A person answers a question correctly and still feels unsure.
Another gives the wrong answer with complete confidence.
A third knows that they usually make mistakes in this domain and lowers their confidence before any external feedback arrives.
Accuracy and confidence are related, but they are not the same variable.
Stephen M. Fleming’s 2024 Annual Review, “Metacognition and Confidence: A Review and Synthesis,” is important because it refuses to treat confidence as a decorative rating added after cognition has already happened. Instead, it asks how confidence is computed, what information it uses, how it becomes available at the personal level and why it can diverge from objective performance.
The article is a review and theoretical synthesis. It does not report one new randomized experiment, and it is not a meta-analysis producing a pooled effect size. Its value lies in connecting research traditions that often use the same words for different levels of explanation.
For the Institute’s Knowledge Library, that distinction is essential. A source-first brief should tell the reader not only what a paper argues, but what kind of evidence can legitimately support that argument.
The problem the review is trying to solve
Research on uncertainty exists at several levels.
Sensory and motor neuroscience can describe subpersonal representations of uncertainty inside a perceptual or action system. Experimental psychology can ask participants to report confidence in a choice. Metacognition research can examine how well those confidence judgments track correctness. Clinical and educational research can ask how people form beliefs about their own abilities.
These literatures overlap, but they are not interchangeable.
Fleming’s review proposes a bridge by focusing on propositional confidence: confidence in the correctness of one’s own hypothetical or actual decisions and actions at a level that can inform deliberate reasoning and report.
This matters because a personal confidence judgment may depend on more than the evidence that drove the first-order decision.
A person can use memory of past performance, beliefs about task difficulty, expectations about their own competence, contextual cues, social information and post-decisional evidence. Confidence is therefore inferential.
The same architecture that makes confidence useful also makes miscalibration possible.
Confidence is built from a model, not merely read from a meter
A simple intuition would say that the brain makes a decision and then reads out how strong the evidence was.
That can be part of the story.
But it does not explain why confidence can be systematically biased, why people can learn that their own confidence is unreliable in a domain, or why global beliefs about ability can shape local judgments.
Fleming’s synthesis emphasises that personal-level confidence is informed by models of the world and of the cognitive system itself.
The self-model may be more or less accurate.
Someone who has repeatedly failed at mental rotation may sensibly discount an otherwise strong feeling of certainty on a new spatial problem. Someone who has been rewarded for fluent answers may overvalue fluency even when fluency is a poor cue to accuracy. Someone working in a familiar domain may use prior knowledge to interpret ambiguous evidence differently from a novice.
Confidence is thus neither pure introspection nor simple external inference. It is a constructed judgment over evidence that includes beliefs about the agent making the judgment.
Why divergence from performance is theoretically useful
Miscalibration is often discussed as noise or failure.
Fleming’s framework makes divergence more informative.
If confidence and performance can dissociate, the dissociation reveals that confidence has its own computational inputs and transformations. The question shifts from “Why did the confidence system fail to copy accuracy?” to “What evidence and self-model produced this confidence judgment?”
That shift is scientifically productive.
It also prevents a common educational mistake. Learners do not merely need confidence to increase. They need confidence to become better aligned with what they know, what they do not know and how stable that knowledge will be under future conditions.
High confidence can be useful when it is earned. Low confidence can be rational when evidence is weak. The target is not confidence maximization.
The target, where relevant, is appropriate confidence.
The paper links personal and subpersonal uncertainty without collapsing them
A major ambition of the review is to connect work on neural and computational representations of uncertainty with explicit metacognitive judgment.
The bridge is not a claim that one brain signal simply is confidence.
Instead, the paper treats lower-level uncertainty signals as possible ingredients available to a higher-level inferential system. Those signals can be integrated with additional information, including beliefs about task structure and self-performance.
This helps explain why experiments sometimes identify neural signals associated with uncertainty before a person makes an explicit confidence report, while other research implicates frontal or parietal systems in metacognitive evaluation and control.
Different signals may participate at different stages.
The literature does not support a single “confidence centre.”
That is an important evidence boundary for popular neuroscience.
Confidence can influence what happens next
Metacognition is not only monitoring.
Confidence can guide control.
A person who is uncertain may gather more information, reconsider a decision, allocate more study time or defer to another source. A person who is confident may stop searching and act.
