Metacognition and Confidence: A Review and Synthesis

The paper addresses how to reconcile two perspectives: (1) neuroscience of confidence, which focuses on subpersonal representations of uncertainty within sensory and motor systems in controlled tasks, and (2) metacognition research, which focuses on personal-level beliefs and knowledge about one’s own performance in real-world settings. The core proposal is to center analysis on propositional confidence—confidence in one’s own (hypothetical or actual) decisions or actions—and to clarify how it differs from other uncertainty signals. The study emphasizes the importance of accurately estimating uncertainty and confidence for adaptive behavior and highlights metacognitive sensitivity (the alignment of confidence with performance) as crucial. By introducing a distinction between world-centered and self-centered reference frames for uncertainty, the paper aims to integrate computational, psychological, and neural findings into a framework that explains when and why metacognitive judgments diverge from task performance and how interventions may affect specific computational components.

The review synthesizes decades of work in metamemory and metaperception, covering behavioral paradigms (prospective judgments of learning, retrospective confidence in recall, confidence forced-choice) and implicit measures (opt-out, wagering, waiting-for-reward) used in humans, animals, and infants. It summarizes measurement approaches to metacognitive sensitivity, including signal detection theoretic meta-d′ (and critiques regarding dependence on bias/performance), model-free mutual information and psychometric slope changes. Empirical literatures are cataloged that demonstrate dissociations between confidence and accuracy via attention, variability in evidence, asymmetrical processing of confirmatory/disconfirming evidence, and response times. Neuropsychological and neuroimaging evidence implicates anterior prefrontal cortex (including BA46/10, frontopolar cortex), medial prefrontal regions, precuneus, and ventral striatum in metacognitive processing, with lesion/TMS studies showing domain-specific impairments (e.g., frontopolar cortex for metaperception; precuneus for metamemory). The review details neural correlates of uncertainty and confidence (e.g., LIP evidence accumulation and opt-out behavior; OFC confidence signatures in rodents; pgACC/vmPFC and pMFC signals in humans), and integrates error monitoring literature (ERN, Pe, CPP) linking postdecisional accumulation to confidence formation. It surveys global broadcast and communication of confidence, including domain-general versus domain-specific signals, social coordination and private–public confidence mappings, and the role of self-models (fluency, interoception, priors) that introduce inferential cues and potential illusions. Finally, it revisits controversies around biases (positive evidence bias), suboptimalities, and domain-generality, and outlines future directions for computational unification, extension from local to global metacognition, and targeted interventions.

This article is a conceptual review and synthesis rather than an empirical study. The methodology consists of: (1) Definitional clarification of metacognition and confidence, introducing a distinction between world-centered uncertainty and self-centered propositional confidence. (2) Development of a computational framework for metacognitive judgments comprising components: representing uncertainty (implicit/distributional), transforming internal states into propositional confidence (Bayesian or self-consistency formulations), global broadcast/communication, and model-based self-influences (priors, beliefs, interoceptive cues). (3) Integration of cross-domain empirical findings from psychophysics (evidence accumulation, postdecisional processes), cognitive psychology (metamemory cues, judgments of learning), social psychology (confidence communication, group decision-making), and neuroscience (single-unit recordings in animals, EEG/fMRI in humans, lesion/TMS studies). (4) Use of canonical tasks/paradigms to illustrate the computational components: random dot motion discrimination with coherence and boundary manipulations; opt-out and wagering tasks; odor/auditory discrimination with confidence-based waiting; prospective/retrospective confidence ratings; global confidence tracking across blocks. (5) Mapping behavioral signatures (e.g., folded X pattern for confidence vs. signal strength; ERN/Pe/CPP dynamics) to brain systems (visual cortex uncertainty decoding; LIP accumulation; pgACC/vmPFC propositional confidence; pMFC error/confidence monitoring; frontopolar cortex broadcast/private–public mapping; precuneus in metamemory). (6) Analytical reevaluation of measurement tools (meta-d′, mutual information, model-free metrics), sources of bias/suboptimality (positive evidence bias, variance misperception, decision-congruent heuristics), and domain-generality through individual-difference correlations and domain-specific neural substrates.

