Background

The strong persistence of delusional beliefs remains insufficiently understood, although current research points to distorted learning mechanisms. Recent findings describe Theory Protection as a selective learning process in which beliefs are maintained by attributing contradicting evidence to ambiguous cues. This selective learning pattern resembles the persistence of systematized delusions. Bayesian models describe how prior beliefs and new information jointly shape learning, while active inference approaches emphasize attentional dynamics in information sampling. These perspectives suggest that persistent beliefs arise from the interplay of prior knowledge, the precision of new information, and attention.

Methods

Across two studies, we investigate the clinical relevance of selective learning in ambiguous contexts and examine its association with risk factors for psychosis. We expect the effect to be associated with subclinical delusional ideation and cognitive vulnerability factors (e.g., stress, working memory, cognitive flexibility). We use a learning paradigm that generates causally ambiguous situations in which selective updating becomes observable and whose variability can be related to these factors. Attentional allocation is assessed using eye-tracking. Trial-by-trial belief updating is modeled using a Bayesian inference model quantifying belief and evidence precision.

Results

Data from two preregistered studies will be presented. Participants engage in selective updating. Analyses will examine the role of prior certainty, clinically relevant predictors, and attentional processes, alongside computational modelling results.

Discussion

The paradigm provides an innovative approach to identifying selective learning processes underlying belief persistence. The findings contribute to both basic models of learning and clinically relevant accounts of maladaptive belief persistence, particularly in psychosis.

Does expecting an aversive event cause simple benign percepts to be experienced as aversive? If so, what neural processes underlie this expectation-induced bias, and does it influence subsequent behavior? To address these questions, we investigated how cues conditioned to evoke aversive expectations shape the processing of simple benign auditory percepts across two preregistered studies (total N = 149). We found that when participants were cued to expect an upcoming noise burst to be of aversively high vs. benign intensity, the same benign intensity noise burst was experienced as more aversive and elicited stronger cardiac acceleration. At the neural level, this was accompanied by potentiated early sensory processing, such that benign percepts were neurally encoded like physically more intense stimuli. Furthermore, as a behavioral consequence of this effect, participants more frequently chose to avoid these benign events despite a monetary cost. These findings characterize a mechanistic pathway propagating from early sensory processing through affective experience to subsequent behavior. This pathway is of clinical relevance, as it demonstrates how aversive expectations may be perceptually confirmed and stabilized during objectively benign conditions, thereby contributing to the persistence of symptoms across several mental disorders.

Beliefs about causal relationships between events guide everyday behavior. Updating these beliefs when confronted with contradictory evidence is critical for adaptive learning. Associative learning theories propose that belief change is driven by prediction error—the discrepancy between expected and observed outcomes—such that violated predictions directly produce updating. However, a more recent account suggests that people often resist revising strongly held beliefs, instead attributing unexpected outcomes to alternative, uncertain causes. Across two human predictive learning experiments, participants learned about the effects of foods (cues) on inflammation (outcome ranging from −10 to +10) in a hypothetical patient, allowing to test which of two foods with different learning histories showed greater updating when presented concurrently. In Experiment 1, cue B prevented increases in inflammation (X+10 XB0), whereas cue A reliably predicted increases (AW+10 EW0), followed by compound training with AB−10. In Experiment 2, cue B was trained identically, whereas cue A trained to have a stronger, more certain association with its outcome (AW+10 W−10) prior to the compound training (AB−10). Results showed that when causal certainty for cues A and B was equal prior to compound training, greater updating occurred for cue A—the cue associated with larger individual prediction error (Experiment 1). In contrast, when participants were more certain about cue A than cue B, greater updating occurred for cue B, indicating preferential learning about the less certain relationship (Experiment 2). We discuss theoretical implications of our findings for the roles of prediction error and certainty in associative learning.

Expectations bias the perception of internal body signals: the same somatic sensation may be experienced as benign or aversive depending on prior expectations. The present within-subject study examined how expectations acquired via observational learning influence the perception of mild dyspnea. Participants (N = 41) first observed videos of a demonstrator experiencing different intensities of dyspnea paired with three different geometric symbols (CS_medium: showing medium intensity of labored breathing; CS_mild: showing mild intensity of labored breathing; CS-: regular breathing). In the subsequent test phase, participants were directly exposed to the same visual symbols. Critically, the CS_medium and CS_mild were followed by the identical mid-intensity inspiratory resistive load (IRL) inducing dyspnea. Before each trial, participants rated the expected aversiveness of dyspnea; after each trial, participants rated the experienced aversiveness of dyspnea. Participants reported greater expected and experienced aversiveness of the dyspnea following the CS_medium compared to the CS_mild. With repeated exposure to the IRL, the subjective experience changed until experience ratings for the CS_medium and CS_mild converged. A comparable convergence was observed for the expectancy ratings as well. This pattern was supported by psychophysiological data: skin conductance responses were greater for the CS_medium than the CS_mild in the first half of the test phase, but this difference disappeared in the second half. Together, these findings demonstrate the strong influence of learned expectations on interoceptive perception and its updating through direct experience. The results have clinical implications for disorders characterized by altered symptom perception, such as panic disorder, asthma, and COPD.

Various forms of psychopathology are characterized by exaggerated threat expectations that defy disconfirming evidence. This has been studied in fear extinction experiments, in which participants initially develop a defensive response – including threat expectations – to a visual cue that repeatedly precedes an aversive outcome (usually a mild electric shock). Then, the extent to which disconfirming evidence results in a decrease in threat expectations is investigated by repeatedly presenting the cue in absence of the shock. Although research on fear extinction has advanced our understanding of why threat expectations persist, it does not take into account that disconfirming events are often ambiguous and leave room for interpretation. This is particularly relevant to paranoia and social anxiety, in which the threat is expected to come from other people. To address this limitation, we developed a novel learning task that used ambiguous disconfirming evidence, i.e. neutral facial expressions. In this task, participants learned to expect threatening faces in the presence of a visual cue that was repeatedly followed by images of angry faces. The cue was then followed by neutral faces during extinction. Finally, participants indicated the extent to which they perceived hostility in the faces shown during the learning task. Consistent with our hypothesis, those who perceived neutral faces as hostile maintained elevated threat expectations during extinction. This shows that misattribution of hostility to neutral faces, which has been observed in people with paranoia or social anxiety, contributes to the persistence of social threat expectations through associative processes.