A person makes a decision and then adds one more piece of information.
“I think that was right.”
Or: “I chose it, but I am not sure.”
That second judgment is the territory at the centre of Stephen M. Fleming's research programme. Fleming is Professor of Cognitive Neuroscience at University College London and leads UCL's MetaLab, which studies the computational and neural bases of subjective experience, self-awareness, metacognition and confidence.
The programme is important because confidence is often treated casually. In ordinary life it is read as charisma, certainty, decisiveness or self-belief. In metacognition research, confidence can be operationalised much more precisely: how well does a person's subjective certainty track the quality of their own decisions?
That shift, from confidence as style to confidence as measurable self-evaluation, has helped connect cognitive psychology, signal detection theory, computational modelling and neuroscience.
The central problem is correspondence, not confidence alone
A person can be highly confident and frequently wrong. Another can be cautious and highly accurate. Mean confidence therefore tells us little by itself about metacognitive quality.
The more revealing question is whether confidence discriminates between a person's correct and incorrect decisions. Does the person tend to be more confident when right than when wrong? Can that relationship be estimated independently of how easy the underlying task is?
Fleming's work repeatedly returns to this separation between first-order performance and second-order evaluation.
A perceptual task might ask which of two intervals contains a visual target. Accuracy on that task reflects first-order perceptual sensitivity. A confidence judgment about the choice provides a second-order report. The relation between the two is one route into metacognitive sensitivity.
This architecture matters because people with the same task accuracy can differ in how well they know when they are likely to be correct.
Early neuroscience work made individual differences visible
In 2010, Fleming and colleagues published a Science paper relating introspective accuracy to individual differences in brain structure. Participants performed a perceptual decision task and rated confidence. The researchers found correlations between introspective accuracy and structural properties of anterior prefrontal cortex.
The study was influential because it linked behavioural variation in metacognition to neuroanatomical variation.
It also illustrates a boundary that needs to remain explicit. The result was correlational. It did not establish that one brain region is a single “centre of introspection,” nor that structural differences cause metacognitive ability in a simple one-directional way.
UCL's own account described a correlation between introspective ability and the structure of a small prefrontal area. That is the appropriate level of claim.
Researcher profiles should preserve that restraint rather than letting memorable brain findings become stronger with each retelling.
Measurement became as important as mechanism
A recurring problem in metacognition is that apparent differences in self-evaluation can be contaminated by differences in task performance.
If one participant sees the stimuli more clearly than another, their confidence has better evidence to work with. A simple confidence-accuracy correlation can therefore mix first-order sensitivity with second-order monitoring.
Brian Maniscalco and Hakwan Lau's 2012 meta-d-prime framework addressed this problem using signal detection theory. Fleming and Lau's 2014 review, “How to measure metacognition,” helped systematise the landscape of metacognitive measures and their confounds for a broader research community.
Later, Fleming developed hierarchical Bayesian methods for estimating metacognitive efficiency across individuals and groups, helping with situations in which trial counts are limited and estimates are noisy.
This sequence is revealing. The field did not progress only by discovering more facts about confidence. It progressed by getting stricter about what a confidence measure can legitimately mean.
Meta-d-prime is part of the programme, not a magic number
The term meta-d-prime often appears wherever Fleming's work is discussed, but attribution needs precision.
The original meta-d-prime method was introduced by Maniscalco and Lau in 2012. Fleming contributed substantially to its adoption, interpretation and extension, including hierarchical Bayesian estimation in HMeta-d.
Meta-d-prime expresses metacognitive sensitivity in signal-detection units, allowing comparison with first-order d-prime. A ratio such as meta-d-prime divided by d-prime can be interpreted as metacognitive efficiency under specified model assumptions.
The value of the approach is not that it makes introspection perfectly measurable. It is that it tries to separate metacognitive information use from underlying task sensitivity more carefully than simpler measures often can.
The limitations remain substantial: model assumptions matter, confidence scale design matters, trial counts matter, and efficiency estimates can be unstable in some conditions.
The dedicated Knowledge Library method page addresses those issues directly.
Confidence is inferential
Fleming's 2024 Annual Review of Psychology synthesis develops a broader theoretical point.