• Confidence relevant to metacognition is best conceived as propositional confidence in one’s own decisions/actions, distinct from distributional uncertainty in world-centered representations. • Neural systems: pgACC/vmPFC encode propositional confidence; pMFC supports error monitoring and postdecisional evidence accumulation; orbitofrontal cortex (rodents) carries modality-general confidence signatures; precuneus is selectively implicated in metamemory; lateral frontopolar cortex contributes to metacognitive efficiency and private–public confidence mapping. • Behavioral and neural signatures: the folded X pattern validates propositional confidence (confidence increases with signal strength on correct trials and decreases on errors); ERN and Pe/CPP reveal postdecisional accumulation and subjective error awareness; LIP firing variability predicts opt-out choices even with fixed stimuli. • Uncertainty informs confidence: trial-by-trial uncertainty decoded from visual cortex correlates with reported confidence; subjects adjust confidence criteria with stimulus uncertainty; multisensory integration weights sources by inverse uncertainty consistent with Bayesian principles. • Model-based influences: fluency, font size/brightness, interoceptive arousal, and priors modulate confidence independently of accuracy; priors can be causally shifted (false feedback) altering confidence calibration without affecting performance or RT. • Domain-generality vs specificity: moderate correlations of metacognitive efficiency across tasks suggest a global resource, but domain-specific neural substrates and lesion/TMS dissociations indicate distinct pipelines for perception vs memory. • Biases/suboptimalities: Positive evidence bias emerges naturally in high-dimensional classification systems and can be reversed by reframing the decision; metacognitive inefficiency arises from meta-uncertainty, constraints on postdecisional accumulation, and maintaining distinct self/other models. • Interventions and individual differences: meditation, pharmacology (noradrenaline blockade), neurofeedback, brain stimulation, and training can alter metacognitive metrics; metacognitive bias is more reliable over time and relates to personality/mental health, while metacognitive efficiency associates with openness and reduced dogmatism.

Framing metacognitive judgments as estimates of propositional confidence provides a principled bridge between subpersonal uncertainty representations and personal-level self-evaluation. This framework explains dissociations between performance and confidence as arising from transformations between world- and self-centered reference frames, postdecisional evidence accumulation, and model-based self-influences. It reconciles findings across species and modalities by assigning sensory cortex and parietal systems to uncertainty representation, and prefrontal circuits (pgACC/vmPFC, pMFC, frontopolar) to confidence formation, monitoring, broadcast, and communication. The approach clarifies measurement concerns (e.g., limitations of meta-d′) and illuminates sources of bias/suboptimality (positive evidence bias, variance misperception), highlighting where deviations from ideal observer models naturally arise. The significance lies in unifying perceptual and memory metacognition, articulating how global confidence priors aggregate local judgments, and identifying computational loci for interventions that may generalize beyond laboratory tasks.

The article synthesizes metacognition and confidence research under the umbrella of propositional confidence, detailing computational components—uncertainty representation, confidence formation in a self-centered frame, global broadcast/communication, and model-based self-influences. It integrates disparate neural findings into a cohesive map of metacognitive circuitry and clarifies how biases and domain-specificities arise. Future research should identify common computational principles across domains, extend from local trial-level judgments to global self-evaluations over time, model rich self-influences and their relation to theory of mind, generalize psychophysical confidence models to naturalistic contexts, and design interventions targeted at specific computational stages to achieve functional benefits.

The review is selective due to scope constraints and does not comprehensively cover development of metacognition, comparative animal metacognition, links with mental health and ageing in detail, or interpersonal/intrapersonal functions beyond targeted discussions. Empirical ambiguities remain in tasks like opt-out that can be solved via world-centered uncertainty without explicit metacognitive confidence. Measurement approaches (e.g., meta-d′) may be confounded by performance and bias, complicating cross-study comparisons. Domain-generality findings can be influenced by confidence scale use and accumulation thresholds. Many neural findings are correlational and task-specific; generalization to naturalistic scenarios and long-timescale global self-beliefs requires further work.