answer or asking for help. Experimental sensitivity captures one mechanism that could contribute to these choices. It does not prove that a participant will revise a mistaken belief, learn more effectively or choose a good strategy. Metacognitive control and downstream behavior must be measured separately. Findings from controlled discrimination tasks may not generalize to open-ended educational and professional choices without additional evidence (Fleming, 2024).

Human–AI decisions

When people consult AI recommendations, a confidence judgment about their own answer may help researchers understand whether they can distinguish their better from worse decisions. A 2025 paper in Lee and colleagues (2025) in PNAS Nexus discussed metacognitive sensitivity as relevant to calibrated trust in AI-assisted decisions. Its argument is a research application of human confidence measurement; a person's M-ratio cannot establish an AI system's consciousness, inner awareness, reliability across domains or entitlement to trust. Experimental study of human reliance must also independently measure actual advice quality, task success and reliance behavior.

Psychological and clinical research

Psychiatric studies sometimes compare confidence bias, discrimination and efficiency across symptom dimensions or groups. Such work may help characterize patterns of decision monitoring, but a metacognitive score is an experimental measure shaped by the protocol, not an individual diagnostic test. A review of confidence measures in psychiatric research stresses the distinction between confidence levels and precision (Hoven et al., 2019). Experimental differences between groups should be interpreted alongside measurement reliability, first-order ability, symptom heterogeneity, clinical assessment standards and replication. A low laboratory M-ratio on its own identifies no disorder.

Research Design Checklist: How to Read a Metacognition Paper

A credible paper states the primary decision task and objective definition of correctness; the number of trials and the distribution of correct and incorrect answers; stimulus difficulty and any adaptive staircase; the confidence prompt, response scale and timing; first-order accuracy and d′; the sensitivity statistic and why it was chosen; model assumptions and fitting method; uncertainty intervals; robustness to task-performance differences, sparse categories and response bias; and whether results concern a group contrast or a stable individual trait. These details tell a reader what the estimate actually means. Rahnev (2025) and Fleming and Lau (2014) are complementary methodological starting points.

When reading a reported “increase in metacognition,” ask whether it reflects higher confidence, better task accuracy, stronger trial-by-trial discrimination, or improved efficiency after accounting for first-order information. Ask whether the study used an identical confidence scale and matched difficulty. Ask whether an estimate obtained in one short session was independently replicated and whether a model-free measure supports the model-based one. Strong interpretation begins with these distinctions rather than with a single significant difference.

Common Misconceptions

A person with very high confidence does not necessarily have high metacognitive sensitivity. A person with high first-order accuracy does not necessarily possess a superior ability to locate their own mistakes. An M-ratio near one refers to a modeled relationship between first-order and confidence evidence; it does not demonstrate complete awareness of one's thinking. A low score is not a disorder or a clinical diagnosis. And a significant group difference need not imply a stable difference between the same two individuals in ordinary life. These statements follow from the conceptual and methodological distinctions elaborated in Fleming and Lau (2014) and the extensive 2025 assessment by Rahnev.

Frequently Asked Questions

What is metacognitive sensitivity in simple terms?

It is how well your confidence lets you tell which of your own answers were right and which were wrong. Someone whose confidence usually rises after correct decisions and falls after errors demonstrates stronger sensitivity than someone who reports similar confidence regardless of correctness. Scientific estimates require many objectively scored decisions and careful analysis rather than one or two striking anecdotes.

Is metacognitive sensitivity the same as metacognitive awareness?

No. Metacognitive awareness can refer broadly to noticing or knowing about cognitive states and strategies. Metacognitive sensitivity is a narrower experimental target: the relationship between a person's confidence reports and actual correctness across trials. The same task cannot by itself measure the entirety of self-awareness, attention, subjective consciousness or reflective understanding.

What is the difference between meta-d′ and d′?

d′ estimates first-order discrimination between stimulus alternatives within a signal detection framework. Meta-d′ expresses the discrimination that would be needed to produce the observed confidence-based separation of correct and incorrect decisions within a related type 2 model. One describes the primary task, while the other describes what confidence reveals about success on that task (Maniscalco and Lau, 2012).

What does meta-d′/d′ measure?

The ratio, usually called M-ratio, compares inferred metacognitive discrimination with first-order discrimination. It estimates efficiency relative to a model-based ideal-observer benchmark. It is informative when the task supports reliable estimation, but cannot be understood independently of d′, task difficulty, confidence criteria and assumptions of the fitting model (Rahnev, 2025).

Can M-ratio be greater than 1?

Yes. Additional evidence available after the choice, sampling variability and violations of the particular model's assumptions can contribute to estimates above one. Such a score is not proof of perfect introspection or an impossible psychological ability. The result calls for inspection of the task timeline and model fit.

Is sensitivity different from confidence calibration?

Yes. Calibration asks whether stated probabilities agree with how often answers are correct. Sensitivity asks whether varying confidence separates a person's correct decisions from errors. A person can be systematically overconfident yet still rank their decisions very well, and calibration by itself does not guarantee strong discrimination.

