Background: Language is a complex aspect of cognition closely related to memory and perception. Current psychosis research using Bayesian frameworks demonstrates an imbalanced weighting of prior information over incoming sensory evidence in hallucinations and delusions. Further investigation of this dynamic is warranted in relation to language capabilities, especially regarding agent certainty under ambiguous perceptual tasks. Generative models using frameworks such as Active Inference may be well-suited to describe and classify agent performance in tasks relevant to cognitive processes such as language. This may be an especially useful model to predict performance differences between low and high schizotypy groups, maintaining an expectation of prior overweighting in the high-schizotypy group during high-entropy or ambiguous stimuli.

Method: We constructed and fitted an Active Inference model of individual agent performance in a predictive language task to investigate variational free energy in relation to self-reported and objective semantic prior clarity. The language tasks requires participants to observe sentences with varying levels of entropy and levels of degradation of the sentence final word. Individual perceptions of words are recorded alongside prior clarity and prediction confidence; individuals are assessed on levels of psychotic-like experiences (PLE).

Results: Preliminary results show that individuals with higher levels of PLEs reveal less optimal behavior and higher free energy.

Conclusion: Bayesian frameworks are well suited to disentangle precision of priors and sensory evidence in perceptual inference tasks, making Active Inference a well-suited model to describe agent performance and perceptual ambiguity within dynamic environments.