The mind that knows and invents

Begin with an ordinary visual judgement. A cloud of dots moves mostly left or mostly right. On an easy trial, nearly every dot agrees. On a hard trial, the pattern is noisy. You choose a direction in both cases. Then you estimate the probability that your answer is correct.

The confidence report need not reveal the detailed computation that produced the choice. It only needs access to information that predicts success: the strength of sensory evidence, conflict between alternatives, response time, remembered difficulty, or recent error history. When high confidence tends to accompany correct answers and low confidence tends to accompany errors, the observer has useful metacognitive sensitivity. Fleming and Lau describe several ways to measure this relation, including signal-detection methods that separate first-order task performance from second-order confidence discrimination.[1]

Now change the question. Instead of asking, “How likely is your answer to be right?”, ask, “Which internal process made you choose left?” The requested evidence is no longer an estimate of success. It is a causal history. The answer would need to identify which representations, memories, goals, habits or social cues made a difference to the action. Much of that history may never have been represented in a verbal form.

Metacognitive monitoring and causal self-explanation are therefore different evidential functions. One maps a judgement to expected correctness. The other maps an action to its producing conditions. They may use overlapping information, but overlap is not identity.

Calibration is not eloquence

Suppose Arin completes four equally weighted trials. His confidences are 0.90, 0.70, 0.60 and 0.40. The corresponding outcomes are correct, correct, wrong and wrong. A simple Brier score averages the squared distance between each confidence and its binary outcome:

A score of zero would be perfect. Arin is not perfect, but his confidence usefully separates likely successes from likely errors. Now imagine Bela answers the same two trials correctly and the same two incorrectly, while saying 0.65 every time. Bela’s Brier score is 0.2725. Accuracy is identical, yet Arin’s uncertainty is more decision-useful.

This example exposes three properties that are often collapsed. Accuracy measures first-order success. Confidence bias asks whether ratings run systematically high or low. Resolution asks whether confidence separates correct from incorrect trials. More advanced measures such as meta-d-prime estimate how much first-order evidence is available to second-order reports, while controlling some performance confounds.[1] No one statistic explains the mechanism, but each prevents confidence from being judged by tone.

What confidence may actually monitor

It is tempting to say that confidence reads truth directly. It cannot. Before the outcome arrives, the decider has only cues that usually, but not always, predict correctness. Those cues can include evidence strength, internal agreement, speed, familiarity and the stability of a response when the question is reconsidered. Koriat develops a serious alternative to the truth-meter picture: subjective confidence may often monitor how replicable a response seems from the sampled cues, rather than accuracy itself.[2] In ordinary environments, repeatability and truth can correlate. In systematically misleading environments, a highly replicable error can feel certain.

Consider a radiology trainee who has repeatedly seen one artefact paired with disease because of a biased teaching archive. The cue is familiar, the response is fast and the same answer would likely be produced again. Confidence can therefore be high even when the diagnosis is wrong in a new hospital. Calibration is never a private property alone. It is a relation among the decider, the sampled task distribution and the outcome definition.

The measurement can also alter what it measures. Asking for confidence may slow a decision, redirect attention or encourage an extra check. A review of such reactivity finds that confidence elicitation can affect cognition and performance, with effects depending on the task and procedure.[3] A confidence protocol must therefore state whether the report is part of the operating system or merely an observation. If every production decision will include a confidence step, reactivity is not automatically a confound. It is part of the configured process. If the research question concerns unprompted cognition, the same step may change the target.

Confidence is trustworthy only at the resolution at which it has been calibrated. A person can be well calibrated for familiar visual inspection and poorly calibrated for rare strategic forecasts. A model can be calibrated among four answer choices and fail when allowed to abstain, use tools or generate free text. The outcome, interface and decision policy belong inside the claim.

Metacognition is also not one global personal trait. Recent work comparing decision-making across domains reports only weak generality and evidence for partially distinct perceptual and cognitive factors.[4] Survey claims such as “I know when I am wrong” can diverge sharply from confidence measured during tasks; one study found self-ratings misaligned, and often negatively associated, with online measures in its tested tasks.[5] The practical lesson is severe: calibrate the monitor in the task family where its confidence will govern action.

flowchart LR
  A["The reason for a choice you<br/>did not make"] --> B["The mind that knows and invents"]
  B --> C["Two questions hidden inside<br/>introspection"]
  C --> D["How reasons become stories"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Two questions hidden inside introspection

Why is causal self-explanation harder than confidence? A decision can be produced by many processes that never appear as reportable objects: perceptual fluency, associative priming, learned motor tendencies, affective valuation, attentional capture, or a slowly accumulated preference. The system may nevertheless have access to the finished choice, its felt ease, the public alternatives and a culturally familiar vocabulary of reasons.

