When a social feed attaches an AI credibility label ("likely false") to a post, many readers stop using their own judgment and simply copy the label. A Bayesian analysis of sequential news-veracity judgments found that about 20% of people weight the AI above their own impression. It also found that following a 55%-accurate AI pushes crowd accuracy below what the crowd achieves with no AI at all, because copied votes carry no independent information (One AI Signal, Many Human Judgments). That paper infers the reader's "private impression" only indirectly, from the votes themselves. Once a reader cascades (copies the label instead of voting their own view), the impression is invisible by construction. We propose to measure it a different way. A reader's pupil dilates involuntarily when incoming information conflicts with what they expected. If the AI label disagrees with what the reader privately believes, the moment the label appears should produce a measurable mismatch dilation, even when the reader then goes along with the label.
The claim is that this label-evoked pupil response carries information about a claim's truth that the reader's explicit vote no longer contains. It should be strongest exactly where cascades do the most damage, on "gray-zone" claims rated half-true or mostly-true, where binary labels fail (A Survey on Automated Fact-Checking). If the claim holds, a pupil-decoded dissent signal could restore independent human information to crowd judgments. If it fails, the failure is still informative: it would show that cascading readers do not merely hide their doubt but never form it.
Core Claim
Nuanced misinformation detection supplies the setting and the ground truth for this work. Physiological stress monitoring supplies the measurement. We put two testable statements at the center of the project. The first is that pupil dilation 0.8 to 3 seconds after an AI veracity label appears grows with the distance between the label and the reader's own pre-label belief, even on trials where the reader's final vote matches the label. The second is that this signal, decoded per trial, adds information about the gold multi-class veracity label beyond everything else available: the reader's vote, their response time and gaze, and the claim text.
Two pupillometry results motivate the mechanism. The first is a picture-word study in which mismatched word-image pairs produced reliably larger dilation starting about 800 ms after stimulus onset. The first principal component of the pupil curve captured 78% of variance as response amplitude (When a Dog is a Cat). The second is a set of reading studies in which the frequency of rapid small dilations indexed semantic and discourse anomalies at about 1 s latency (The Frequency of Rapid Pupil Dilations). Two neurophysiology results support reading the pupil as a measure of internal state. Pupil fluctuations track cortical noradrenaline and acetylcholine (Reimer et al.), and they reflect latent population states in cortex (Johnston et al.).
One reading result constrains the design. These latencies make word-level attribution in continuous reading unreliable. For that reason we deliberately do not try to localize deceptive tokens with the pupil. We use a single, time-locked event per trial, the onset of the AI label, which matches the event-locked designs where pupillometry is known to work.
Relation to Prior Work
The closest work, The Pupil Becomes the Master, turns fixation and scanpath features into preference pairs for tuning language models on sentiment and sarcasm. It does not touch factual veracity, it does not use pupil dilation, and it treats gaze as a training signal for a model. We instead treat the pupil as a measurement of a latent human belief inside a social aggregation problem. That paper also reports that simple ocular metrics alone gave no significant gains, which warns us that cheap gaze features may be weak. We therefore test them explicitly as a baseline.
Are current models "fact-checking" or "gut-checking"? uses "gut-checking" as a metaphor for models that exploit affective wording in the text. That raises a real confound for us: arousal could reflect emotionally charged claim wording rather than belief conflict. We control for it by locking the analysis to label onset, by baseline-correcting against the pre-label period, and by entering the text affect score as a covariate.
CognitIDE shows that pupil, skin-conductance and EEG streams can be synchronized with on-screen content through Lab Streaming Layer, a standard protocol for time-aligning sensor streams. We reuse that infrastructure pattern; the application is new.
Text-side work fixes the stimulus regime. MultiFC provides real claims with fine-grained fact-checker labels, and its best evidence-based model reaches 49.2% macro F1. X-Fact shows that fine-grained classifiers collapse to roughly 15–16% macro F1 out of domain. Together these results show that the AI labels readers see in practice will often be wrong on gray-zone claims, which is the regime where an independent human signal matters most.
Research Program
We build a laboratory paradigm around a sequential social-voting task modeled on the design analyzed in One AI Signal, Many Human Judgments. The stimuli are about 160 English PolitiFact and Snopes claims drawn from MultiFC. They are balanced across the six-point scale from "true" through "half-true" to "pants on fire," and matched for length and for NRC affect intensity.
Each trial runs as follows. The claim appears on a display with fixed luminance and fixed layout. After reading, the participant fixates a central point while a brief, luminance-matched AI label appears. The participant then gives a three-way vote (true, mixture, false) and a confidence rating. We set the AI label's accuracy by design and run two blocks, one at 55% accuracy and one at 80%. Some labels are uninformative ("AI unavailable"); these serve as a control for generic arousal at label onset.
In half the blocks, participants also give a private rating before the label appears. These elicitation blocks serve as the calibration set for the decoder. The other half, with no pre-label rating, represent the deployment condition, since a real feed cannot ask for a private rating before showing a label.
