Eye of the beholder: Pupillary response reflects how subjective prior beliefs shape reinforcement learning with fake news

Significance

In an age where misinformation spreads rapidly and online environments favor belief-consistent information, understanding how people learn from and act upon misleading content is a scientific and societal priority. This study integrates fake news evaluation, reward-based learning, and pupillometry to uncover how veracity judgments and confidence shape choice behavior. We show that participants readily follow reinforcement when it aligns with their prior beliefs, but struggle to adapt when rewards favor confidence. Crucially, pupil dilation reveals that these effects emerge even before decisions are made, reflecting early engagement of confidence-related mechanisms. These findings suggest that rigidity in belief-related choices may arise from predecisional reinforcement dynamics that stabilize existing epistemic structures, highlighting cognitive pathways through which misinformation can persist.

Abstract

Information selection plays a crucial role in how individuals navigate online content. While confirmation bias has been implicated in this phenomenon, its interaction with reinforcement learning dynamics and internal confidence signals remains poorly understood. Here, we examined how veracity judgments and confidence shape choices when probabilistic rewards are tied to different epistemic attributes of news headlines. Participants completed a three-phase paradigm that combined news classification, a probabilistic learning task with varying reward contingencies, and a final reevaluation phase. Using real and false headlines judged for veracity and confidence, we created personalized sets of stimulus categories that were later used in a two-armed bandit task. In different blocks of trials, reinforcement was probabilistically associated with either the perceived truthfulness or confidence of each item. Across all experimental phases, pupil dilation provided neurophysiological signatures of belief-related processing. At a behavioral level, participants showed higher accuracies and learning rates when rewards were contingent on their previous judgments of veracity, whereas performance was markedly reduced when reinforcement favored confidence, especially low-confidence options. Pupillometric data revealed predecisional modulations tied to subjective confidence, while computational modeling showed that participants relied on feature-based generalization when veracity predicted reward and shifted toward valence-sensitive updating when contingencies no longer matched their prior epistemic structure. Together, these results reveal how veracity and confidence jointly guide reinforcement-driven choices and modulate the flexibility of belief-related decisions. By integrating cognitive, computational, and physiological data, our study provides a mechanistic understanding of how prior beliefs shape learning in complex and misinformation-rich contexts.

Social networks are often used as a primary source of news and political information (1), even though they often provide news without filtering options or preliminary control for quality standards. As a result, we are witnessing an era characterized by an increased spread of “fake news” articles containing false, discrediting, or misleading information that heavily impacts politics, democracy, and society (2). For these reasons, disinformation has become a serious and urgent concern worldwide. Policies like user reports and fact-checking have been adopted in response to these adverse effects. However, these methods alone have so far proven unsuccessful in combating disinformation (3, 4), and critical questions remain unanswered: Why do people believe in fake news?

Humans often tend to search for, interpret and recall information in a way that confirms their prior beliefs rather than considering contradicting information (5). Such a phenomenon is defined as “confirmation bias” (CB) and significantly impacts domains ranging from politics (6) to social science (7), finance (8), justice (9), and education (10). Additionally, CB is considered one of the leading causes of disinformation in social media (11). This is because users of social networks are not allowed to intentionally choose the source of articles presented on their newsfeed, which are instead selected by “algorithmic editing” designed to display only information users are more likely to agree with while excluding opposing views and contents (12).

Previous psychological studies showed that CB is a “pervasive” (13), “ineradicable,” and “ubiquitous” feature of human thinking (14) found in both concrete and abstract reasoning (15), independently of subjects’ intelligence, cognitive ability, or motivation to avoid it (16), and even low-level, automatic perceptual processing could be susceptible to CB (17).

Why such a pervasive bias, often leading us to hold inaccurate beliefs, has not been selected away by evolution, remains controversial. Some philosophers argue that CB might merely be a by-product of the evolution of another adaptive trait (18). Others, however, believe that such bias might have significant advantages, being somehow “adaptive” at both the individual and social levels (19). A different and more interesting scenario suggests that CB is rooted in reinforcement learning (RL) mechanisms that prioritize consistency between expected and observed outcomes. For instance, previous studies have shown that humans display a learning asymmetry caused by a significant confirmatory bias (13). From this perspective, selective weighting of belief-congruent information may serve adaptive functions in noisy environments, where uncertainty makes consistent reinforcement a useful heuristic for stabilizing predictions (20). For example, Hoxha et al. (21) demonstrated that learning asymmetries, such as positivity bias and gradual perseveration, can spontaneously emerge in reinforcement-learning (RL) systems, thereby improving learning stability under uncertainty or providing smaller, context-dependent gains. In the same vein, the influential theory of active inference (22) frames human behavior as a constant attempt to fulfill “optimistic” predictions based on (sensory) evidence that is evaluated in relation to a generative model of (observed) outcomes.

However, in structured ecological contexts, such as those encountered during media consumption, confirmation-driven information selection may limit exposure to belief-incongruent content, thereby promoting the persistence of inaccurate or misleading representations. Repeated encounters with false or biased information (e.g., political misinformation, health-related conspiracy theories) can further encourage a preferential weighting of belief-consistent cues or interpretations (11, 23, 24), even when competing information is available. Although such preferential selection may support efficient decision-making in noisy environments, it can become maladaptive in informational ecosystems designed to amplify preexisting preferences rather than challenge them.

Taken together, these insights raise a broader question: How do preexisting beliefs shape learning when individuals must integrate external feedback with their own epistemic expectations? RL contexts provide an ideal setting for examining this interaction, as they pit belief-guided expectations against reward-guided adaptation.

Here, we hypothesized that veracity judgments and confidence can jointly modulate how prior beliefs interact with reward contingencies within a RL framework. Specifically, we expected belief-congruent learning to be enhanced when external rewards aligned with participants’ initial veracity judgments, reflecting belief-congruent reinforcement, and to be reduced when reinforcement favored confidence-based choices, particularly under low-confidence conditions. These latter cases would reveal the limitations of adaptive learning in the presence of strong epistemic assumptions/priors. To these aims, we designed a multiphase experimental paradigm inspired by real-world scenario. In a first phase, participants judged the veracity (True vs. Fake) of news headlines and indicated their level of confidence by wagering a virtual amount of money. This allowed us to categorize each item into one of four subsets: true or fake news judged with either high or low confidence. Participants then engaged in a RL task in which they chose between pairs of previously seen headlines and received probabilistic feedback based on each headline’s category. Afterward, participants were shown all previously judged headlines along with their initial responses and wagers and were given the opportunity to revise their judgments.

To model learning dynamics and formalize our predictions, we implemented a feature-based Rescorla–Wagner model that learned separate values for the constituent dimensions of each headline, namely veracity and confidence. Second, drawing on prior work on confirmation-biased RL (13), we implemented an asymmetric variant of this model, which includes distinct learning rates for positive and negative prediction errors. Together, these models enable us to distinguish whether belief-congruent choice dynamics arise from feature-based generalization mechanisms or from asymmetric reinforcement dynamics, providing complementary insights into how confidence and veracity influence learning.

