Feeling of Knowing

Definition

The capacity to judge whether unrecalled information is stored and will be retrievable later. The judgment is not a direct readout of the store. It is computed from how much pertinent information the retrieval attempt makes accessible, and it is calibrated because accessible information tends to be correct (⁠Koriat, 1993⁠). The tip-of-the-tongue experience overlaps this capacity without being contained by it: there the accessible information is the target's form, its opening letters, its syllable count and its stress pattern, and the expectation is of imminent recall rather than of later recognition (R. Brown & McNeill, 1966; A. S. Brown, 1991).

Contested Boundaries: The account above is one side of a live dispute. A competing view holds that the feeling of knowing rests on familiarity with the question rather than on anything retrieved about the answer, and Schwartz and Metcalfe (1992) tested the two against each other across four experiments and found for cue familiarity. In their first two experiments a generation manipulation raised recall, and raised recognition of the unrecalled items when the two studies were pooled, while leaving feeling-of-knowing judgments untouched; priming the cue in an unrelated pleasantness-rating task raised the judgments in all four experiments. Their fourth experiment, which primed before encoding and used unrelated paired associates so as to remove two artifacts present in their third, found target priming raising recognition without raising the judgments and cue priming doing the reverse. They also argue against the accessibility account specifically: if cue priming worked by increasing the amount of information retrieved, commission errors should have risen alongside the judgments, and in their first two experiments they did not.

Koriat (1993) takes up cue familiarity and argues it may itself reduce to accessibility, since an unfamiliar question is one that leaves the person blank while a familiar one brings associations to mind, though he concedes the alternative that familiarity contributes independently and early, before accessibility takes over. He also notes that the cue-familiarity account is awkward to apply to his own experiments, which used no distinctive cue. A later cascaded proposal treats the two as sequential stages, an initial familiarity check that either terminates search or licenses it, followed by an accessibility assessment of whatever the search produces (⁠Schnyer et al., 2004⁠).

One scope limit follows from this and bears on the tip-of-the-tongue clause in the Definition. Tip-of-the-tongue is the phenomenological anchor for the target-retrievability account, which Schwartz and Metcalfe name as its source and which this row does not follow. Koriat's own evidence is not tip-of-the-tongue evidence: his materials were nonsense consonant strings, chosen to minimise preexperimental familiarity and content-based inference, which are precisely the effects the tip-of-the-tongue experience depends on. Where the row follows the accessibility account, it is following the source, not a settled field.

Neural Substrate

The evidence is human throughout. The most direct result comes from lesion work: 14 patients with focal frontal damage and 18 matched controls read sentences, attempted cued recall of the final word, rated their confidence in each recall attempt, and then made feeling-of-knowing judgments for the items they had missed before a six-alternative recognition test (⁠Schnyer et al., 2004⁠). The patients were impaired on the feeling-of-knowing judgments by both indices used, F(1, 29) = 6.95, p \< .01 for gamma and F(1, 30) = 8.59, p \< .01 for the Hamann coefficient, while their confidence ratings about the recall attempts themselves were indistinguishable from controls. Their memory was also impaired, but entering recognition performance as a covariate left the group difference in feeling-of-knowing marginal rather than eliminating it, so the deficit is not simply a memory deficit. Lesion overlap among the four most impaired patients converged on right medial prefrontal cortex. The same patients rated unstudied filler sentences no differently from controls, which the authors read as intact familiarity assessment alongside a broken accessibility assessment, mapping the deficit onto the second stage of the cascaded account rather than the first.

The dissociation runs in both directions and is not confined to this construct. Fleming and Dolan (2012) review evidence that metacognitive accuracy is separable from task performance and varies across individuals, and that prefrontal damage can selectively degrade the accuracy of metacognitive reports while leaving performance relatively intact. The converse pattern is also on record: patients whose amnesia arises from medial temporal damage show impaired memory with preserved feeling-of-knowing, while patients with frontal atrophy show impaired feeling-of-knowing at matched recognition performance, which is what first pointed the literature at the frontal lobes rather than the memory system itself. Fleming and Dolan propose a specific division within prefrontal cortex, lateral regions for retrospective confidence and medial regions for prospective judgments such as feeling-of-knowing, but they present it as a synthesis of a small and partly indirect literature rather than a settled localisation.

