An answer can arrive before certainty does.
A person chooses one option, then pauses over a second question: "How confident are you?"
That second response is a confidence judgment.
Confidence judgments are explicit estimates of the probability or credibility a person assigns to their own decision, memory, perception or answer. They are among the most widely used observable reports in metacognition research because they make subjective evaluation available trial by trial.
Their usefulness depends on a crucial fact: confidence and correctness are related, but they are not the same variable.
Confidence is a personal-level inference
Stephen Fleming's 2024 review draws an important distinction between personal-level confidence and subpersonal representations of uncertainty. Neural or computational systems may represent uncertainty without a person forming an explicit proposition such as "I am 70 percent confident that this answer is correct."
A confidence judgment is therefore more than raw sensory evidence. It can incorporate beliefs about the task, prior knowledge, decision evidence, response time, memory cues and models of one's own reliability.
This helps explain why confidence can be informative and still depart from performance.
The person is not reading accuracy directly from the world. They are constructing an estimate from evidence about the decision and about themselves.
High confidence is not the same as high metacognitive ability
Two people can have the same task accuracy and very different average confidence. One may use the top of the scale routinely. The other may be conservative.
That difference is often called metacognitive bias or confidence bias. It should not be confused with metacognitive sensitivity, which asks whether confidence discriminates correct from incorrect responses.
A chronically overconfident person can still show meaningful trial-by-trial discrimination. A cautious person can have poor discrimination.
This distinction matters in journalism, education and professional evaluation because "confident" is often treated as a proxy for "knows when they are right." Research does not license that shortcut.
Calibration asks a different question again
Calibration concerns the match between stated confidence and observed accuracy.
If decisions assigned 80 percent confidence are correct about 80 percent of the time across an appropriate set of cases, the judgments are well calibrated at that level. A person who says 90 percent while being correct 70 percent is overconfident in an absolute sense.
Calibration is especially useful when confidence is meant to guide action thresholds.
But calibration and resolution can diverge. A person may be roughly calibrated on average yet poor at distinguishing stronger from weaker individual decisions. Another may rank their decisions well but use a confidence scale with systematic bias.
The term "confidence accuracy" is therefore too vague unless the metric is specified.
Confidence can be shaped by cues that have nothing to do with truth
Confidence often tracks evidence strength, but it can also be influenced by processing fluency, familiarity, response time and contextual cues.
This is not irrational by definition. In many environments these cues correlate with accuracy. The problem arises when the environment changes and the cue becomes less diagnostic.
A familiar name may feel more credible. A fluent explanation may feel better understood. A fast response may feel like expertise. Repeated exposure can increase familiarity without increasing truth.
Metacognitive research is valuable partly because it reveals these cue dependencies instead of treating confidence as a transparent readout.
Measurement can alter decision-making
Confidence reports are often treated as if they merely reveal an internal state that already exists.
That assumption is not always safe.
Double and Birney's 2024 Scientific Reports paper reported three experiments in which contemporaneous confidence judgments impaired perceptual decision accuracy. Other work has found mixed reactivity effects, so the correct conclusion is not that confidence ratings always harm performance.
The methodological point is narrower and stronger: eliciting a judgment can change the cognitive process being studied.
A research design, classroom tool or decision interface that repeatedly asks for confidence should therefore consider whether the prompt is measurement, intervention or both.
Confidence scales are not self-interpreting
Researchers use percentages, verbal categories and ordinal scales. Each format carries assumptions.
Does "70 percent confident" mean subjective probability of correctness, degree of evidence, willingness to bet, or merely a point on a familiar slider? Do different participants use the scale similarly? Does the end of the scale feel socially acceptable?
Signal-detection approaches such as meta-d prime aim to extract metacognitive sensitivity from patterns of confidence and accuracy. Calibration curves evaluate another aspect. Mean confidence measures bias.
No single score answers every confidence question.
Rahnev's 2025 assessment of 17 metacognitive measures reinforces this pluralism. Different measures have different dependencies and reliability properties. Choosing a metric is part of the substantive claim, not a technical afterthought.
Confidence communicates uncertainty socially
Confidence is not only private.
People use expressed confidence to decide whose advice to follow, whether consensus is strong, and whether a decision should be reopened. Groups can benefit when confidence carries information about accuracy. They can also be misled when social dominance, style or incentives distort the signal.
The social role creates a second calibration problem. A person can be internally calibrated yet communicate certainty strategically. Another may have accurate private doubt but understate confidence because of status or culture.
Institutional procedures that rely on expressed confidence need to distinguish epistemic information from social performance.
This is particularly important where speaking with certainty is rewarded.
Expertise does not eliminate confidence error
Expertise often improves first-order performance because experts possess richer representations and domain knowledge. It can also improve the cues available for metacognitive judgment.
But expertise does not create perfect self-knowledge.
Complex domains have incomplete feedback, changing base rates and rare cases. Confidence can be miscalibrated even when average performance is high. In some settings, specialists may be confident for good reasons but still wrong because the case lies outside ordinary experience.
The practical implication is not to distrust expertise. It is to avoid turning confidence into a substitute for outcome data, external validation or explicit uncertainty communication.
AI creates confidence without a human confidence state
Generative models often produce assertive language, caveats or numerical scores. It is tempting to treat these as directly analogous to human confidence.
The analogy is risky.
A model's textual certainty style, a probability derived from token distributions, and a person's propositional confidence are different objects. Even when a system exposes a calibrated uncertainty estimate for a particular task, the semantics need to be specified.
For human users, AI output can also alter their own confidence. Fluent explanations may make an answer feel more credible. Direct evidence on these complex interactions is growing but remains task dependent.
The Institute should therefore distinguish model uncertainty, expressed linguistic certainty, human confidence, and actual correctness rather than collapsing them into one certainty dimension.