In addition, we sought to explore whether psychophysiological markers, specifically pupil dilation, could provide a continuous, implicit index of the subjective confidence underlying the interaction between beliefs and RL. Prior research has shown that pupil size reflects attention, arousal, and evidence accumulation during decision-making, and is sensitive to value-based learning and belief commitment (25, 26). In this light, pupil-linked signals may serve as a window into the metacognitive processes that support confidence-based reinforcement (27). This notion is further supported by Rollwage et al. (28), who demonstrated that subjective confidence drives a neural CB, such that higher confidence increases the weighting of belief-congruent information during decision-making. Consistent with this view, Pajkossy et al. (29) demonstrated in a probabilistic reversal-learning task that pupil-linked arousal tracks violations of learned contingencies, reflecting ongoing adjustments to the internal model.

This integrated design offers a comprehensive approach to examining the joint contribution of reinforcement contingencies, confidence, and physiological arousal on belief-congruent learning. By linking behavioral, computational, and pupillometric evidence, it can shed light on the mechanisms through which subjective confidence biases information processing and value learning in uncertain or deceptive contexts.

1. Materials and Methods

1.1. Participants.

The required sample size to obtain statistically significant and reliable results was determined in advance using G*Power software (version 3.1.9.6) (30). Considering the minimum number of levels in each dependent variable (main effects) and an effect size of f² = 0.25, the analysis indicated that at least 24 participants would be required to achieve a power of 0.80, with α = 0.05 (two-tailed) for repeated-measures ANOVAs. Similarly, for one-sample and paired-sample t tests, power analyses showed that a minimum sample size of 22 and 28 participants was respectively required to detect an effect size of d = 0.5. To ensure these estimates remained robust under more realistic assumptions of noise and individual variability, additional simulations were conducted using the simr package (31) in R (v. 4.5.1). A series of linear mixed-effects models (LMMs) with experimental block as within-subject levels and random intercepts for participants were used to estimate the statistical power to detect a medium effect size (f² ≈ 0.25). These simulations, run over 1,000 iterations per sample size, confirmed that a minimum of 28 participants was sufficient to achieve 80% power, even under nonideal conditions such as unequal differences between blocks and moderate measurement noise (SI Appendix, Fig. S1).

Based on these analyses, 31 right-handed healthy young adults aged between 18 and 36 y (mean = 23.77, SD = 4.24; M = 10, F = 21) were recruited. None of the participants were using psychotropic medications or substances of abuse, nor were they undergoing treatment or had a history of hospitalization for psychiatric disorders or neurological diseases. During the first experimental phase (Section 1.2), three participants did not reach the minimum required number of judged news items per category (n = 48) and were therefore excluded from ensuing analyses. Consequently, the final sample consisted of 28 individuals (mean = 23.71, SD = 4.37; M = 9, F = 19). All participants provided informed consent to participate in the study, which was approved by the Sapienza University Ethics Committee (protocol no. 43/2023) and was conducted following the principles outlined in the Declaration of Helsinki.

1.2. Tasks and Stimuli.

1.2.1. Behavioral computerized task.

All stimuli consisted of textual headlines from both real and fake news sources. The stimulus set included 324 fake (n = 162) and real (n = 162) news circulated on Facebook and Twitter between 2020 and 2023. Fake news were identified using reliable international and national fact-checking websites (e.g., “Snopes.com,” “Facta.news,” and “Open.online”). To ensure that real and fake news items had similar levels of engagement across the selected platforms, all news were compared using BuzzSumo.com. All headlines did not differ in length and number of characters and were categorized into nine thematic sections (SI Appendix). All stimuli were isoluminant and presented as white against a gray background on an LCD monitor (23,8’) with a 60 Hz refresh rate and a 1,920 × 1,080 resolution.

The experiment was structured into three distinct phases, each designed to assess different aspects of belief processing: i) Mapping Phase—evaluating the credibility of news items (Fig. 1A), ii) Learning Phase—examining the role of CB in RL (Fig. 1C), iii) Feedback Phase—assessing eventual belief revision (Fig. 1E).

Fig. 1.

Outline of the three phases used in our task: (A and B) Mapping phase, in which participants judged the veracity of news headlines (real/false) and rated their confidence by placing a monetary bet, yielding to four news categories (TH, TL, FH, FL); (C and D) Learning phase, where participants chose between headline pairs and received probabilistic monetary feedback tied to specific category contingencies; and (E) Feedback phase, where participants made final veracity judgments and received final feedback.

1.2.1.1. Mapping phase.

In each trial, participants were randomly presented with one of the headlines and asked to judge whether the headline was real or fake. Successively, they were asked to wager a virtual amount (from 0 to 99 cents) on the accuracy of their judgment, serving as a confidence measure. While there was no overall time limit for responses, participants could only respond after 6 s from the headline’s presentation to encourage careful reading. A change in the fixation point’s color (isoluminant green) signaled when they could provide their answer. Confidence scores above the 75th percentile and below the 25th percentile were classified as high and low confidence, respectively, allowing to generate four subsets of 48 items each (Fig. 1B), including news judged as real or fake, with high (TH and FH) or low (TL and FL) confidence.

1.2.1.2. Learning phase (two-armed bandit task).

In each trial participants were presented with two news headlines selected among the four subsets generated in the mapping phase. Headlines were positioned 7 cm to the left and right of a central fixation point, corresponding to ~6.66° of visual angle at a 60 cm viewing distance. Participants were instructed to freely choose one of the two headlines in the attempt to obtain a virtual monetary reward of 20 cents. After 6 s from the news presentation, a change in the fixation point’s color (isoluminant green) signaled that participants could provide their responses. Also in this case, no overall time limit for responses was provided.

All headlines were pseudorandomly paired in six combinations (Fig. 1D) with a counterbalanced side of presentation. The task comprised five experimental blocks, each consisting of 96 trials (16 repetitions per comparison). Within each block, the probability of receiving a reward (83%) was implicitly linked to either the subjective veracity of the headline (True or Fake; T-Block and F-Block) or the individual’s confidence level (High or Low; H-Block and L-Block) independently (SI Appendix, Table S1). As a control condition, the reward probability was entirely provided at chance in one block (R-Block). Participants were not informed about the reward probability associations and were asked to maximize their earnings. The order of blocks was counterbalanced across participants.

1.2.1.3. Feedback phase.

In the final phase, participants were shown all the headlines from the four subsets created during the Mapping Phase, along with their initial veracity judgments and confidence ratings (Fig. 1E). They were then asked to confirm or modify their initial decision regarding the veracity of each headline, knowing that, depending on their accuracy, they could either retain or lose the virtual money wagered during the Mapping Phase.