The imaging picture is looser than that proposal implies. An activation likelihood estimation meta-analysis of 47 neuroimaging studies, 88 contrasts, 739 activation foci, and 2,215 participants found a domain-general network for metacognitive judgments spanning posterior medial frontal cortex, bilateral insula and inferior frontal gyrus, bilateral dorsolateral prefrontal cortex, ventromedial prefrontal cortex, right dorsal precuneus, and right ventral striatum (⁠Vaccaro & Fleming, 2018⁠). Restricting to metamemory left left dorsolateral prefrontal cortex, posterior medial frontal cortex, bilateral insula and inferior frontal gyrus, and bilateral parahippocampal gyrus. Splitting metamemory by temporal focus, the 14 studies of prospective judgments, which is where feeling-of-knowing sits, converged on posterior medial frontal cortex, left dorsolateral prefrontal cortex, and right insula, while the 17 retrospective studies converged on bilateral parahippocampal cortex and left inferior frontal gyrus. That is not the clean medial-prospective and lateral-retrospective split, and the authors say so, noting that direct imaging evidence for a distinction between judgment types is limited. Two limits bound all of this. The meta-analysis collapses across feeling-of-knowing, judgments of learning, and confidence ratings, which may not share a mechanism, and its analysis of metacognitive sensitivity, the property that actually corresponds to accuracy rather than to judgment magnitude, yielded no surviving clusters from only 11 studies. The imaging literature localises how confident a person feels considerably better than it localises whether that feeling is any good.

In people the feeling of knowing predicts whether a currently unretrievable item will be recognised later, and it does so not by inspecting the store but by reading off how much information the failed retrieval attempt brought to mind, which works because what comes to mind is usually correct (⁠Koriat, 1993⁠). Models can be prompted or trained to produce a comparable prospective estimate, which discriminates above chance in most systems tested but is calibrated in few and is rarely acted on (⁠Cao et al., 2026⁠; ⁠Kadavath et al., 2022⁠), and the estimate is about the weights only: no component in an adaptive-retrieval system judges whether the external document store holds a target before searching it (⁠Moskvoretskii et al., 2025⁠).

• Weights: the numerical parameters fixed at the end of training, where whatever a model retains from its training data is stored (Mallen et al., 2023).
• Adaptive retrieval: a system that decides case by case whether to consult an external document collection before answering, rather than always or never doing so (Mallen et al., 2023).

How Far It Has Been Measured

Where It Lives in the System

01 Weights 02 Context 03 Memory layer

What Has Been Measured

• Not yet compared: no study found through September 2026 gives an AI the recall–judgment–recognition procedure (fail to produce an answer, rate whether it would be recognized, then take a recognition test), with or without a human baseline, so there are no AI results on the ability as psychologists measure it (⁠Hart, 1965⁠).
• A 2026 preprint that uses the FOK name told seven language models not to solve each question but to rate how likely they were to answer it correctly; on 720 factual test questions, discrimination (area under the ROC curve, where 0.5 is chance) ranged from 0.56 (Qwen3-7B) to 0.67 (Qwen3-32B), with Claude Sonnet 4.6 at 0.66 and GPT-5.4 at 0.57, and no human baseline (⁠Kale, 2026⁠).
• That task is a different measure: it predicts the model's own recall from the question alone, with no failed-recall step and no recognition test, so it resembles the rapid initial FOK or a prospective confidence rating, and its score mainly reflects telling answerable from unanswerable questions, which is exactly what FOK for unrecalled items leaves out; people who rate recall instead of predicting recognition score very differently, so the two cannot stand in for each other (⁠Meunier-Duperray et al., 2025⁠).
• The scale conversion itself is published (gamma = 2 × area − 1), and one human FOK study reports the ROC-based value: 0.14 to 0.26 against the best model's 0.34, inside the human spread; this is not used, because the tests measure different judgments and the direction of any advantage is unclear (the models were told not to attempt retrieval, which supplies the partial information human FOK relies on, but were also not restricted to unrecalled items) (⁠Yeung, 2022⁠).
• The earliest close test trained a 52-billion-parameter model to estimate, from the question alone, the probability that it knows the answer; this separated known from unknown trivia questions well (0.864), transferred unevenly to other tasks (0.606 to 0.928), and rose when relevant text was placed in its context window, meaning what it can see in the current conversation (⁠Kadavath et al., 2022⁠).
• The gap FOK depends on does exist in models: in three 7–9-billion-parameter models, an internal readout ranked the correct answer above every wrong one for 9% of questions whose answer the model never produced in 1,000 attempts, and internal knowledge exceeded expressed knowledge by 14% to 57% across models; the models were not asked to sense this themselves (⁠Gekhman et al., 2025⁠).
• For saved or retrieved memory, systems use a model's sense of what it already knows to decide when to look information up in an external store, but they are scored on final question-answering accuracy, not on how well that sense predicts later recognition (⁠Wang et al., 2023⁠).
• Same-question human–AI studies exist only for confidence given after answering, which bears on retrospective confidence rather than FOK: on trivia, 110 people's ratings discriminated right from wrong answers at an area of 0.50 and four chatbots (each run 100 times) at 0.48 to 0.65 (⁠Cash et al., 2025⁠).
• Likewise, on Japanese general-knowledge questions rated after each answer, 87 people discriminated right from wrong answers at an area of 0.76 (SD 0.14) and GPT-4, run 87 times, at 0.63 (SD 0.20); again a retrospective confidence result, not FOK (⁠Yoshizawa et al., 2026⁠).