Confidence can guide control when its meaning is clear
The reason confidence matters is not only measurement. It can regulate behaviour.
Low confidence can trigger information seeking, checking or deferral. High confidence can justify stopping. In team settings, confidence can influence weighting of advice. In learning, judgments of confidence can determine what is restudied.
These functions depend on calibration and on cost.
If checking is expensive, an adaptive threshold may tolerate some uncertainty. If errors are catastrophic, the threshold may be more conservative. Confidence becomes useful when it is connected to the consequences of being wrong.
That is a control problem, not simply a feeling.
Confidence can be local or global
A person can be confident about one answer and uncertain about their overall ability. These levels should be kept separate.
Local confidence concerns a particular decision or memory. Global confidence concerns a broader self-assessment such as "I am good at statistics" or "I understand this chapter." The two influence one another, but they are not interchangeable.
Global beliefs can bias local reports. Repeated local outcomes can update global self-beliefs. Educational feedback often operates across both levels.
This matters when institutions ask broad confidence questions and interpret them as if they were trial-level calibration data. A survey item about confidence in leadership, for example, does not reveal how well a person discriminates correct from incorrect judgments in actual decisions.
Social confidence can become detached from epistemic confidence
People do not report certainty in a vacuum.
They speak to audiences, negotiate status and manage impressions. A person may sound more certain than they feel because the role demands decisiveness. Another may hedge a well-supported claim because social norms punish assertiveness.
For group decision-making, this means expressed confidence can be a noisy channel for private confidence.
Procedures can reduce the distortion. Anonymous forecasts, independent initial estimates and explicit probability ranges can sometimes separate epistemic information from performance style. They do not eliminate every bias, and their effectiveness depends on context.
The broader lesson is that confidence has both cognitive and communicative functions.
Confidence has consequences even when it is wrong
Confidence can determine whether a person checks, whether others follow them, whether a learner stops studying, and whether a team escalates a decision.
A miscalibrated confidence judgment therefore changes more than a score on a questionnaire. It changes the trajectory of action.
This is why metacognitive control and confidence belong close together in the knowledge graph. The value of confidence is partly instrumental: it allocates attention, verification and social trust.
A system that improves first-order accuracy but produces badly distorted confidence can still create risk if people no longer know which outputs deserve scrutiny.
That principle is especially relevant to automated decision support, where human users may become more accurate on average while becoming less able to identify the remaining errors.
Confidence feedback needs outcome quality
Calibration feedback sounds straightforward: show people whether they were right. In complex domains, "right" may not arrive cleanly.
A policy outcome has many causes. A medical diagnosis may be revised months later. A creative decision may have multiple defensible standards. A strategic forecast can be affected by the very action taken because of the forecast.
This means confidence learning is easiest in environments with repeated, interpretable outcomes and harder where feedback is sparse or endogenous.
For institutions, the temptation is to manufacture certainty by scoring what is easy to score. That can produce a beautifully measured proxy and leave the consequential judgment untouched.
Confidence systems should therefore begin with the quality of the outcome variable. A precise confidence number paired with an ambiguous outcome is not precise metacognition.
Confidence should sometimes remain qualitative
Numerical probabilities can improve clarity, but forcing every judgment into a number can create false precision. In domains where the evidence base is weak or the event is poorly specified, a range, scenario set or explicit statement of unknowns may be more honest.
The scientific lesson is not "always quantify certainty." It is to represent uncertainty in a form matched to what can actually be known.
What the evidence does not show
High confidence does not prove correctness.
Low confidence does not prove error.
Average confidence is not the same as metacognitive sensitivity.
Calibration is not the same as resolution.
One confidence scale is not a universal measurement of self-knowledge.
A confidence report may alter the decision process rather than simply reveal it.
Laboratory confidence measures do not justify labeling people globally "well calibrated" across domains.
And human confidence research does not mean that a language model's confident phrasing is a psychological confidence state.
Open questions
How stable are confidence biases across tasks and time? Which components of confidence generalise across domains? When does eliciting confidence improve reflection and when does it produce harmful reactivity? How do social incentives reshape expressed certainty? Which forms of feedback improve calibration without encouraging defensive certainty management?
Human-AI systems add further questions. When should a tool expose uncertainty? Which uncertainty representation helps users rather than becoming another misleading cue? How does model explanation affect human confidence independent of correctness? Can people learn when AI fluency is diagnostic and when it is not?
These are empirical questions. Their answers will differ by domain and design.
The second number
The first number is the answer.
The second is how much trust the person places in it.
Confidence judgments make that second number visible.
At their best, they allow cognition to allocate checking, learning and social trust. At their worst, they give style the appearance of evidence.
For individuals, the challenge is not to become less confident. It is to make confidence answerable to outcomes.
For institutions, the challenge is to create conditions in which uncertainty can be expressed without becoming status loss.
For a civilization increasingly surrounded by fluent machine output, the challenge is to remember that certainty is a signal whose value depends on calibration.
Confidence is not truth.
It is a claim about how much truth a decision believes it has earned.
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
- 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
- Double KS, Birney DP. Confidence judgments interfere with perceptual decision making. Scientific Reports. 2024;14:14133. https://doi.org/10.1038/s41598-024-64575-7
- Koriat A. Monitoring one's own knowledge during study: A cue-utilization approach to judgments of learning. Journal of Experimental Psychology: General. 1997;126(4):349-370. https://doi.org/10.1037/0096-3445.126.4.349
- 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
- Lin LM, Zabrucky KM. Calibration of Comprehension: Research and Implications for Education and Instruction. Contemporary Educational Psychology. 1998;23(4):345-391. https://doi.org/10.1006/ceps.1998.0972