1.3. Experimental Procedure.

All participants, successively to eye-tracking calibration, performed the computerized behavioral experiment, which was programmed and executed using PsychoPy (v2023.2.3). Participants were seated in an acoustically isolated room with fixed ambient light, having their heads positioned on a chin rest throughout the experiment to maintain fixation at a distance of 57.3 cm from the monitor. Between different experimental phases, participants were given a 15- to 20-min break; the entire experiment lasted approximately 3 h. At the end of the experimental procedure, a debriefing phase was conducted to investigate any unexpected aspects noticed by participants and to ensure that the experimental objectives had not been misunderstood.

1.4. Pupil Dilation Measurement.

A wearable Pupil Core eye-tracker (Pupil Labs GmbH, Berlin, Germany) with a sampling rate of 120 Hz, was used for recording gaze position and pupil diameter of both eyes in mm using a 3D eye model. Calibration and validation procedures were performed using the built-in programs provided with the eye tracker (Pupil Capture v.3.5.1). All data analyzed in the present study were obtained from recordings with a tracking accuracy of less than 1°.

For each trial, the eye with higher mean 3D model confidence was selected for analysis. To minimize data loss, missing signals caused by eye blinks were reconstructed using a two-stage detection algorithm based on the estimation of the pupil’s velocity profile and linear interpolation (32) (SI Appendix). Baseline normalization was systematically adapted to the temporal structure of the analyzed event. For all stimulus-locked analyses (e.g., feedback-related pupil activity), the baseline was defined as the mean pupil diameter during the 500 ms preceding stimulus onset. For response-locked analyses (e.g., predecisional activity during news evaluation), which were affected by variable duration, pupil traces were aligned to the keypress, and a global baseline was computed by averaging pupil size across the 500 ms fixation period preceding headline presentation. In all cases, trials with baseline values exceeding ±2 SD (computed separately for each eye) were discarded to exclude artifacts related to blinks or transient luminance changes.

Percentage signal change (PSC) was then computed as

$$\mathrm{PSC}=\frac{X-{\overline{X}}_{\mathit{bas}}}{{\overline{X}}_{\mathit{bas}}}·100,$$ [1]

where X is the pupil signal at each time point, and ${\overline{X}}_{\beta a_s}$ is the baseline mean.

Statistical differences between conditions were assessed using cluster-based permutation testing (33), a nonparametric method that controls for multiple comparisons without assuming data normality. For each comparison, t-values were computed at each time point, and clusters of contiguous significant values (P < 0.05) were identified. Cluster-level statistics (T-sum) were then compared against a null distribution generated from 1,000 random permutations of condition labels.

2. Statistical Analyses

2.1. Mapping Phase.

Individual performance in judgments of news headlines was analyzed using the Signal Detection Theory (SDT) model (34). The d’ and c indexes were computed and tested for significance through one-sample t tests against 0. A second set of analyses examined individual average bets as a function of truthfulness judgment (true, false) using a paired-sample t test.

Pupillary responses during news evaluation were analyzed by examining the 2-s period preceding participants’ decisions. In a first analysis, we examined whether pupillary responses immediately preceding each decision reflected the subsequent judgment of a headline as true or false, and the associated level of confidence (high vs. low).

To rule out the possibility that pupillary differences were driven by intrinsic properties of real vs. fake news (i.e., objective veracity rather than participants’ judgments), we conducted a control analysis examining pupil size during the 2 s preceding each decision, comparing objectively real and fake headlines. We also addressed the potential effects of low-level stimulus features on news evaluation by fitting two LMMs, which included news length, operationalized as the number of syllables or characters, as continuous covariates (SI Appendix).

2.2. Learning Phase (Two-Armed Bandit Task).

In the learning task, the accuracy of individual choices was assessed by calculating the percentage of selections of headlines belonging to the category most likely associated with reinforcement in each specific block of trials (SI Appendix, Table S1). These values were then entered in a first LMM including the within-subject factor Block (four levels) as fixed factor, and the random intercepts modeled for each subject to account for repeated measures [Accuracy ~ Block + (1 | Subjects)]. Pairwise post hoc comparisons among all blocks were performed using the emmeans package with Tukey correction for multiple comparisons [robustness confirmed with false discovery rate (FDR) correction].

Complementarily, we fitted an additional LMMs using participants’ choices distribution as the dependent variable, computed as the proportion of selected items as a function of Veracity (T, F) and Confidence (H, L) evaluation across blocks. This model included Block, Veracity, Confidence, and their interaction as fixed effects, and random intercepts for subjects [Choice ~ Block * Truthfulness * Confidence + (1|Subj)].

For all models, marginal and conditional R² values were computed to quantify the variance explained by fixed effects and by the full model (fixed + random effects), respectively. Effect sizes were further estimated using Cohen’s f², and P-values were FDR-corrected.

Selection preferences across blocks were also examined for each of the six headline pairings (Fig. 1D). A “preference index” was calculated by subtracting the selection percentage of the second item from the first (e.g., TH vs. RL: 74 to 26% = 48%), with positive values indicating preference for the first item. One-sample t tests assessed whether each index differed significantly from zero.

Pupillary responses during the Learning phase were analyzed both in the 2,000 ms preceding each decision and during feedback presentation. In the first case, we examined pupil dilation prior to the participant’s choice, comparing responses across the six possible pairwise comparisons between the four news categories (TH, TL, FH, FL) within each block. In the second case, pupil traces time-locked to feedback onset were categorized based on expectancy and outcome type, namely, Expected Feedback, Unexpected Feedback, Expected No Feedback, and Unexpected No Feedback, according to the reinforcement structure specific to each block. Again, we controlled for potential effects of low-level stimulus features by fitting additional LMMs for both behavioral and pupillary data, which incorporated news length (in syllables or characters) as continuous covariates (SI Appendix).

2.2.1. Computational modeling.

To characterize the learning dynamics underlying participants’ decisions during the probabilistic RL task, we implemented two computational models based on the Rescorla–Wagner framework combined with a SoftMax choice rule. Both models were designed to capture how participants integrated reward feedback with prior veracity and confidence beliefs, differing in whether learning was symmetric or asymmetric with respect to the sign of the prediction error (PE).

The models were programmed in MATLAB (v. R2023b) and executed locally. For each participant and block, trial-wise vectors containing individual choices (1 = left stimulus, 2 = right stimulus), headlines comparisons (1 to 6, indicating the specific pair among the six fixed comparisons), and reward delivery (1 = reward received, 0 = no reward) served as model inputs. Because stimulus positions (left/right) were counterbalanced, choices were remapped to a fixed headline category pairs based on the comparison index (1 to 6), where 1 = FH–TH, 2 = FL–TH, 3 = TH–TL, 4 = FH–FL, 5 = FH–TL, and 6 = FL–TL. In all cases, a response coded as “1” corresponded to the first element of the comparison, and a response coded as “2” to the second element, ensuring consistent alignment between behavioral and model data.

