Restoring sight in choice blindness: pupillometry and behavioral evidence of covert detection

Frontiers in Psychology 16:1598254, doi 10.3389/fpsyg.2025.1598254; open access (CC BY 4.0), PMC12713319. Read for research direction R2, 4 October 2026. Provenance: papers/amadeus/grassi2025_choice_blindness_pupil.provenance.json.

What was read

• Read in full: the PMC XML converted to text: abstract, introduction, methods for both experiments, results, figure captions and alt texts 1 to 5 (figures are images), discussion, limitations, conclusion, statements and the reference list.
• Not read: the online supplement, which holds the verbatim transcripts (Tables S1 to S5), pupil statistics tables (S6, S7), trial counts (S8, S9), the power simulation and per-subject pupil plots.

Background (as the paper states it)

"Choice blindness": people choose between two options, are shown the option they did not choose as if it were theirs, and are asked why they chose it. In the original face study only 13% of switches were called out. Later studies extended this to taste, smell, moral attitudes and political attitudes, and report preference change and memory distortion afterwards. The standard reading, after Nisbett and Wilson (1977), is that people lack introspective access to their own choices and confabulate reasons. Kurisutina's P6 relies on this literature in spirit: people misreport themselves.

What they did

• Experiment 1: 41 adults (35 with usable pupil data). 30 trials of "which photo was rated more beautiful":
  • 8 manipulated trials at fixed positions: the unchosen photo shown as the choice for 5 s, then "please explain your choice";
  • 8 unmanipulated trials with explanation;
  • 14 baseline trials.
  • Half the sample got a small financial incentive. Photo pairs were similar or dissimilar in rated beauty. Eye tracking ran throughout.
  • After the task: "did you notice anything?", debriefing, then a questionnaire on noticing and on reasons for not reporting, and a surprise memory test to pick out the manipulated pairs.
• Experiment 2: 23 adults (21 with pupil data), monochrome female faces (attractiveness), the classic stimulus.
• Coding. Verbal reports were coded by two raters (κ 0.94 and 0.83) as concurrent report, possible detection or no report. Correct identification in the memory test counted as a retrospective report.
• Pupil analysis. Responses 2 to 4 s after the outcome, with mixed models, plus a leave-one-subject-out threshold classifier.

Main results (verified)

Experiment 1 (colour photographs)

• People reported about half the switches but knew most of them. 56% of manipulated trials (185 of 328) were reported during the task, yet 84% were picked out afterwards.
  • Together 92% were reported one way or the other; only 7% were never reported.
  • Of the switches not reported during the task, 83% were correctly identified afterwards.
  • Retrospective discrimination was d′ = 2.73 (hit rate 0.84, false alarms 0.03).
• Why people stayed silent.
  • Those who reported some switches mostly thought they had pressed the wrong key or that the experiment had a fault.
  • 9 of the 10 who never reported thought the switches were part of the experiment.
  • 8 of those 10 said they gave reasons for the presented option anyway. They knowingly gave reasons for a choice they had not made.
  • Participants estimated their own number of concurrent reports almost exactly (mean error 0.25).
  • Up to 33 of 41 people (80%) showed covert detection.
• What did and did not matter. Reporting rose with trial number (odds ratio 1.90) and with confidence (1.48). The incentive and the stimulus similarity had no effect.
• The pupil registered the switch whether or not the person said anything.
  • Manipulated against unmanipulated trials: β = 0.58.
  • Reported switches β = 0.65; retrospectively reported β = 0.53.
  • Never-reported switches: β = 0.25, not significant, from only 16 trials in 11 people.
• Classification across people: switch against no switch 65.5% (p < .001). Whether the person would report the switch: 51.4% (chance).

Experiment 2 (faces)

• Concurrent report 60%. Retrospective identification only 43% (d′ = 1.23); faces are hard to remember. 26% of switches were never reported.
• Similar faces were reported less often (odds ratio 0.28).
• The pupil dilated on switches even when they were never reported (β = 0.53, p = .01).
• Classification of switch against no switch 60.8%; prediction of reporting at chance (45.9%).

Across both experiments, an estimated 65.5% of unreported switches (135 of 206) were covert detections.

Limits

• Small samples for the post-hoc subgroups. The "never reported" pupil estimate in Experiment 1 rests on 16 trials.
• Retrospective identification is an indirect measure of detection at the time. It depends on how memorable the stimuli are, as Experiment 2 shows.
• These were computerised tasks. Classic choice-blindness studies used sleight of hand, where reasons for silence may differ, for example assuming one misremembered or not wanting to accuse the experimenter. The authors argue covert detection should still occur there, but this is untested here.
• The pupil effect is a generic surprise or arousal signal: about 61 to 66% trial classification, not specific to self-knowledge.

What it means for Amadeus (inference)

• The headline evidence for "people confabulate their own reasons" needs a footnote. Much apparent introspective blindness here is detection without report: people notice, then go along out of politeness or assumption, or change their minds when indifferent. That changes what the "say" channel is. A report reflects the person's knowledge filtered through the social cost of saying it, not their knowledge alone.
• A bodily signal knew more than the words, and only that much. Pupil dilation flagged the switch at trial level, including unreported ones. It could not tell who would speak up. Neural and physiological data can show that something registered. They cannot show what the person believes about it or why they stayed silent. That is the general shape of what brain evidence can add to P6.
• Three cases P6 must keep apart. A gap between what a person said and what their body registered can be:
  1. not detected (a genuine belief);
  2. detected but not said (strategic or social, i.e. contamination);
  3. a change of mind (a real update). P6 treats self-deception as part of the person and stores it as believed. Only case 1 fits that. Asking later, privately and with low pressure, recovered most of case 2: 83% in Experiment 1.
• Elicitation design rule (the authors' own suggestion, transferable): make disagreeing or correcting as cheap as going along. In the interview, the person should be able to correct a summary of themselves at no social cost. This is the human analogue of the drift battery: people, too, accept an outcome presented to them when contradicting it costs something.