The research, led by Elisabeth F. Sterner and Franziska Knolle at the Technical University of Munich, set out to test one of the most talked-about frameworks in modern computational psychiatry: the idea that autism and schizophrenia sit at opposite ends of a predictive processing continuum. According to this account, autistic perception is dominated by incoming sensory evidence — the world feels ‘too real,’ contributing to sensory overload and literal language comprehension — while schizophrenia-spectrum perception is dominated by overly strong prior beliefs, which may underpin delusions that resist contradictory evidence and hallucinations in which priors override sensory input. If true, the two conditions should show diametrically opposed computational alterations when processing language, a domain impaired in both.

To probe this, the team recruited 55 right-handed native German speakers aged 18 to 35 with normal hearing. Participants completed the Schizotypal Personality Questionnaire and the Autism Spectrum Quotient, which measure subclinical schizotypal and autistic traits in the general population, along with subtests of the Wechsler Adult Intelligence Scale IV to derive a verbal comprehension index. Crucially, autistic and schizotypal traits were significantly correlated in this sample, echoing large-scale psychometric findings that the two trait dimensions overlap substantially rather than diverge.

The experimental task was an auditory sentence comprehension paradigm designed to independently manipulate the two ingredients of Bayesian inference. The strength of high-level semantic prior beliefs was varied by presenting sentence beginnings with three levels of Shannon entropy — high, medium, and low predictability — normed in an independent sample. For highly predictable sentences, half were completed with the most expected word, while the other half ended with a syntactically plausible word with a cloze probability of zero, creating a high mismatch condition. The precision of low-level sensory evidence was manipulated by noise-vocoding the sentence-final words at four clarity levels, from one channel (unintelligible) to twelve channels (highly intelligible). After each trial, participants rated the clarity of the final word, said the word they understood, and rated their confidence.

Behavioral results replicated the task manipulations: clarity and confidence ratings decreased with greater degradation and lower predictability. Notably, words following highly predictable contexts were rated as clearer regardless of degradation, indicating that subjective clarity reflects not only sensory precision but also the precision of the semantic prior. Task-based hallucinations — saying a word that was never presented — increased dramatically as sentence predictability fell, demonstrating that rising contextual uncertainty drives misperceptions of the kind theorized to contribute to clinical hallucinations.

The team then applied a Bayesian belief updating model to estimate, trial by trial, the relative weight of semantic prior beliefs versus sensory evidence in each participant’s responses. The model formalizes comprehension as a precision-weighted update of expected correctness probabilities, with a parameter ν capturing the implied number of previous observations — the weight of the prior relative to the incoming evidence. The results were unambiguous for schizotypy: increasing SPQ scores induced a significant shift toward the prior belief, with a 10-point increase in SPQ producing a 3.8% increase in prior weighting. Autistic traits and verbal comprehension did not significantly alter the prior weight, and post-hoc analyses confirmed the SPQ effect was significantly stronger than the AQ effect.

Electrophysiological evidence came from EEG recordings focused on the N400, an event-related potential measured over 15 centro-parietal electrodes between 300 and 500 milliseconds after word onset. The N400 is operationalized within predictive processing as a lexico-semantic prediction error signal: less predictable words elicit more negative amplitudes than highly predictable ones. Linear mixed-effect models with participant and electrode position as random intercepts revealed a graded predictability effect, with the most negative amplitudes in the high mismatch condition, followed by low, medium, and high predictability.

The trait interactions told a more surprising story than the continuum hypothesis predicted. Increasing schizotypy was associated with a reduced N400 difference between highly predictable sentences and both low-predictability and high-mismatch sentences — an attenuated semantic prediction error signal, consistent with overweighted priors. Follow-up simple slope analyses showed the reduced effect was driven specifically by impaired processing of predictable input: N400 amplitudes became more negative with increasing schizotypy in the high and medium predictability conditions, suggesting that individuals with higher schizotypal traits assign greater precision to semantic priors that are less accurate, so that even expected words deviate from their overly precise but faulty predictions. Increasing autistic traits showed a similar but weaker pattern, with a reduced N400 difference only for the high mismatch condition, and post-hoc power analyses estimated 93.7% power to detect the schizotypy interaction versus 70.4% for the autism interaction.

Verbal comprehension ability showed the opposite electrophysiological tendency: higher verbal skills were associated with an enhanced N400 difference between predictable and unpredictable endings following highly predictable contexts, indicating a larger prediction error response and greater sensitivity to unexpected input. Importantly, schizotypy showed no association with verbal comprehension, ruling out the possibility that prior overweighting simply reflects poorer language ability.

Taken together, the findings do not support diametrically opposed predictive processing profiles in subclinical expressions of autism and schizophrenia. Instead, both trait dimensions were associated with attenuated N400 prediction error responses, suggesting a shared tendency to overweight semantic prior beliefs and comparable difficulties flexibly using contextual information to guide comprehension. The authors propose that psychotic symptoms such as hallucinations and delusions may more clearly differentiate the two conditions, and that multiple predictive processing alterations can coexist within each spectrum, mapping onto different symptoms and distinct neurobiological substrates. Future work with larger samples and clinical cohorts, combining single-trial ERP analyses with computational parameters, will be needed to disentangle disorder-specific mechanisms — but for now, the brain’s broken predictions appear to be a flaw that autism and schizophrenia share rather than a mirror that divides them.

Subject of Research: Neuroscience

Article Title: Altered semantic prediction error processing with increasing schizotypal and autistic traits

Article References: Sterner, E. F., Demler, V. F., Lenz, R., MacGregor, L. J., Mathys, C., & Knolle, F. (2026). Altered semantic prediction error processing with increasing schizotypal and autistic traits. Translational Psychiatry, 16(1), Article 502. https://doi.org/10.1038/s41398-026-04479-4

Image Credits: AI Generated

DOI: 10.1038/s41398-026-04479-4

Keywords: predictive processing, N400, EEG, schizotypy, autism, language comprehension, Bayesian inference, psychiatry, Altered, semantic, prediction, error