Since our task involved immediate, nonsequential decisions between fixed stimulus pairs both models implemented a standard Rescorla–Wagner learning rule without temporal discounting (γ = 0) (35):

$$Q\left(t+1\right)=Q\left(t\right)+\alpha \left[R\left(t\right)-Q\left(t\right)\right],$$ [2]

where Q(t) is the estimated value of the chosen option at time t, α is the learning rate (0 < α ≤ 1), and R(t) is the reward received (1 or 0). Initial Q-values were set to 0.5, corresponding to the a-priori expectation of a 50% chance of receiving a reward (R = 1) vs. no reward (R = 0). Action selection was modeled using a SoftMax choice rule, which transforms Q-values into choice probabilities according to the following function:

$$P\left(a_t=i\right)=\frac{e^{\mathrm{\beta }{Q}_i\left(t\right)}}{{\sum }_j{e}^{\mathrm{\beta }{Q}_j\left(t\right)}},$$ [3]

where β is the inverse temperature parameter controlling choice stochasticity (higher β = more deterministic choices).

2.2.1.1. Feature-based symmetric model.

In this model, each headline was represented by two latent features, i.e., veracity (True, Fake) and confidence (High, Low). Separate Q-values were maintained for each feature [Q(True), Q(Fake), Q(High), Q(Low)], and the total value of a given headline was computed as the normalized (36) additive combination of its feature values:

$${\mathrm{Q}}_{\text{headline}}=\frac{{\mathrm{Q}}_{\text{veracity}}+{\mathrm{Q}}_{\text{confidence}}}{2}.$$ [4]

When a headline was chosen, both relevant feature values were updated according to the Rescorla–Wagner rule. This formulation allowed the model to generalize learning across stimuli sharing the same belief-relevant features (e.g., all “high-confidence” or all “true” headlines). Here, learning was governed by two feature-specific learning rates, ${\alpha }_V$ and ${\alpha }_C$, which scaled the update of veracity- and confidence-related Q-values, respectively. After each choice, the PE was computed:

$${\delta }_t=R\left(t\right)-Q_{\text{headline}}\left(t\right).$$ [5]

and both relevant feature values were updated proportionally to δ, using the same sign for positive and negative PEs:

$$\begin{array}{c}{Q}_v\leftarrow Q_v+{\alpha }_v{\delta }_t\\ Q_c\leftarrow Q_c+{\alpha }_c{\delta }_t\end{array}.$$ [6]

2.2.1.2. Feature-based asymmetric model.

The asymmetric model shared the same feature-based architecture as the symmetric model but introduced distinct learning rates for positive and negative PEs, following the formulation used in Palminteri et al. (13). For each trial, the model selected the appropriate learning rate based on the sign of the PE:

$$\alpha \left(t\right)=\left\{\begin{array}{cc}{\alpha }_{\text{pos}},& \delta \left(t\right)>0\\ {\alpha }_{\text{neg}},& \delta \left(t\right)<0\end{array}\right..$$ [7]

And the update rule became:

$$Q_f\left(t+1\right)=Q_f\left(t\right)+\alpha \left(t\right)\delta \left(t\right),f\in V,C.$$ [8]

This formulation preserves feature-based generalization while allowing learning to be valence-asymmetric, capturing well-documented human tendencies to differentially weight positive and negative outcomes.

2.2.1.3. Parameter estimation and model comparison.

Model parameters (${\alpha }_V$, ${\alpha }_C$, ${\alpha }_{+}$, ${\alpha }_{-}$, β) were estimated for each participant and block by minimizing the negative log-likelihood of observed choices given model-predicted probabilities. Optimization used MATLAB’s fmincon function with 50 random initializations, uniformly sampled within the bounds (0.01, 1) for learning rates and (0.01, 10) for β. Alpha estimated parameters were analyzed, for both models, as a function of dimension/valence and experimental block through a repeated-measures ANOVA with dimension (veracity, confidence) or valence (positive, negative) and Experimental Block (T, F, H, L, R) as the main factors. Beta parameters were analyzed through a repeated-measures ANOVA across and within models as a function of Experimental Block (T, F, H, L, R). Akaike Information Criterion and Bayesian Information Criterion (BIC) were computed per subject, block, and model. Group-level model evidence was evaluated using Bayesian Model Selection (BMS) (random-effects), reporting protected exceedance probabilities. Subject-wise best-model attribution (based on BIC) was also computed, and subgroup-level BMS analyses were performed to examine the stability of individual strategies across blocks (SI Appendix).

2.2.1.4. Simulation analyses, parameter/model recovery.

To evaluate the robustness of the computational framework, we performed three complementary validation analyses including parameter recovery, model recovery, and simulated accuracy by computational agents (SI Appendix).

2.3. Feedback Phase.

In the last experimental phase, we first computed d’ and c indices and tested for their significance using one-sample t tests against zero. These indices were then compared with those from the Mapping phase using paired-sample t tests. Successively, we investigated potential opinion changes by comparing with a paired-sample t test the percentage of confirmations of initial judgments with that of opinion changes. Moreover, to examine the factors influencing both belief updating and task accuracy, we estimated two LMMs separately for each dependent variable (DV; Belief Update and Accuracy), including the interaction of Veracity (T, F) and Confidence [H, L; DV ~ Veracity * Confidence + (1 | Subject)] and the random intercepts for Subject to account for individual-level variability. For each model, marginal and conditional R² values were computed to quantify the variance explained by fixed effects and by the full model (fixed + random effects), respectively. Effect sizes were also quantified using Cohen’s f², and P-values were corrected for multiple comparisons using FDR where appropriate.

Pupillary responses during the Feedback phase were again analyzed during headlines evaluation preceding individual decisions and during feedback presentation. In a first step, we compared pupil dilation in the 2,000 ms preceding the response between trials in which participants confirmed their initial judgment (Opinion Confirmed) and those in which they changed their mind (Opinion Changed). Second, predecisional pupil responses were also examined as a function of the headline category associated with the opinion being expressed during the initial Mapping phase. In a third analysis, we focused on trials in which participants confirmed their initial decision and received feedback. Here, pupil responses were categorized based on feedback valence: Positive Feedback indicated correct confirmations (i.e., alignment between belief and truth), while Negative Feedback indicated incorrect confirmations. Due to the low number of trials in which participants changed their mind, it was not possible to conduct the same feedback-locked analysis for Opinion Changed trials. Also in this case, control LMMs were run on behavioral and pupillary data, considering the number of syllables and the number of characters as continuous covariates (SI Appendix).