Pupil traces are blink-interpolated and baseline-corrected to the 500 ms before label onset. We summarize them with functional principal components as in When a Dog is a Cat. We also compute an open-source wavelet estimate of rapid-dilation frequency, because the original Index of Cognitive Activity depends on proprietary software. A hierarchical decoder takes these pupil features plus participant random effects and is trained on the elicitation blocks to predict the signed distance between the private rating and the label. We then apply it to the no-elicitation blocks. Finally, we feed the decoded "dissent" into a cascade-aware Bayesian aggregator. In that aggregator, a vote judged likely to be a cascade has its weight reallocated toward the decoded private belief.
We plan 60 participants. This is about ten times the sample of the picture-word study, a margin chosen because the effect on factual claims is likely smaller and noisier than on lexical mismatches.
Baselines
The naive reference point is plain majority voting over the explicit votes of 11-person sequences. We construct these sequences by resampling our participants in the label order used by Lu et al. Under the 55%-accurate label, this baseline should reproduce the reported effect that following the AI degrades crowd accuracy.
The strongest existing method is the behavior-only Bayesian cascade model from One AI Signal, Many Human Judgments. It estimates each participant's weighting of the AI versus their own impression from their choice history and discounts likely cascade votes. The file reports its fitted population structure: 20% of participants are cascade-prone, 46% are learners, and the AI is weighted as 3–6 peer judgments. No reported result was found on Lacuna for its crowd-accuracy recovery. We will therefore fit it ourselves on the same sequences, with the same cross-validation folds and the same hyperparameter search budget as our aggregator. Beating it means a statistically reliable gain in macro F1 against the collapsed three-way gold labels on held-out items, concentrated in the 55%-accurate block and on half-true and mostly-true claims.
The cheap-behavior baseline adds response time, total dwell time, regression count and fixation count to the decoder in place of pupil features. These are the features that The Pupil Becomes the Master found weak when used alone.
The text-only baseline predicts per-item cascade rate and gold label from claim text, using a MultiFC-style evidence model. The file reports 49.2% macro F1 for the best such model; on our three-way collapse we will report our own figure.
The ceiling is the explicit pre-label private rating, which is available only in the elicitation blocks.
Two ablations isolate the key ingredient, which is label-locked mismatch. The first shuffles pupil traces across trials within each participant. The second replaces mismatch-locked features with the response to the uninformative label. Pupil features pass the test only if they help in the full model and fail to help after shuffling.
Why not just…?
Why not just ask people for their private rating before showing the label? This is the real threat, and on raw accuracy it will probably win. That is why it is our ceiling, not a competitor we claim to beat. The proposal's value rests on two conditions that the experiment tests directly. The first is that real feeds cannot elicit a rating before showing a label. The second is that elicitation itself changes behavior: committing to a rating may make people less susceptible to cascades, so comparing cascade rates between elicitation and no-elicitation blocks measures this. If the pupil decoder recovers little of the gap between the behavior-only model and this ceiling, we will report that physiology adds nothing in practice.
Why not just use response time and gaze, which webcams can capture? Hesitation alone could plausibly signal doubt. The cheap-behavior baseline and a combined model settle this: the pupil must add predictive power on top of response time and gaze, not merely match them. If it does not, the practical conclusion is to use the cheaper signals.
Why not just model cascades from item text and behavior? Item-level difficulty explains part of who cascades on which claims. But the text-only and Bayesian baselines can only capture differences between items or between people. The pupil claim is about variation within a person on a single item. Our decoder is evaluated on the trial-level residual left after item and participant effects are removed, so this alternative cannot account for any gain we report.
Decisive Checks
Three results would confirm the idea. In the elicitation blocks, label-evoked dilation amplitude should rise monotonically with private-belief–label distance, including on trials where the final vote matched the label. That effect should disappear for uninformative labels. And in the no-elicitation blocks, pupil-augmented aggregation should beat the behavior-only cascade model on held-out gray-zone items under the 55%-accurate label.
A null on the first check, with distance effects present only when readers vote against the label, would show that cascading readers form no private dissent. That result would bear directly on how credibility labels should be designed. A positive result whose pupil effect vanishes once claim affect is controlled would show the signal is "gut-checking" in the literal sense rather than belief conflict.
References
- One AI Signal, Many Human Judgments: A Bayesian Cascade Analysis of AI-based Credibility Indicators in Online Information Spread
- A Survey on Automated Fact-Checking
- When a Dog is a Cat and How it Changes Your Pupil Size: Pupil Dilation in Response to Information Mismatch
- The Frequency of Rapid Pupil Dilations as a Measure of Linguistic Processing Difficulty
- Pupil fluctuations track rapid changes in adrenergic and cholinergic activity in cortex
- The eyes reflect an internal cognitive state hidden in the population activity of cortical neurons
- The Pupil Becomes the Master: Eye-Tracking Feedback for Tuning LLMs
- Automatic Fake News Detection: Are current models “fact-checking” or“gut-checking”?
- Using CognitIDE to Capture Developers' Cognitive Load via Physiological Activity During Everyday Software Development Tasks
- MultiFC: A Real-World Multi-Domain Dataset for Evidence-Based Fact Checking of Claims
- X-Fact: A New Benchmark Dataset for Multilingual Fact Checking