Confidence is not necessarily a transparent readout of an internal certainty variable. It can be understood as an inference informed by models of the world and of one's own cognitive system.
That helps explain why confidence can diverge from performance. People do not simply “feel the probability of being correct.” They use internal evidence, prior beliefs, task knowledge and self-models to arrive at a report.
Fleming distinguishes subpersonal representations of uncertainty from personal-level confidence judgments. The distinction prevents a common reduction: finding uncertainty-related neural activity does not automatically explain the conscious proposition “I am probably wrong.”
The bridge between levels needs a model.
This is one reason his work is especially relevant to the Institute's interest in metacognitive literacy. Confidence is not merely an emotion attached to a decision. It is evidence about how a cognitive system is representing its own reliability, and that evidence can itself be biased.
Better monitoring does not mean omniscience
The phrase “metacognitive ability” can sound grander than the tasks used to estimate it.
A participant may show good confidence-accuracy correspondence in visual perception and weaker correspondence in memory or reasoning. Domain generality is an empirical question, not a default assumption.
Measurement also creates a reactivity problem. Asking for confidence may change the way a person performs the first-order task. Recent work by Double, Goldwater and Birney found that eliciting confidence ratings altered category-learning strategy and impaired rule learning in a preregistered study.
That means the report used to observe metacognition can sometimes perturb the cognition being observed.
This does not invalidate confidence research. It makes experimental design more demanding.
The 2025 measurement comparison complicates simple rankings
Dobromir Rahnev's 2025 Nature Communications assessment compared 17 measures of metacognition using a common evaluation framework.
The paper found that all 17 measures passed the proposed validity test and that many had similar precision, while also showing important complications. Many measures depend on first-order performance. Split-half reliability was generally high, but test-retest reliability was poor for most measures in the datasets examined.
The practical implication is not that one metric has won.
It is that measurement choice should follow the research question, task design, model assumptions and available data.
This complements rather than displaces Fleming's methodological programme. A maturing field learns to compare its tools rather than treating a favourite index as the construct itself.
Neuroscience is most useful when it constrains theory
Fleming's laboratory combines behaviour, computational models and neuroimaging. That combination can tempt public summaries toward localisation stories: a region “does” confidence, self-awareness or introspection.
The stronger scientific contribution is more relational.
Neural data can help distinguish models of where confidence-related information becomes available, how first-order and metacognitive computations interact, and why particular clinical conditions alter self-evaluation.
UCL's MetaLab also studies implications for conditions involving distorted self-awareness and confidence, including affective disorders, dementia and psychosis, in collaboration with clinical researchers.
A profile should not convert these research interests into clinical claims. Studying a mechanism in a clinical population is not the same as having a validated treatment or diagnostic tool.
Europe is not a decorative angle here
Fleming's programme is itself part of a major European research environment.
The MetaLab sits at UCL's Institute of Cognitive Neuroscience and Department of Experimental Psychology and is affiliated with the Max Planck-UCL Centre for Computational Psychiatry and Ageing Research. This provides a substantive European authority connection rather than an added geographic citation.
The research also sits within a broader European tradition in computational psychiatry, consciousness science and metacognition, where questions about subjective report are linked to formal modelling and clinical neuroscience.
For the Institute's Northern European knowledge architecture, this is exactly the kind of regional connection that belongs because it changes the intellectual map, not because a quota requires it.
Confidence research matters beyond the laboratory
Modern institutions ask people for confidence constantly.
Clinicians estimate diagnostic certainty. Analysts assign probabilities. Forecasting teams communicate confidence intervals and subjective probabilities. Students predict whether they know an answer. AI systems may display model confidence or uncertainty proxies.
These contexts differ enormously, but they share a structural question: what relationship does the confidence report have to the underlying accuracy of the judgment?
Fleming's work does not provide one calibration method for all of them. It does help resist the idea that confidence can be interpreted without reference to performance.
The Institute extends this into a broader synthesis: cultures of responsible judgment should value calibrated uncertainty, not confidence performance. That phrase is a normative and institutional framing, not a result attributable to Fleming alone.
AI creates a new confidence problem
When people work with AI, there may be at least three levels of uncertainty in play: the system's internal or reported uncertainty, the human user's confidence in the task, and the human user's confidence in their ability to evaluate the system.