When asked “Why?”, it can build a plausible bridge from those accessible materials. The bridge may be informative. It may identify an explicit rule that really guided the decision. It may also be a post-choice reconstruction that fits the outcome without tracking the producing cause.

Nisbett and Wilson’s classic review argued that people often lack direct introspective access to higher-order cognitive processes and instead report theories about likely causes.[6] That claim should not be inflated into “people know nothing about their own minds”. Ericsson and Simon offered a more selective account: carefully elicited concurrent reports can reveal information currently held in accessible memory, while explanations that require inference or long retrospective reconstruction deserve weaker treatment.[7]

The difference is temporal as well as conceptual. A report captured while evidence is active can name what the person notices. A report requested after commitment, feedback and social interaction can incorporate all three. The word “introspection” hides that changing instrument.

The swapped choice

Choice-blindness experiments make the dissociation vivid. In an influential study, participants selected the more attractive of two faces. On some trials, the experimenter covertly presented the rejected face as if it had been chosen. Participants often did not openly report the mismatch and nevertheless gave reasons for the displayed face.[8] The result demonstrates that a coherent reason can be generated for an outcome that does not match the recorded choice.

Yet even this case is not a simple verdict on introspective blindness. A later pupillometry study found increased pupil dilation on manipulated trials even when participants did not report the mismatch. The authors argued that covert detection, reporting hesitation and experimental design could explain part of the original pattern, and they challenged broad claims that people failed to notice.[9]

This boundary strengthens rather than weakens the central thesis. An overt statement such as “yes, that was my choice” can fail for at least two reasons: the mismatch was not detected, or it was detected but not reported. Silence is behaviour, not transparent absence of awareness. A disciplined study must separate detection, report policy and causal explanation.

What a verbal report can still do

Rejecting introspective omniscience does not make self-report disposable. A report can directly reveal current experience, accessible content, intended policy, uncertainty, pain, imagery, or a consciously rehearsed rule. Schooler’s distinction between ongoing experience and explicit meta-conscious re-representation shows why reflection may capture, distort or interrupt what it represents.[10] The correct response is not silence. It is typing the claim.

A report becomes disciplined when its type, time and validation path are explicit. That rule protects first-person evidence from two opposite mistakes: treating it as an oracle, or discarding it as noise.

Confabulation is not a synonym for lying

A lie requires a relation between belief, assertion and an intention to mislead. Confabulation, as used here, is an evidential diagnosis: an account is offered with more causal authority than its support warrants. The speaker may be sincere. The account may even contain real causes. The failure is that the available method cannot distinguish those causes from a story generated after seeing the outcome.

This distinction changes how a reviewer should respond. Accusing the speaker of dishonesty invites defence and adds another social pressure to the report. A better question is operational: “What observation did you have before the choice, and what change would have made you choose differently?” The first request moves the account closer to time-locked evidence. The second turns a story into a candidate counterfactual.

Disciplined first-person methods exploit this asymmetry. They shorten the interval between event and report, ask for concrete particulars before abstract interpretation, repeat sampling and preserve uncertainty. Such methods still do not reveal every cause. They improve the conditions under which experience and accessible content can constrain a theory. Their value lies in narrowing the claim, not in granting private certainty.

How reasons become stories

We can now state the mechanism without pretending to have located two literal organs. Call it the monitor-narrator split. The monitor is any process that turns decision-relevant signals into an estimate of likely correctness. The narrator is any process that turns accessible cues, the chosen outcome, background theories and conversational demands into an account that makes the action intelligible.

Both are useful. Both can fail. The decisive point is that they have different input contracts.

The notation does not explain a brain. It compresses a testable separation. If an intervention changes accessible cues or social context while leaving decision evidence stable, reasons may change while confidence stays fixed. If outcome feedback improves calibration while access to the producing trace remains unchanged, confidence may improve without better causal explanations.

The signature dissociation

Plot metacognitive calibration on one axis and reason fidelity on another. Four regions appear. A system can be calibrated and causally traceable. It can be calibrated yet untraced. It can offer stable reasons while misjudging correctness. Or it can fail both.

This two-axis view is more informative than a single “self-awareness” score. It also creates a clean experimental prediction: manipulations that reduce access to decision traces should lower reason fidelity more than calibration; manipulations that remove outcome feedback should degrade calibration even when a person’s preferred explanation remains fluent.

A fidelity score that asks the right question

A causal reason should respond to changes in a factor that really influenced the choice and resist changes in irrelevant presentation. Let the causal sensitivity be the proportion of reason reports that change when the true driver is intervened upon. Let surface sensitivity be the proportion that change when only wording, colour or display order changes. A simple diagnostic is:

A score near 1 is strong under this particular test. A score near 0 means the report changes no more for the cause than for the decoration. A negative score means the story follows surface features more readily than the putative driver. This is not a universal psychological scale. It is a property of a configured contrast.