The quality of confidence therefore affects downstream behaviour.
This is one reason miscalibration has consequences beyond subjective feeling. Overconfidence may terminate inquiry too early. Underconfidence may lead to unnecessary checking, deference or avoidance. Well-calibrated confidence can support efficient allocation of cognitive effort.
Yet the causal chain is context-dependent. More checking is not always better. Search has costs. Some environments reward speed. Others punish false certainty. A rational control policy depends on what is at stake and what additional evidence is available.
The review’s contribution is to make confidence part of a control architecture rather than an isolated self-report score.
Intervention is not the same as boosting confidence
The paper also helps sharpen what a metacognitive intervention would need to change.
If confidence is inferential, an intervention can act at several loci.
It might improve the quality of first-order evidence. It might teach a person which cues are diagnostic. It might change beliefs about their own competence. It might improve mappings between uncertainty and confidence reports. It might change what a person does after noticing uncertainty.
Those are not the same intervention.
A programme that raises confidence after successful practice may improve self-efficacy without improving metacognitive sensitivity. Feedback can improve calibration in one task without creating a domain-general metacognitive ability. Training people to use confidence scales consistently can improve measurement without changing underlying monitoring.
Fleming’s review makes these distinctions easier to see.
Measurement remains a central problem
The review sits next to an expanding literature on how metacognition should be measured.
Confidence ratings are observable reports. Researchers can transform them into calibration scores, discrimination measures, type-2 ROC metrics, meta-d-prime, M-ratio and other quantities. Each operationalization makes assumptions.
Rahnev’s 2025 systematic empirical comparison of 17 metacognitive measures, published after Fleming’s review, reinforces this point. Measures can differ in dependence on task performance, precision and reliability even when they are all called measures of metacognition.
This does not undermine the 2024 review.
It strengthens one of its implicit lessons: theories of confidence and measurements of confidence must be kept conceptually connected but methodologically distinct.
A theory can be sophisticated while a particular score is noisy.
Global self-beliefs and local confidence should not be merged
People also hold broad beliefs such as “I am good at mathematics” or “I am bad at remembering names.”
These global beliefs can influence local confidence, but a global self-assessment is not the same as a trial-level confidence judgment.
Global beliefs aggregate across memories, social comparison, identity and history. Local confidence is attached to a particular decision or anticipated action.
The relationship can run both ways. Repeated local errors may update the global self-model. The global model can then bias future local confidence.
For education, this creates a delicate problem. A learner may be accurately uncertain about one problem while carrying an inaccurate global identity such as “I am not a science person.” Correcting the identity and calibrating the immediate judgment require different evidence and different interventions.
The review offers a vocabulary for separating them.
Why this matters for AI-assisted cognition
The paper predates some of the strongest 2025 and 2026 evidence on everyday generative-AI use, but its framework is highly relevant to it.
When an AI system produces a fluent answer, the user must estimate not only whether the output looks plausible but whether they are in a position to verify it.
That second judgment is metacognitive.
A novice may be uncertain about the answer but overconfident in their ability to detect hallucinations. An expert may distrust a correct output because they know the model’s failure modes. Repeated successful assistance may alter a user’s self-model, making them feel more capable even if unaided performance has not changed.
The Institute’s synthesis is that AI literacy requires a confidence architecture, not merely a fact-checking checklist. Users need some capacity to distinguish confidence in the tool, confidence in the output, confidence in their own verification and confidence in the final decision.
This is an application of the review’s logic, not a direct claim tested by Fleming’s paper.
The review is not evidence that confidence is a unitary trait
Because the paper integrates multiple literatures, it can be misread as establishing one stable metacognitive faculty.
It does not.
Metacognitive ability varies across tasks and domains. Measurement reliability can be limited. Confidence can depend on local evidence and learned priors. Different components may generalize differently.
The existence of a useful theoretical framework does not guarantee a single latent trait that can be measured once and used to rank people.
This is especially important for organisations tempted to build “confidence accuracy” assessments into leadership or talent systems.
The literature does not justify that leap without population-specific validity evidence.
Accuracy, bias and sensitivity are different questions
One reason confidence research becomes confusing is that several desirable properties are often folded into the word “accuracy.”