3. Results

3.1. Mapping Phase.

3.1.1. SDT indexes.

During the mapping phase, on average, participants were able to distinguish true from false news successfully, as indicated by a discriminability index (d’) significantly different from zero [t₍₃₀₎ = 10.4, P < 0.001, Cohen’s d = 1.87; mean = 0.64; Fig. 2]. Despite a generally conservative response tendency (i.e., a bias toward responding “false”) the decision criterion (c), was not different from that of an “ideal observer” [t₍₃₀₎ = 1.5, P = 0.1, Cohen’s d = 0.28; mean = 0.09], and they tended to bet larger amounts of virtual money when judging news as false [t₍₃₀₎ = −2.7, P = 0.01, Cohen’s d = −0.48; Fig. 2].

Fig. 2.

Signal detection and confidence measures during the Mapping phase. Left: Group means (±95% CI) and individual values of signal detection indices (d′, criterion c). Middle: Receiver operating characteristic (ROC) curve illustrating average sensitivity (d′ = 0.64). Right: Mean accuracy estimation from betting behavior, showing higher confidence (bet amount) for true than for false headlines (P = 0.01).

3.1.2. Pupil size.

At the neurophysiological level, pupil dynamics during the Mapping phase revealed that belief-related processing begins well before an explicit decision is made. In the two seconds preceding the response, pupillary responses varied systematically depending on how participants would later judge and endorse each headline. Specifically, cluster-based permutation analyses uncovered increased dilation (less constriction) for headlines later judged as fake with high confidence (FH) compared to both true/low-confidence (TL) headlines (P = 0.01; –2,000 to –575 ms; Fig. 3) and fake/low-confidence (FL) headlines (P = 0.04; –2,000 to –742 ms; Fig. 3). When collapsing across veracity and focusing on confidence alone, high-confidence trials were consistently associated with reduced pupillary constriction relative to low-confidence trials throughout the entire preresponse window (P = 0.01; Fig. 3). This pattern suggests that self-confidence exerts an early influence on autonomic arousal or attentional engagement, potentially reflecting anticipatory processes tied to internal epistemic evaluation. Notably, no significant effects were found when pupil size was analyzed as a function of objective truthfulness (real vs. fake), indicating that these physiological signals reflect internal judgments rather than stimulus-driven properties.

Fig. 3.

(A) Mean PSC of pupil size during the 2 s preceding participants’ decision about news veracity as a function of news subjective category (True High, TH; True Low, TL; Fake High, FH; Fake Low, FL). (B) Same data collapsed across veracity, showing larger predecisional dilation (less constriction) for high-confidence than low-confidence trials (P = 0.01). Boxes indicate SEM; dots represent individual participants.

3.2. Learning Phase: Two-Armed Bandit Task.

3.2.1. Accuracy.

The LMM used to evaluate whether response accuracy varied across experimental blocks explained a moderate portion of variance, with a marginal R² of 0.13 and an effect size of f² = 0.16. Accuracy significantly differed as a function of the reinforcement condition [Block main effect: (χ²(3) = 24.9, p_FDR < 0.001, Fig. 4A and SI Appendix)]. Relative to T-Block, where reinforcement was associated with news items participants had judged as true, accuracy was significantly lower in F-Block (Est. = –7.87, p_FDR =0.01), H-Block (Est. = –8.93, p_FDR = 0.009), and L-Block (Est. = –14.51, p_FDR < 0.001). Additionally, accuracy was higher in F-Block as compared to L-Block (Est. = 6.64, p_FDR = 0.04), while no differences were found between F-Block and H-Block (Est. = 1.06, p_FDR = 0.71), as well as between H-Block and L-Block (Est. = 5.58, p_FDR = 0.07).

Fig. 4.

RL performance across experimental blocks. (A) Mean accuracy (±95% CI) and individual values across the first four blocks, where specific headline categories were reinforced; Please note that the random-reward block (R-Block) was excluded because no objective accuracy could be defined. (B) Left panels: percentage of headline-category choices across the five blocks. The gray line indicates the random-reward baseline (R-Block) used for comparison. Right panels: headline comparison choice indices, expressing the relative preference between paired categories (positive values = preference for the first term; negative = for the second). Error bars represent SEM; asterisks indicate P < 0.05.

3.2.2. Choices distribution across blocks.

To better understand how participants adapted their choices to different reinforcement contingencies, we reestimated the LMM using R-Block, where reward was assigned randomly, as the baseline. This allowed us to examine how structured reinforcement in T-, F-, H-, and L-Blocks systematically altered headline selection relative to a neutral reference (Fig. 4B and SI Appendix). The analysis revealed a robust Block × Veracity x Confidence [χ²(4) = 28.5, P < 0.001], with moderate explanatory power (marginal R² = 0.24) and a medium effect size (Cohen’s f² = 0.31).

More specifically, in T-Block, where rewards were linked to headlines previously judged as true, choices shifted significantly. High-confidence true headlines (TH) were selected significantly more often (Est. = 12.50, p_FDR < 0.001), while high-confidence false headlines were chosen less frequently (FH: Est. = –13.46, p_FDR < 0.001). In the absence of any incentive, participants preferred false headlines judged with low rather than high confidence (FH vs. FL: –31%, P < 0.001), indicating that confidence alone influenced decisions even when reward was not at stake (Fig. 4B). This pattern reversed in F-Block, which rewarded headlines previously judged as false. Participants now selected false headlines with high confidence more frequently (Est. = 13.17, p_FDR < 0.001), while choices of true headlines declined (Est. = –10.04, p_FDR = 0.004). Notably, participants only preferred FL when contrasted to TL, i.e., when both items shared similarly low confidence (TL vs. FL: –12.1%, P = 0.01; Fig. 4B). Given equal reward probability (FH vs. FL), participants reliably chose the more confident option (18.7%, P = 0.01; Fig. 4B). In H-Block, where rewards were tied to high confidence regardless veracity, no overall increase in selection rates of TH (Est. = 4.09, p_FDR = 0.27) and FH (Est. = 5.73, p_FDR = 0.10) was observed. At the same time, a marginal suppression of low-confidence headlines did not survive FDR correction (TL: Est. = –4.54, p_FDR = 0.29; FL: Est. = –5.28, p_FDR = 0.33). Within veracity-matched comparisons, participants reliably favored high-confidence items (TH vs. TL = 23.2%, P = 0.001; FH vs. FL = 15.2%, P = 0.04; Fig. 4B). Yet, when veracity and confidence varied simultaneously (e.g., TH vs. FL or TL vs. FH), this preference disappeared. Likewise, in comparisons without differential incentives (e.g., TH vs. FH; TL vs. FL), no preference for truth over falsity emerged. These mixed results suggest that the reward–confidence association was only partially internalized. L-Block reversed the contingency, rewarding low-confidence headlines. Yet participants did not significantly adapt their choices. Selection rates for all headline types remained statistically indistinguishable from the random baseline of R-Block (e.g., TH: Est. = 1.71, p_FDR = 0.60; FH: Est. = –3.05, p_FDR = 0.34; FL: Est. = –1.93, p_FDR = 0.73; TL: Est. = 3.27, p_FDR = 0.31). When comparing specific headline pairs, a single significant effect emerged: Participants preferred TL over FH (preference index: 19.6%, P = 0.007). Overall, however, the failure to adjust behavior suggests that the inverse reward–confidence structure was not effectively learned. Finally, in R-Block, where no contingency-guided behavior, choices reflected participants’ default epistemic tendencies. No overall preference was observed for any specific headline category. Yet, when contrasting extreme conditions, TH vs. FL and TL vs. FH, participants showed a consistent preference for headlines they had previously judged as true (preference indices: 9.8% and 13.8%, both P = 0.01), even in the absence of reward.

