Mechanisms
Metacognition · Confidence
Confidence Calibration
Example
You give a colleague two figures, sure of one and hedging on the other, and you're usually right about which is which, even though you can't say why one felt firmer.
Confidence Calibration in Humans
Definition
Judging, after producing an answer, how likely it is to be correct. The judgment is not privileged access to memory strength. It is inferred from the consistency of the representations sampled, essentially an estimate of whether sampling the same problem again would yield the same answer (Koriat, 2012), a view that current syntheses of the metacognition literature adopt as the general computational definition of confidence (Fleming, 2024). Two aspects of the judgment are conventionally separated: resolution, the within-person ability to assign higher confidence to correct than to incorrect answers, and calibration, the absolute match between stated confidence and observed accuracy. The two are independent, so calibration can be near perfect while resolution is near zero, and the reverse (Koriat, 2012).
*Framing note: one of four metacognitive monitoring rows, with Source Monitoring, Reality Monitoring, and Feeling of Knowing. Distinguished from Feeling of Knowing by direction: feeling-of-knowing is prospective, about an answer not yet retrieved; confidence is retrospective, about one already produced.*
Neural Substrate
Prefrontal cortex, on human evidence. A coordinate-based meta-analysis of 47 neuroimaging studies of metacognitive judgment, covering 88 contrasts, 739 activation foci and 2,215 participants, found convergent involvement of posterior medial frontal cortex, bilateral insula and inferior frontal gyrus, left dorsolateral and right anterior dorsolateral prefrontal cortex, ventromedial prefrontal cortex, right dorsal precuneus, and right ventral striatum (Vaccaro & Fleming, 2018). Contrasts in which activity scaled with the level of confidence reported were the dominant driver of that composite pattern.
A lateral-for-retrospective, medial-for-prospective division of labour was proposed on neuropsychological grounds, largely from frontal lesion cases in which feeling-of-knowing accuracy and retrospective confidence dissociate (Fleming & Dolan, 2012). The meta-analytic data do not reproduce that division cleanly. Prospective metamemory judgments converged on posterior medial frontal cortex, left dorsolateral prefrontal cortex and right insula, while retrospective metamemory judgments converged on bilateral parahippocampal cortex and left inferior frontal gyrus (Vaccaro & Fleming, 2018). The anatomical separation between prospective and retrospective judgments is real, but it is not the lateral-versus-medial one usually stated.
Metacognitive accuracy is dissociable from performance on the task being monitored. With first-order accuracy held at roughly 71 percent by an adaptive staircase, grey matter volume in right rostrolateral prefrontal cortex predicted individual differences in metacognitive accuracy; metacognitive accuracy correlated across two perceptual tasks at r = 0.71 while performance on those same two tasks did not correlate at all (r = 0.05, p = 0.83); and repetitive transcranial magnetic stimulation to bilateral dorsolateral prefrontal cortex lowered metacognitive accuracy while leaving task performance intact (Fleming & Dolan, 2012). This dissociation is better established than its anatomy: the meta-analysis of studies reporting neural correlates of metacognitive sensitivity specifically drew on only 11 studies and 61 foci and yielded no significant clusters after correction for multiple comparisons (Vaccaro & Fleming, 2018).
AI at a Glance
In people, confidence is a retrospective judgment about an answer already produced, inferred not from privileged access to memory strength but from how consistently the sampled evidence favoured that answer, and it discriminates correct from incorrect answers only because consistently favoured answers are usually the correct ones (Koriat, 2012; Fleming, 2024). In models the function is present and heavily measured, but split: a reasonably well calibrated estimate of correctness sits in the trained weights and can be extracted either by resampling the model and measuring disagreement (semantic entropy) or by probing internal activations, while the confidence the system states in words is a partly separate quantity that tracks the internal one imperfectly (Kuhn et al., 2023; Farquhar et al., 2024; Kumar et al., 2024; Kumaran et al., 2026).
Confidence Calibration in AI
How Far It Has Been Measured
Measurement in AI: Directly Measured
AI Memory Performance: Comparable to Humans provisional
Where It Lives in the System
What Has Been Measured
- A peer-reviewed 2026 study gave 87 Japanese-speaking adults (recruited on X, answering an online form in September–October 2023) and 87 separate runs of GPT-4 the same 60 general-knowledge questions about Japan as open-ended recall, four-choice and two-choice items, each followed by a 0–100% confidence rating; GPT-4 was queried in December 2025 through OpenAI's programming interface under the setting "gpt-4", which points to a June 2023 version, one question per fresh request, so it answered only from what it learned in training (Yoshizawa et al., 2026).