Does higher metacognitive sensitivity mean higher IQ?

The metric targets confidence-based monitoring of performance, not general intelligence. It can vary across tasks and measurement contexts. Research should test relationships with other traits directly, using reliable measures and adequate samples rather than assuming that two construct labels are equivalent.

Can I measure my own metacognitive sensitivity with a quick quiz?

A few subjective questions cannot provide a dependable meta-d′ estimate. Research requires objectively scored trials, repeated confidence reports, enough errors, appropriate modeling and quantified uncertainty. A short personal quiz may help someone notice habits of certainty, but it cannot establish a scientifically validated personal metacognitive efficiency score (Rahnev, 2025).

Can metacognitive sensitivity be trained?

Some studies examine whether monitoring or confidence discrimination changes after feedback and practice. Whether an intervention improves M-ratio specifically, generalizes between tasks, and persists over time are separate empirical questions. Evidence for a strategy improving task accuracy does not establish that it improves metacognitive efficiency. Claims should identify the studied intervention, outcome and follow-up window rather than promising a global training effect (Fleming, 2024).

Does a low metacognitive sensitivity score indicate a mental health condition?

No. A score from a confidence task is a research measurement, not a screening diagnosis. It may depend on fatigue, task difficulty, interpretation of the response scale and estimation error. When researchers investigate clinical populations, they need independent diagnostic procedures and careful controls. No one should use an online meta-d′ result to diagnose themselves or another person.

Related Articles

The wider construct and monitoring-versus-control distinction are developed in Metacognition: What It Is and How Thinking About Thinking Works and Metacognitive Monitoring vs. Control: Key Processes. For the separate question of confidence bias and correctness, see Metacognitive Confidence vs. Accuracy: Why Certainty Can Mislead. The broader context of applied AI-supported reflection is addressed in Metacognition in the Age of AI: How to Monitor Your Thinking When AI Helps You Think. These links concern distinct published English search intents.

References

Arnold, D. H., Johnston, A., Adie, J., & Yarrow, K. (2023). On why we lack confidence in some signal-detection-based analyses of confidence. Consciousness and Cognition, 113, 103532. https://doi.org/10.1016/j.concog.2023.103532

Dayan, P. (2023). Metacognitive information theory. Open Mind, 7, 392–411. https://doi.org/10.1162/opmi_a_00091

Fitousi, D. (2025). Information-theoretic measures of metacognitive efficiency: Empirical validation with the face matching task. Entropy, 27(4), 353. https://doi.org/10.3390/e27040353

Fleming, S. M. (2017). HMeta-d: Hierarchical Bayesian estimation of metacognitive efficiency from confidence ratings. Neuroscience of Consciousness, 2017(1), nix007. https://doi.org/10.1093/nc/nix007

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., & Dolan, R. J. (2012). The neural basis of metacognitive ability. Philosophical Transactions of the Royal Society B, 367, 1338–1349. https://doi.org/10.1098/rstb.2011.0417

Fleming, S. M., & Lau, H. (2014). How to measure metacognition. Frontiers in Human Neuroscience, 8, 443. https://doi.org/10.3389/fnhum.2014.00443

Hoven, M., Lebreton, M., Engelmann, J. B., Denys, D., Luigjes, J., & van Holst, R. J. (2019). Abnormalities of confidence in psychiatry: An overview and future perspectives. Translational Psychiatry, 9, 268. https://doi.org/10.1038/s41398-019-0602-7

Lee, D., Pruitt, J., Zhou, T., Du, J., & Odegaard, B. (2025). Metacognitive sensitivity: The key to calibrating trust and optimal decision making with AI. PNAS Nexus, 4(5), pgaf133. https://doi.org/10.1093/pnasnexus/pgaf133

Lund, A. E., et al. (2025). Domain generality in metacognitive ability: A confirmatory study across visual perception, episodic memory, and semantic memory. Journal of Experimental Psychology: Learning, Memory, and Cognition. https://doi.org/10.1037/xlm0001462

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

Mazancieux, A., Fleming, S. M., Souchay, C., & Moulin, C. J. A. (2020). Is there a G factor for metacognition? Correlations in retrospective metacognitive sensitivity across tasks. Journal of Experimental Psychology: General, 149(9), 1788–1799. https://doi.org/10.1037/xge0000746

Rahnev, D. (2025). A comprehensive assessment of current methods for measuring metacognition. Nature Communications, 16, 701. https://doi.org/10.1038/s41467-025-56117-0

Rahnev, D., & Fleming, S. M. (2019). Correction: How experimental procedures influence estimates of metacognitive ability. Neuroscience of Consciousness, 2019(1), niz010. https://doi.org/10.1093/nc/niz010

Rahnev, D., & Fleming, S. M. (2019). How experimental procedures influence estimates of metacognitive ability. Neuroscience of Consciousness, 2019(1), niz009. https://doi.org/10.1093/nc/niz009

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