3.2.3. Computational modeling.

Consistent with behavioral performance, parameter estimates from the symmetric feature-based model revealed higher learning rates for the veracity than for the confidence dimension [0.33 vs. 0.24; F_(1,27) = 4.2, P = 0.04, η_p² = 0.13; Fig. 5A], although no reliable modulation across blocks was observed. Conversely, for the asymmetric model, a significant Valence × Block interaction [F_(4,108) = 2.6, P = 0.04, η_p² = 0.09] indicated that positive PEs were weighted more strongly than negative ones, particularly in the H- (0.42 vs. 0.17, P = 0.01) and L-Blocks (0.48 vs. 0.24, P = 0.005). The asymmetric model was also characterized by more deterministic choices [F_(1,27) = 7.52, P = 0.01, η_p² = 0.22], with β values particularly higher in the T-Block relative to the other conditions [F_(4,108) = 2.60, P = 0.04, η_p² = 0.08; Fig. 5A].

Fig. 5.

(A) Means, 95% CI, and individual values distribution of estimated Alpha (learning rates) and Beta (inverse temperature) parameters by block and model. (B) Means, 95% CI, and individual values of accuracy at the bandit task for experimental human participants and computational agents generated by symmetric and asymmetric models.

BMS revealed a clear block-dependent pattern (SI Appendix, Fig. S2). In blocks where reward contingencies aligned with headline veracity (T- and F-Blocks), the symmetric model was favored (expected posterior probability: 0.73 and 0.81, respectively; protected exceedance probability, pxp: 0.87 and 0.90). In contrast, when confidence rather than veracity determined reward outcomes (H-Block), or when feedback was driven by low-confidence or random contingencies (L- and R-Blocks), the asymmetric model provided a superior account of the data (expected posterior probability: H = 0.83; L = 0.48; R = 0.53; pxp: H = 0.98; R = 0.52). At the individual level, participants displayed heterogeneous but internally coherent computational profiles. 16 participants were best fit overall by the symmetric model, and 12 by the asymmetric one. Subgroup-specific BMS confirmed these patterns: Symmetric-dominant participants strongly favored the symmetric model in the T- and F-Blocks (pxp = 0.99 to 0.99) but switched to the asymmetric model in the H-Block (pxp = 0.96). Conversely, asymmetric-dominant participants showed a pronounced preference for the asymmetric model in H-, L-, and R-Blocks (pxp = 0.99, 0.75, 0.97, respectively), while favoring the symmetric model only in the F-Block (pxp = 0.98).

Finally, simulation analyses revealed that both symmetrical (60.3%, P = 0.01) and asymmetrical (60.6%, P = 0.005) models achieved higher overall accuracies than human participants [58.1%; F_(2,54) = 4.9, P = 0.01, η_p² = 0.15], even though no interaction with block factor emerged, indicating that the two models faithfully reproduced the block-wise structure of learning and did not differ from human accuracy across all the experimental block (Fig. 5B). Trial-by-trial learning curves confirmed that both models closely tracked the progression of human performance across blocks (SI Appendix, Fig. S3).

3.2.4. Pupil size.

Predecisional pupillary responses during the Learning phase revealed significant differences in T and F-Blocks (Fig. 6A). In the T-Block, cluster-based permutation tests identified stronger dilation in trials involving strongly endorsed true headlines (TH), with significant contrasts between TH vs. FH and TH vs. FL (P = 0.03; –1,222 to 0 ms), and between TH vs. FL and FH vs. FL (P = 0.04; –987 to 0 ms). These results suggest that belief-congruent choices involving high-confidence truths elicited distinct anticipatory arousal compared to epistemically incongruent or uncertain items. In the F-Block, where false headlines were rewarded, pupil dilation was greater when evaluating TH vs. FH pairs than TL vs. FH (P = 0.03; –1,096 to 0 ms), indicating higher cognitive load when a strongly held belief conflicted with feedback. Comparisons involving low-confidence items (e.g., TL vs. FH) were associated with more constricted pupils, consistent with reduced decisional conflict. No significant effects were found in the remaining Blocks or in postdecision, feedback-related, analyses (Fig. 6B). Overall, these findings indicate that belief strength modulates anticipatory arousal during learning, with high-confidence priors amplifying cognitive effort in the face of competing reward signals.

Fig. 6.

(A) Mean PSC of pupil size during the 2 s preceding individual choices, shown for all six possible headline comparisons in T-Blocks and F-Block. Bars indicate group means ± SEM; dots represent individual data. (B) Mean PSC time courses following feedback presentation for T-, F-, H-, and L-Blocks, separately for expected vs. unexpected feedback (ExpFB/UnexpFB) and for trials with vs. without feedback (FB/NoFB). Shaded areas represent SEM.

3.3. Feedback Phase.

During the Feedback phase, participants showed a significant tendency to confirm rather than disconfirm the judgments expressed during the initial Mapping phase [t₍₂₇₎ = 19.2, P < 0.001, Cohen’s d = 3.62; Fig. 7A]. This attitude led to a marginally significant shift in response criterion during the Feedback phase [t₍₂₇₎ = 1.95, P = 0.06, Cohen’s d = 0.36; Fig. 7B] leading participants toward a more liberal response bias (c = −0.03). At the same time, this change in response criterion was accompanied by a significant increase in discriminability (d′), which rose from 0.64 to 0.75 [t₍₂₇₎ = −2.13, P = 0.04, Cohen’s d = 0.40; Fig. 7B].

Fig. 7.

(A) Percentage of trials in which participants confirmed or changed their initial judgments, shown overall and by headline category (TH, TL, FH, FL). Bars represent group means ± 95% CI; dots indicate individual data. (B) Signal detection indices (d′, c) during Mapping and Feedback phases, and corresponding ROC curves illustrating higher sensitivity (d′) during Feedback (P = 0.04).