- GPT-4 answered more questions correctly than people (78–88% vs 52–73% across formats) but reported 93–97% confidence throughout, overconfident by 8–15 percentage points, while people's average confidence was within 4 points of their accuracy; individual people varied widely (SD of over- or underconfidence 15–16 points, computed from the authors' public data), so GPT-4's bias sits within the human range, about 0.6–0.75 SD above the human mean (Yoshizawa et al., 2026).
- For resolution, scored as type-2 AUROC (the chance that a correct answer drew higher confidence than a wrong one, .5 = chance), GPT-4 matched people on two-choice (.73 vs .70, SD .14) and open-ended recall (.78 vs .81, SD .12) but fell below chance on four-choice items (.40 vs .76, SD .13), where it was more confident in its wrong answers (99%) than its right ones (96%); across questions, people's confidence tracked item difficulty far more closely than GPT-4's (r = .83 vs .33) (Yoshizawa et al., 2026).
- The study's efficiency measures are information-based (meta-I and meta-I2r), not meta-d′ or M-ratio, and they do not line up with its AUROCs: on open-ended recall AUROC shows no reliable difference while both information measures favour people, and a mutual-information score counts GPT-4's inverted four-choice confidence as positive information; the text also reverses the direction of a format effect shown in the authors' own statistical output, and its "All" AUROC averages the three format scores rather than pooling trials (Yoshizawa et al., 2026).
- A peer-reviewed five-study paper (online 2025, in print 2026) gave online US adults and public chatbots (ChatGPT, Bard/Gemini and, in Studies 4–5, Claude Sonnet and Haiku, 2023–2025 versions, each run repeatedly through its chat interface) identical items and confidence scales, but its items were NFL and Oscar forecasts, sketch guessing, 22 more-than/fewer-than trivia questions built to require piecing facts together (110 adults), and 33 questions about an unnamed university's unpublished records (110 adults), so no study isolates confidence in answers recalled from memory (Cash et al., 2026).
- On the trivia questions the best system, Claude 3.5 Sonnet (released 2024), missed its own prospective score by 2.15 items versus 3.51 for people (SD 3.15), with AUROC .66 versus .55 (SD .10), while systems ranged from 2.15 to 4.94 items and .57 to .66; on university questions Claude 3.7 Sonnet (2025) missed by 1.08 items versus 4.84 (SD 3.74), yet people discriminated right from wrong answers better (.57 vs .53), and over all 24 calibration comparisons LLMs were better 13 times, equal 4 and worse 7 (Cash et al., 2026).
- Both are direct comparisons and both put AI at or above people, so the rating is Comparable: in the forecasting, trivia and records studies the best systems matched or beat people's calibration, and on general-knowledge recall GPT-4 was more accurate than people and within the human range on calibration and on two of three resolution formats. The important caveat is overconfidence: GPT-4 was 8–15 points more confident than it was accurate where people were within 4, and on four-choice items it was more confident in its wrong answers than its right ones. The rating is provisional because the recall study tests one 2023 system whose 87 runs were near-duplicates, and the other study's items are forecasts and estimates rather than answers from memory (Yoshizawa et al., 2026).
- Neither study examined saved memory; the closest test, a benchmark of chat assistants' long-term memory of past sessions, scores whether they decline questions their memory cannot answer (abstention), not whether confidence tracks accuracy, and has no human comparison (Wu et al., 2025).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Dorsolateral prefrontal cortex
Cortex in the prefrontal cortex
The upper outer surface of the prefrontal cortex, on and around the middle frontal gyrus, roughly Brodmann areas 9 and 46.
The dorsolateral prefrontal cortex sits on the lateral convexity of the frontal lobe, centered on the middle frontal gyrus and covering approximately Brodmann areas 9 and 46; its borders are drawn differently by different labs and no sulcus marks them. It is densely connected with posterior parietal cortex, the dorsal striatum, and the mediodorsal thalamus. Activity here tracks working-memory delays, task switching, and rule maintenance, and it is the usual target for prefrontal transcranial stimulation. Whether it stores the contents of working memory, holds the rules for acting on them, or simply reflects task difficulty is genuinely contested, and "executive function" names the dispute rather than settling it.
Frontal lobe
Lobe in the cerebral hemisphere
The lobe in front of the central sulcus and above the lateral fissure, the largest of the four in humans.