These can diverge.
A highly fluent answer may increase user confidence without increasing answer quality. A cautious user may defer to a confident-looking system. A domain expert may know enough to recognise when the model's uncertainty signal is uninformative.
Metacognition research offers conceptual tools for separating these layers. It does not yet provide a complete science of human-AI joint calibration.
That is an open research frontier.
Confidence can guide control as well as report belief
Metacognitive research is sometimes described as though confidence were only an output: the decision happens, confidence is reported, and the experiment ends. Fleming's synthesis places confidence in a larger control architecture. A confidence estimate can affect whether a person checks again, changes a decision, seeks more evidence, asks for help or moves on.
That makes confidence consequential even when it is imperfect. Low confidence can trigger information seeking. High confidence can terminate search. If confidence is systematically miscalibrated, the error is therefore not only descriptive. It can change the sequence of actions that follows.
This is especially relevant to institutions that ask people to escalate uncertain decisions. An escalation policy depends on the person's ability and willingness to recognise uncertainty. If subjective confidence is badly related to actual accuracy, a formally sensible policy can fail at the monitoring stage.
The research does not imply that every decision should be accompanied by explicit numerical confidence. Repeated ratings can be reactive, and some domains do not support meaningful probability judgments. The point is structural: self-evaluation can become input to cognitive control.
Intervention claims need a mechanism
The idea that metacognition can be improved is attractive, but “improving metacognition” can mean several different things. An intervention might improve first-order performance, sharpen confidence discrimination, reduce confidence bias, teach better strategy selection, or alter how confidence is communicated. These outcomes should not be collapsed.
Fleming's 2024 review is useful here because it asks where a metacognitive intervention acts. If confidence depends on a model of the world and a model of the self, an intervention could change evidence quality, change the inference drawn from evidence, or change the mapping from an internal state to a report.
That distinction matters for transfer. A person can learn the confidence scale used in one laboratory task without developing a broadly transferable capacity for wiser uncertainty. Conversely, an educational intervention might improve monitoring behavior without moving a particular meta-d-prime estimate.
The Institute therefore treats “metacognitive training” as a family of intervention claims that must name their mechanism and outcome, not as one validated treatment category.
What the evidence does not show
Fleming's work does not show that confidence is a direct readout of truth.
It does not show that one brain area contains introspection.
It does not show that meta-d-prime is a context-free trait measure.
It does not show that good metacognition in one task guarantees good metacognition across domains.
It does not show that asking for confidence is measurement-neutral in every task.
It does not show that a neural correlate establishes a causal mechanism.
And it does not establish that confidence interventions will improve judgment outside the contexts in which they have been tested.
A research programme about the space between decision and self-knowledge
The distinctive contribution of Fleming's work is not one score or one scan.
It is a programme that keeps asking how a cognitive system moves from doing a task to evaluating its own doing, how that evaluation can be measured, why it sometimes diverges from performance, and what the divergence tells us about self-awareness.
That space is small in time. A decision happens; a confidence judgment follows.
But scientifically, the space is large.
It contains the difference between being right and knowing that one is right, between uncertainty and the report of uncertainty, and between a mind that acts and a mind that can inspect the reliability of its own action.
Research references
Fleming SM, Weil RS, Nagy Z, Dolan RJ, Rees G. Relating Introspective Accuracy to Individual Differences in Brain Structure. Science. 2010;329(5998):1541–1543. https://doi.org/10.1126/science.1191883
Fleming SM, Lau HC. How to measure metacognition. Frontiers in Human Neuroscience. 2014;8:443. https://doi.org/10.3389/fnhum.2014.00443
Fleming SM. HMeta-d: hierarchical Bayesian estimation of metacognitive efficiency from confidence ratings. Neuroscience of Consciousness. 2017;2017(1):nix007. https://doi.org/10.1093/nc/nix007
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
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
Double KS, Goldwater MB, Birney DP. Reactivity to confidence ratings: evidence of impaired rule-learning. Metacognition and Learning. 2025;20:8. https://doi.org/10.1007/s11409-025-09413-5
University College London. MetaLab. https://www.ucl.ac.uk/brain-sciences/pals/research/experimental-psychology/research/cognition-and-decision-making/metalab