3.3.1. Belief update.

Participants revised their beliefs primarily as a function of confidence level, χ²(1) = 8.71, P = 0.003. As expected, belief updating was significantly higher for low- than for high-confidence news (Est. = 6.77, p_FDR = 0.01). No main effect of veracity emerged, χ²(1) = 0.01, P = 0.89, but a significant Veracity × Confidence interaction, χ²(1) = 4.35, P = 0.03, indicated that under low-confidence conditions, participants were more likely to revise their beliefs for news initially judged as false than for those judged as true (Est. = –7.06, p_FDR = 0.005). The model accounted for a moderate proportion of variance (marginal R² = 0.17) and yielded a moderate effect size (Cohen’s f² = 0.20).

3.3.2. Accuracy.

As observed for belief updating, accuracy varied primarily as a function of confidence level, χ²(1) = 10.24, P = 0.001. Specifically, accuracy was significantly lower for low- compared with high-confidence news (Est. = –7.82, p_FDR = 0.006). No significant main effects of Veracity, χ²(1) = 0.03, P = 0.86, or Veracity × Confidence interaction, χ²(1) = 0.24, P = 0.62, were found. The model accounted for a moderate proportion of variance (marginal R² = 0.14) and yielded a moderate effect size (Cohen’s f² = 0.17).

3.3.3. Pupil size.

Analysis of predecisional pupil size revealed no significant difference between trials where participants confirmed their initial judgment and those where they changed their mind. However, when pupil dynamics were examined by headline category, Fake High (FH) items elicited significantly greater pupil size than both True High (TH) (P = 0.01; –1,450 to 0 ms) and True Low (TL) (P = 0.004; –2,000 to 0 ms; Fig. 8A). This suggests that reaffirming a confidently held but false belief is associated with increased anticipatory arousal, possibly reflecting internal conflict or epistemic instability.

Fig. 8.

(A) Mean PSC in pupil size during the 2 s preceding confirm/disconfirm decisions, shown for each news category (TH, TL, FH, FL). Boxes represent group means ± SEM; dots indicate individual participants. (B) Mean PSC following feedback presentation, shown separately for positive (green) and negative (red) outcomes; shaded areas represent ± SEM. The black bar along the x-axis indicates the time window of the significant cluster (P = 0.001).

Postdecision, a significant effect of feedback valence emerged for confirmed judgments: Pupil dilation was greater following negative than positive feedback to confirmations (P = 0.001; 804 to 2,000 ms; Fig. 8B), indicating surprise or expectancy violation. This aligns with the idea that CB reduces sensitivity to disconfirming evidence, making unexpected errors, especially those contradicting recently reaffirmed beliefs, particularly cognitively demanding.

4. Discussion

The present study investigated how the accessibility of decision attributes related to news headlines, specifically veracity and confidence, interacts with RL dynamics within a learning paradigm. By combining pupillometry, a personalized stimulus set, and computational modeling, we provide a multifaceted account of the neurophysiological, cognitive, and behavioral mechanisms underpinning belief-guided reinforcement strategies. Findings across the three task phases converge in describing participants’ epistemic sensitivity, their reinforcement-based learning strategies, and their eventual resistance or openness to revising their beliefs.

4.1. Mapping Phase: Individual Characteristics and Fake News Discriminability.

Pupillometric data during the Mapping phase revealed early signs of autonomic engagement during belief evaluation. More in detail, trials involving high-confidence judgments, particularly for headlines judged as false, were associated with increased predecisional pupil dilation.

This pattern is consistent with evidence that pupil responses scale with the strength of internal decision signals. Prior studies have shown that pupil dilation can track not only arousal but also subjective confidence and metacognitive monitoring (25, 27, 28). In our task, greater dilation prior to confidently judging a headline may therefore reflect a combination of heightened arousal, increased cognitive effort, and the recruitment of confidence-related monitoring mechanisms when applying strict internal criteria to evaluate potentially misleading content. In other words, predecisional pupil dilation in high-confidence trials may reflect the subjective salience or motivational weight of the belief itself. Importantly, no significant differences in pupil dilation emerged when comparing objectively true vs. false headlines, suggesting that pupillary responses are shaped by internal belief states rather than external stimulus properties. This finding supports the view that the autonomic nervous system, and the pupil specifically, can encode epistemic signals rooted in subjective certainty (26, 37).

From a behavioral standpoint, participants demonstrated above-chance accuracy in distinguishing real from fake headlines, as reflected by a significantly positive d′ and a conservative response criterion (c > 0). This cautious stance was also evident in wagering behavior: Participants reported higher confidence when rejecting a headline as false, consistent with a “better safe than sorry” heuristic in uncertain contexts (38). Recent meta-analytic evidence (39) further corroborates this pattern, showing that analytical thinking and age are associated with a mild bias toward rejecting potentially false content, suggesting that skepticism, rather than credulity, can serve as a default epistemic posture when incentives are absent or ambiguity is high.

4.2. Learning Phase: RL Dynamics and Choice Behavior under News Consumption.

During the Learning phase, behavior was strongly modulated by the match between reinforcement contingencies and participants’ preexisting beliefs. When the reward structure aligned with the veracity categories established during the mapping stage, as in T- and F-Blocks, choices reliably shifted toward headlines previously judged as true or false, particularly when endorsed with high confidence. This pattern suggests reliance on a belief-congruent reinforcement strategy [i.e., preferential selection of belief-consistent options under reward uncertainty; (13, 39, 40)], whereby internal epistemic priors guide action selection when the environment remains structurally coherent with them (41). By contrast, when contingencies required prioritizing confidence over veracity, lower accuracies were observed. In H-Block, participants internalized the high-confidence rule only when it aligned with veracity judgments, whereas in L-Block adaptation was largely absent, with choice distributions indistinguishable from the random baseline. These failures to invert well-established belief structures are consistent with findings that behavior defaults to model-free, habitual strategies under counterintuitive or weakly structured contingencies (42).

Computational modeling further clarifies learning mechanisms underlying individual performances. The symmetric feature-based model exhibited higher learning rates for the veracity dimension than for the confidence dimension, indicating a reliance on belief-based generalization when reward contingencies aligned with participants’ epistemic priors. This is consistent with evidence that humans efficiently generalize over abstract features that are perceived as reliable predictors of value (43, 44) and with theories of credition framing beliefs as higher-level priors that constrain information integration. In contrast, the asymmetric model was characterized by an overweighting of positive relative to negative PEs, and provided a superior account of behavior, especially in H-and L-blocks. This valence asymmetry mirrors classic findings in affective RL (45) and suggests a shift toward a more model-free, surprise-driven updating regime when veracity-based abstraction can no longer support prediction. The concurrent increase in inverse temperature in these blocks further indicates that participants stabilized their policy through more deterministic choices under conditions of epistemic conflict or low environmental reliability, consistent with prior work on choice determinism in volatile settings (46). BMS and individual-level analyses confirmed this dual-process structure, revealing consistent computational profiles across participants. Simulated agents reproduced human performance and trial-by-trial dynamics, reinforcing the view that behavior emerges from the interaction between stable belief-based priors and flexible prediction-error updating. Taken together, these findings indicate that no single learning mechanism uniformly explains behavior across all conditions. In our task, belief-guided decision-making seems to emerge from an adaptive arbitration between epistemic structure and reinforcement signals: Individuals rely on abstract belief categories when they remain predictive, but revert to valence-sensitive updating when environmental feedback requires abandoning their prior epistemic organization.