The frontal lobe runs from the frontal pole back to the central sulcus, with the lateral fissure marking its lower boundary against the temporal lobe. Its posterior strip is the precentral gyrus, and in front of that lie premotor and supplementary motor areas; everything anterior to those is prefrontal cortex. Its lateral surface folds into superior, middle and inferior frontal gyri, its ventral surface into orbital gyri and the gyrus rectus. Frontal damage is associated with changes in planning, inhibition and social conduct, but "frontal lobe function" covers many dissociable deficits rather than one.
Frontopolar cortex
Cortex in the prefrontal cortex
The most anterior cortex in the brain, the rounded tip of the frontal lobe, corresponding to Brodmann area 10.
The frontopolar cortex is the forward tip of the frontal lobe and the foremost cortex in the human brain, approximately coextensive with Brodmann area 10, which reaches beyond the pole ventromedially and is proportionally larger in humans than in other primates, with an unusually low density of cells per unit volume, leaving more space for connections. Its inputs come almost entirely from other association cortex rather than from sensory areas. It is activated when people hold one goal in mind while pursuing another, evaluate their own performance, and reason about possibilities. Proposals range from relational integration to branching control to prospective memory, and no single account has displaced the others.
Grey matter
Division
Tissue dominated by neuronal cell bodies, dendrites and synapses, darker in appearance than the myelinated axon bundles around it.
Grey matter is where neuronal cell bodies, dendrites and most synapses sit, together with glia and a dense capillary bed. It looks grey-brown because it lacks the lipid-rich myelin that makes axon bundles appear white. In the cerebrum it forms the outer cortical sheet plus the deep nuclei; in the spinal cord the arrangement inverts, with grey matter central. Structural MRI studies routinely report "grey matter volume" or "cortical thickness" differences between groups, but those are model-derived measures sensitive to scanner, segmentation algorithm and even hydration, and small reported differences have often failed to replicate.
Inferior frontal gyrus
Cortex in the frontal lobe
The lowest lateral frontal gyrus, divided into pars opercularis, triangularis and orbitalis by the ascending rami of the lateral fissure.
The inferior frontal gyrus runs above the lateral fissure and is cut into three parts by that fissure's ascending and horizontal rami: pars opercularis behind, pars triangularis in the middle, pars orbitalis in front. In the left hemisphere the first two correspond roughly to Broca's area, and damage there is associated with non-fluent aphasia. The right IFG is repeatedly reported in response-inhibition tasks such as stop-signal, though whether its contribution is specifically inhibitory or reflects detecting behaviorally relevant signals is a live argument. Its cytoarchitectonic boundaries do not align neatly with the gross parts.
Insula
Cortex in the cerebral hemisphere
A triangular expanse of cortex buried inside the Sylvian fissure, covered over by the frontal, parietal and temporal opercula.
The insula is the only cortical lobe with no exposed surface: it sits at the bottom of the Sylvian fissure, hidden by the opercula that grow over it during development, and is divided by its central sulcus into a larger anterior and smaller posterior part. Its architecture grades from agranular at the front to granular at the back. It receives visceral, thermal, nociceptive and gustatory input. Popular accounts name it the seat of interoception, disgust, empathy, or consciousness in turn; the region appears in a large share of published fMRI contrasts, which makes reverse inference from insular activity to any one of those especially unsafe.
Parahippocampal cortex
Cortex in the parahippocampal gyrus
The posterior part of the parahippocampal gyrus, responding strongly to scenes and spatial layouts and feeding medial entorhinal cortex.
Cortex behind the entorhinal region on the medial temporal surface, connected to retrosplenial and posterior parietal areas and projecting into medial entorhinal cortex. Human imaging finds a patch here that responds far more to images of places than of objects or faces: the parahippocampal place area. What that response encodes is unsettled: candidate accounts include scene geometry, spatial context for navigation, and the presence of large landmark-like objects, and each explains some of the contrast that named the region.
Precuneus
Cortex in the parietal lobe
The medial parietal cortex in front of the parieto-occipital sulcus, above the posterior cingulate and behind the paracentral lobule.
The precuneus occupies the medial surface of the parietal lobe, bounded by the parieto-occipital sulcus behind, the marginal branch of the cingulate sulcus in front, and the subparietal sulcus below. It is densely connected to posterior cingulate and retrosplenial cortex, to the superior parietal lobule, and to the thalamus. Metabolically it is one of the most active regions of the resting brain and among the earliest to show hypometabolism in Alzheimer's disease. It is reported in episodic retrieval, mental imagery, perspective-taking and self-referential tasks; no single account of what unites those is established.