A person can be generally underconfident or overconfident. That is a bias question. A person can also give higher confidence to correct answers than to incorrect answers even while being globally overconfident. That is a discrimination or sensitivity question. Calibration asks how stated confidence corresponds to observed correctness, but the answer depends on whether one is examining absolute agreement, relative discrimination, or efficiency after controlling for task performance.
The review’s value is partly that it resists turning these into one psychological virtue. Different computational and behavioural measures capture different aspects of the monitoring problem. Later measurement work, including Rahnev’s 2025 comparison of 17 metacognitive measures, makes the same caution concrete: apparent metacognitive quality depends in part on the metric chosen and on first-order task performance.
For readers outside the laboratory, this distinction prevents a common error. Someone who says “I’m only 60% sure” and is correct may be modestly calibrated, poorly calibrated or simply cautious depending on the pattern across many judgments. A single confidence statement cannot diagnose a stable capacity.
Confidence also coordinates social action
Confidence is not only private information about one’s own cognition. It is also something people display to one another.
Teams decide whose recommendation to follow partly from expressed certainty. Teachers use student confidence to decide whether to probe or advance. Clinicians and analysts may communicate uncertainty so another person can decide how much verification is needed. In these contexts, confidence becomes a coordination signal.
That creates a second layer of calibration. A person must not only monitor their own uncertainty but translate it into language another person can interpret. Social incentives can distort that translation. Some environments reward certainty; others punish it. Status can make hesitant but accurate people less influential than confidently wrong ones. Cultural norms change how certainty is expressed.
The focal review is not a social-institutions paper, so this extension should be labelled as synthesis. Yet the implication follows naturally from treating confidence as a propositionally structured representation rather than a raw feeling. Once confidence is communicated, institutions begin acting on it.
For the Institute, this is one reason metacognitive literacy belongs in organisational design as well as individual training. A healthy system needs ways to say “I know,” “I suspect,” and “I do not know yet” without making all three socially equivalent or morally loaded.
Intervention should target the monitoring-control loop
If confidence is imperfect, the tempting intervention is to make people less confident.
That is not the goal.
Underconfidence can be costly too. A well-calibrated learner may abandon a correct strategy if uncertainty is treated as failure. A professional may over-verify routine decisions and exhaust limited attention. The more defensible target is the coupling between evidence, confidence and action.
Useful interventions therefore ask whether confidence changes appropriately when evidence changes, whether low confidence triggers proportionate checking, and whether high confidence can still be revised when contradictory evidence arrives. The desired outcome is not a particular emotional tone. It is a better monitoring-control relationship.
What the evidence does not show
The review does not show that confidence comes from one brain region.
It does not establish that all uncertainty signals become conscious confidence.
It does not show that self-reported confidence is always accurate, nor that low confidence is inherently bad.
It does not prove that metacognitive training transfers broadly across domains.
It does not establish that one metacognitive metric is universally superior.
And because it is a review and synthesis, it should not be cited as though it were one new experiment demonstrating a causal intervention effect.
What changed with this paper
The contribution is architectural.
Fleming provides a way to connect confidence in a choice with models of the world and of the self, while preserving distinctions between subpersonal uncertainty and personal-level judgment.
That changes the question researchers and practitioners can ask.
Instead of treating confidence as a number attached to an answer, they can ask where the number came from, which evidence informed it, what model of self shaped it, how reliable the report is and what control decision follows from it.
That is a richer scientific object.
It is also a more humane one.
A person who says “I’m not sure” is not displaying a deficit by default.
They may be showing one of the most useful capacities a cognitive system has: the ability to represent the limits of its own certainty.
Research references
Fleming SM. Metacognition and Confidence: A Review and Synthesis. Annual Review of Psychology. 2024;75:241–268. https://doi.org/10.1146/annurev-psych-022423-032425
Fleming SM, Lau HC. How to measure metacognition. Frontiers in Human Neuroscience. 2014;8:443. https://doi.org/10.3389/fnhum.2014.00443
Maniscalco B, Lau H. A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition. 2012;21(1):422–430. https://doi.org/10.1016/j.concog.2011.09.021
Rahnev D. A comprehensive assessment of current methods for measuring metacognition. Nature Communications. 2025;16:701. https://doi.org/10.1038/s41467-025-56117-0