Pupillometric data nicely complemented behavioral and computational results. During the learning phase, pupil dilation increased in trials involving strong motivational conflict, particularly TH vs. FH comparisons in T- and F-Blocks, where participants’ prior veracity beliefs clashed with the reward structure. These anticipatory responses are consistent with evidence showing that pupil size reflects value conflict, decision uncertainty, and PE (25, 26), and likely reflect the engagement of the LC-NE system in situations requiring adaptive reevaluation of belief strength vs. external incentives (47). Notably, this arousal occurred before feedback was delivered, suggesting that participants had internally anticipated the difficulty of reconciling their epistemic convictions with reward-based contingencies.

4.3. Feedback Phase: Conditions and Limits of Belief Revision.

Importantly, in our experimental design, the preceding learning phase did not assess explicit epistemic updating in response to corrective evidence. Accordingly, changes observed in the Feedback phase should not be expected to depend on prior learning dynamics; however, they may reflect additional processes of belief evaluation and decision reassessment. Consistent with this interpretation, participants displayed a clear pattern of belief-congruent persistence, often reaffirming their initial judgments, particularly when those judgments had been made with high confidence. This effect was comparable for confidently held true and false headlines, indicating that high-confidence beliefs, regardless of their subjective veracity, were equally resistant to change, consistent with the notion that confidence itself, rather than veracity, stabilizes belief representations (48). Importantly, although this pattern of persistence does not in itself demonstrate CB in the classical sense of asymmetric evidence integration, such confidence-weighted resistance constitutes a key mechanism through which confirmation-like effects may emerge.

Despite this overall stability, some evidence of belief updating emerged: Participants shifted toward a slightly more liberal decision criterion and showed improved discriminability (d′), suggesting partial recalibration of epistemic judgment under feedback. As one might expect, belief revision was more likely when initial confidence was low, especially for headlines initially judged as false, supporting the idea that uncertainty facilitates openness to new evidence (28, 46). Alternatively, this improvement may be partly driven by increased motivation or attentional engagement during the final phase, where outcomes of the wagers were displayed. Indeed, during the Feedback phase, participants received a final summary of their accumulated gains, which may have encouraged more deliberate scrutiny of the headlines, especially given the simultaneous display of their initial responses and confidence wagers. These two explanations are not mutually exclusive: Increased effort may have amplified opportunities for metacognitive reassessment, particularly for low-confidence items. Taken together, these findings support dynamic models in which belief persistence and revision emerge from the interplay of prior expectations, reinforcement history, and confidence-based inertia (13).

Pupillometric data offered converging physiological evidence. During predecisional intervals, trials with confidently held but false beliefs (Fake High) elicited larger pupil sizes than true or low-confidence items, suggesting increased cognitive and affective load. Postdecision, disconfirming feedback following belief confirmation evoked marked pupil dilation, consistent with expectancy violation and surprise-driven arousal (49). Our findings are consistent with those of Lefebvre et al. (50), who showed that confirmatory feedback decreases visual attention to unchosen outcomes, whereas disconfirmatory feedback produces enhanced pupil dilation, which reflects cognitive surprise and adaptive reevaluation. Together, these results suggest a common mechanism by which arousal dynamics shape the integration of feedback valence and belief coherence during learning. These effects highlight the role of the pupil as a marker of epistemic conflict, integrating arousal signals into computational accounts of belief-congruent reinforcement and epistemic rigidity (26). Overall, pupillary responses appear to index both anticipatory cognitive load and reactive arousal following epistemic surprise, embedding physiological evidence within models of metacognitive monitoring and belief stability.

5. General Conclusions, Limitations, and Future Perspectives

Our findings underscore the robust influence of prior beliefs on how individuals learn from feedback. Learning in belief-relevant contexts is not driven solely by reward maximization but by an interaction between representational strategies and the structure of prior beliefs. Notably, veracity and confidence played distinct, phase-dependent roles. Confidence was crucial during the initial evaluation and later governed the stability of belief. In contrast, during the learning process, it was veracity rather than confidence that structured reinforcement-based generalization. Confidence-contingent rewards failed to increase confidence-based learning, instead pushing the system toward valence-driven updating. This dissociation suggests that confidence primarily regulates the formation and persistence of beliefs, whereas veracity shapes how those beliefs are used during value-guided learning, revealing a coherent division of labor across the epistemic cycle.

Across all task phases, pupil dilation served as a sensitive physiological marker of epistemic salience and anticipatory arousal. During initial judgments, greater pupil size in high-confidence trials suggested enhanced metacognitive engagement. In the learning phase, pupil size reflected motivational conflict and reevaluation when prior beliefs clashed with reward contingencies. During feedback, disconfirming outcomes elicited robust postdecisional dilation, consistent with surprise and expectancy violation. These patterns align with models linking LC-NE–mediated arousal to belief updating, stimulus salience, and value-driven conflict (25).

Critically, our paradigm models belief dynamics across their full lifecycle, from initial classification to feedback-driven learning and eventual revision, preserving the semantic ambiguity typical of real-world misinformation. While this design offers high ecological validity, it also raises limitations. The within-subject manipulation of reward structures may have induced carryover effects; future studies could employ between-subjects designs to isolate contingency-specific influences. Moreover, the exclusive use of news stimuli limits generalizability; adding nonepistemic or neutral content could test domain specificity. A further conceptual limitation concerns the structure of PEs. In our task, “prior-incongruent” outcomes do not correspond to genuine violations of epistemic priors but only of block rules, making it impossible to meaningfully decompose learning rates by prior congruency. Future paradigms could elicit personal prior beliefs independently of ground-truth veracity and explicitly violate them through reinforcement, enabling a full dissociation of prior-congruent vs. prior-incongruent learning signals.

In summary, while the current paradigm does not involve direct exposure to disconfirming evidence, it does reveal the learning and metacognitive dynamics that promote epistemic stability and belief-congruent reinforcement. These processes represent the cognitive architecture upon which CB and resistance to correction can build. Understanding these dynamics is essential for developing interventions that foster cognitive flexibility, epistemic humility, and resistance to disinformation in complex informational environments.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Anonymized Raw Behavioural and Pupillary data have been deposited in OSF (DOI: 10.17605/OSF.IO/QHBSM) (51).