Prefrontal cortex
Cortex in the frontal lobe
The frontal cortex in front of the motor areas, defined classically by its connections with the mediodorsal thalamus rather than by any one function.
The prefrontal cortex occupies the front of the frontal lobe, anterior to premotor and motor cortex, and accounts for roughly a sixth to a quarter of human cortex depending on where the posterior border is drawn. Three competing definitions are in use (projection from the mediodorsal thalamus, the presence of a granular layer IV, and the absence of movement on electrical stimulation), and they do not enclose the same tissue, which is why rodent-to-primate homology is still argued over. It is reliably engaged by tasks with delays, rules, and conflicting options; what it contributes to those tasks is the open question, not whether it participates.
Ventral striatum
Subcortical in the striatum
The lower striatal territory (nucleus accumbens and neighboring olfactory tubercle) receiving limbic and prefrontal input.
Heimer's term for striatal tissue defined by its inputs rather than its position above or below the anterior commissure: it takes projections from amygdala, hippocampus, and orbital and medial prefrontal cortex, and sends output to ventral pallidum rather than to the internal pallidal segment. It is the limbic end of the cortico-striatal loops, and the usual localization in human imaging for signals scaling with anticipated outcomes. Human studies frequently report "ventral striatum" where the resolution cannot distinguish accumbens from adjacent caudate or putamen.
Ventromedial prefrontal cortex
Cortex in the medial prefrontal cortex
The lower medial corner of the prefrontal cortex, where the medial wall meets the orbital surface above the eye sockets.
The ventromedial prefrontal cortex occupies the junction of the medial and orbital surfaces of the frontal lobe, covering parts of Brodmann areas 10, 11, 14, 25 and 32; its overlap with medial orbitofrontal cortex is substantial and inconsistently labeled. It is heavily connected with the amygdala, ventral striatum, and hypothalamus. Damage here (the Phineas Gage and Elliot cases, and later lesion series) produces poor real-world decisions with intact intelligence, and the region is central to the somatic marker hypothesis. Its popular description as a moral-decision region is a strong reading of weak evidence: the same tissue tracks subjective value, regulation of fear responses, and memory schemas.
Cash et al. (2025)
Quantifying uncert-AI-nty: Testing the accuracy of LLMs’ confidence judgments
In
Memory & Cognition, 54(2), 375-400
Abstract
Abstract The rise of Large Language Model (LLM) chatbots, such as ChatGPT and Gemini, has revolutionized how we access information. These LLMs can answer a wide array of questions on nearly any topic. When humans answer questions, especially difficult or uncertain questions, they often accompany their responses with metacognitive confidence judgments indicating their belief in their accuracy. LLMs are certainly capable of providing confidence judgments, but it is currently unclear how accurate these confidence judgments are. To fill this gap in the literature, the present studies investigate the capability of LLMs to quantify uncertainty through confidence judgments. We compare the...
Find it
Farquhar et al. (2024)
Detecting hallucinations in large language models using semantic entropy
In
Nature, 630(8017), 625-630
Abstract
Large language model (LLM) systems, such as ChatGPT1 or Gemini2, can show impressive reasoning and question-answering capabilities but often ‘hallucinate’ false outputs and unsubstantiated answers3,4. Answering unreliably or without the necessary information prevents adoption in diverse fields, with problems including fabrication of legal precedents5 or untrue facts in news articles6 and even posing a risk to human life in medical domains such as radiology7. Encouraging truthfulness through supervision or reinforcement has been only partially successful8. Researchers need a general method for detecting hallucinations in LLMs that works even with new and unseen questions to which humans...
Find it
Fleming (2024)
Metacognition and Confidence: A Review and Synthesis
In
Annual Review of Psychology, 75(1), 241-268
Abstract
Determining the psychological, computational, and neural bases of confidence and uncertainty holds promise for understanding foundational aspects of human metacognition. While a neuroscience of confidence has focused on the mechanisms underpinning subpersonal phenomena such as representations of uncertainty in the visual or motor system, metacognition research has been concerned with personal-level beliefs and knowledge about self-performance. I provide a road map for bridging this divide by focusing on a particular class of confidence computation: propositional confidence in one's own (hypothetical) decisions or actions. Propositional confidence is informed by the observer's models of...