The same distinction in human-AI work

A language model can output a probability-like confidence and a rationale in the same response. That proximity tempts us to treat both as readings of one internal state. The evidence does not justify that shortcut. Some studies find that model confidence can be calibrated under particular elicitation formats, while other work finds strong overconfidence and method dependence in verbalised confidence.[11][12] Calibration procedures using held-out labelled examples can improve the mapping, but prompting alone does not guarantee it.[13]

The machine comparison is structural and bounded. A human confidence report and a model confidence output need not share mechanism, experience or ontology. The preserved relation is narrower: both can be evaluated against outcomes, and both can be mistaken for causal provenance. A rationale generated after an answer is a model proposal about cause until an intervention connects it to the computation.

stateDiagram-v2
  [*] --> Claim
  state "Calibration is not eloquence" as Claim
  state "What confidence may actually<br/>monitor" as Test
  state "One accuracy, two monitors" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

A discipline for self-report

The practical response is a dual-ledger protocol. The uncertainty ledger records what the decider expected before the outcome. The provenance ledger records candidate causes and tests them against interventions. Neither ledger is allowed to borrow authority from the other.

This protocol is intentionally modest. It does not promise complete access to human cognition or model internals. It prevents a more common error: treating one fluent retrospective paragraph as confidence estimate, causal mechanism, moral justification and experience report all at once.

The seven moves

The smallest useful implementation

A team does not need eye tracking, brain imaging or a complete causal model to improve its evidence. Start with one repeated decision class. Define the outcome that makes confidence meaningful. Capture a probability before feedback, along with the leading alternative and one sentence naming the evidence expected to change the decision. Store those fields separately from the later rationale.

After enough cases, produce a reliability diagram and a Brier score. Do not pool unlike tasks merely to create a larger sample. Then select a small number of disputed decisions for matched replay. Change one claimed driver while preserving presentation, and change one presentation feature while preserving the driver. Record whether the choice and reason move. The result is not a personality score. It is an assurance result for one decision process.

The data model can remain compact: decision identifier, actor or configured model, task class, alternatives, pre-outcome confidence, evidence references, chosen action, outcome, retrospective account, intervention identifier and typed conclusion. Identity and timestamps matter because later discussion can otherwise be mistaken for earlier knowledge. For machine systems, record model and prompt versions. For human teams, record role and information access without pretending that a form captures the whole person.

Failure conditions should be designed in advance. If outcomes are delayed or disputed, confidence cannot yet earn release authority. If a causal feature cannot be changed safely, use simulation or retain the reason as an unverified hypothesis. If prompting for confidence changes performance materially, evaluate the prompted workflow rather than projecting its calibration onto an unprompted one. If repeated review causes strategic gaming, rotate test cases and separate coaching from assessment.

The protocol is valuable even when it returns “unknown”. Unknown means the organisation has located the missing bridge: outcomes, temporal records, discriminating interventions or an independent path. That is more useful than a persuasive rationale whose evidential role was never specified.

function simulateTrial(config, random) {
  const causalCue = random() < 0.5 ? -1 : 1;
  const surfaceCue = random() < 0.5 ? -1 : 1;
  const evidence = config.skill * causalCue + normal(random);
  const choice = evidence >= 0 ? 1 : -1;
  const correct = choice === causalCue ? 1 : 0;

  const idealConfidence = 0.5 + 0.49 * Math.tanh(Math.abs(evidence) / 2);
  const confidenceNoise = normal(random) * (1 - config.monitor) * 0.18;
  const confidence = clamp(0.5, 0.99, idealConfidence + confidenceNoise);

  const reasonMatchesCause = random() < config.trace;
  const statedReason = reasonMatchesCause ? "causal cue" : "surface cue";

  const swapped = random() < 0.20;
  const internalDetection = swapped && random() < clamp(0, 1, 0.55 + 0.35 * config.monitor);
  const reportedSwap = internalDetection && random() < config.assertiveness;

  return { causalCue, surfaceCue, choice, correct, confidence,
           statedReason, reasonMatchesCause, swapped,
           internalDetection, reportedSwap };
}

How to apply it outside a laboratory

Ask for confidence before asking for reasons. Preserve the alternative that nearly won. Record what evidence was available at that moment. Later, score confidence against outcomes. For reasons, select one claimed driver and one irrelevant surface feature, then replay matched cases with each changed separately. Do not reward rhetorical richness until the causal contrast has passed.

In high-stakes work, the operational rule is simple. Confidence may determine review depth only after calibration. A reason may determine attribution, remediation or release only after an independent causal test. When intervention is impossible, retain the report as a hypothesis and lower its authority.