Find it
Koriat (2012)
The self-consistency model of subjective confidence.
In
Psychological Review, 119(1), 80-113
Abstract
How do people monitor the correctness of their answers? A self-consistency model is proposed for the process underlying confidence judgments and their accuracy. In answering a 2-alternative question, participants are assumed to retrieve a sample of representations of the question and base their confidence on the consistency with which the chosen answer is supported across representations. Confidence is modeled by analogy to the calculation of statistical level of confidence (SLC) in testing hypotheses about a population and represents the participant's assessment of the likelihood that a new sample will yield the same choice. Assuming that participants draw representations from a commonly...
Find it
Kuhn et al. (2023)
Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
In
arXiv
Abstract
We introduce a method to measure uncertainty in large language models. For tasks like question answering, it is essential to know when we can trust the natural language outputs of foundation models. We show that measuring uncertainty in natural language is challenging because of "semantic equivalence" -- different sentences can mean the same thing. To overcome these challenges we introduce semantic entropy -- an entropy which incorporates linguistic invariances created by shared meanings. Our method is unsupervised, uses only a single model, and requires no modifications to off-the-shelf language models. In comprehensive ablation studies we show that the semantic entropy is more...
Find it
Kumar et al. (2024)
Confidence Under the Hood: An Investigation into the Confidence-Probability Alignment in Large Language Models
In
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 315-334
Abstract
As the use of Large Language Models (LLMs) becomes more widespread, understanding their self-evaluation of confidence in generated responses becomes increasingly important as it is integral to the reliability of the output of these models. We introduce the concept of Confidence-Probability Alignment, that connects an LLM's internal confidence, quantified by token probabilities, to the confidence conveyed in the model's response when explicitly asked about its certainty. Using various datasets and prompting techniques that encourage model introspection, we probe the alignment between models' internal and expressed confidence. These techniques encompass using structured evaluation scales to...
Find it
Kumaran et al. (2026)
How do LLMs Compute Verbal Confidence
In
arXiv
Abstract
Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from black-box models. However, how LLMs internally generate such scores remains unknown. We address two questions: first, when confidence is computed -- just-in-time when requested, or automatically during answer generation and cached for later retrieval; and second, what verbal confidence represents -- token log-probabilities, or a richer evaluation of answer quality? Focusing on Gemma 3 27B (across TriviaQA, BigMath, and MMLU), Qwen 2.5 7B, and the reasoning model Magistral Small 24B, we provide convergent evidence for cached retrieval. Activation...
Find it
Vaccaro & Fleming (2018)
Thinking about thinking: A coordinate-based meta-analysis of neuroimaging studies of metacognitive judgements
In
Brain and Neuroscience Advances, 2
Abstract
Metacognition supports reflection upon and control of other cognitive processes. Despite metacognition occupying a central role in human psychology, its neural substrates remain underdetermined, partly due to study-specific differences in task domain and type of metacognitive judgement under study. It is also unclear how metacognition relates to other apparently similar abilities that depend on recursive thought such as theory of mind or mentalising. Now that neuroimaging studies of metacognition are more prevalent, we have an opportunity to characterise consistencies in neural substrates identified across different analysis types and domains. Here we used quantitative activation...
Find it
Wu et al. (2024)
LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
In
arXiv
Abstract
Recent large language model (LLM)-driven chat assistant systems have integrated memory components to track user-assistant chat histories, enabling more accurate and personalized responses. However, their long-term memory capabilities in sustained interactions remain underexplored. We introduce LongMemEval, a comprehensive benchmark designed to evaluate five core long-term memory abilities of chat assistants: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. With 500 meticulously curated questions embedded within freely scalable user-assistant chat histories, LongMemEval presents a significant challenge to existing long-term memory...
Find it
Yoshizawa et al. (2026)
Metacognition of ChatGPT in confidence judgements
In
Frontiers in Artificial Intelligence, 9
Abstract
Recent advances in Large Language Models (LLMs) have raised critical concerns regarding AI alignment and safety, particularly with respect to the reliability of their outputs. In humans, metacognition plays a key role in making cognition robust and adaptive. LLMs frequently express high confidence in their responses, raising the question of whether such confidence reflects human-like metacognitive capability. In this study, we systematically compared humans and GPT-4 across multiple task formats to examine how confidence relates to performance. GPT-4 consistently outperformed humans in task accuracy. This advantage was not accompanied by human-like confidence behavior: Human confidence...
Find it