When Expectation Drowns the Signal: Overweighted Priors as a Mechanism of the Psychotic Spectrum
- Across two studies (N = 109 and N = 55), people higher in schizotypal traits leaned more heavily on high-level semantic expectations and less on the actual acoustic input when decoding noise-degraded speech.
- This overweighting of priors predicted task-based hallucinations – hearing words that matched the expectation but were never present in the signal – linking a computational parameter directly to a perceptual symptom.
- The behavioural pattern from Study 1 was replicated in an independent Study 2 sample, and the same imbalance correlated with higher glutamate levels in the cingulate cortex measured by proton magnetic resonance spectroscopy.
- The effect held along a continuous subclinical dimension of psychosis proneness rather than appearing only in a diagnosed group, supporting a graded mechanistic account of how perception drifts toward prediction.
The dominant computational story about psychosis is not really about psychosis as a category – it is about a dial. Perception, in predictive-coding accounts, is a negotiation between what the brain expects (priors) and what the senses report (sensory evidence). Turn the dial too far toward expectation and the world starts to confirm the prediction regardless of the input; turn it too far the other way and every stray noise demands explanation. Franziska Knolle and colleagues at the Technical University of Munich, working with collaborators in Cambridge and Aarhus, set out to measure where that dial sits in people whose personalities sit closer to the psychosis end of the spectrum.
Their instrument is elegant in its ordinariness: a language-comprehension task. Participants heard spoken sentences degraded by noise. The researchers independently varied how predictable the final word was, how badly the sound was degraded, and how surprising the eventual word turned out to be. A Bayesian belief-updating model then estimated, for each person, how much weight they gave the high-level semantic prior versus the muddy acoustic evidence. This is the methodological move worth noticing. Rather than asking whether someone has aberrant inference, the design yields a number – the prior weight – that can be correlated with traits, symptoms, and brain chemistry.
The result is a clean mechanistic chain. Higher schizotypy went with heavier reliance on priors. Heavier reliance on priors went with more task-based hallucinations: a participant confidently reporting a word that fit the sentence but was acoustically absent. In other words, the same overweighting that helps a healthy listener reconstruct speech in a noisy café becomes, when pushed, a manufacturing process for percepts that have no external referent. The model did not merely describe the symptom; it generated it under controlled conditions.
What lifts this from a behavioural curiosity to a translational finding is Study 2. The same overweighting of priors tracked elevated cingulate glutamate. Glutamate is the principal excitatory neurotransmitter and a leading suspect in models of prediction-error signalling; an association between a computational parameter and a neurochemical measure begins to bridge the gap between an abstract dial and a biological substrate. It is correlational, and the sample is modest, but it is the kind of convergence the field has been demanding – a mechanism specified at the algorithmic level and anchored, however tentatively, in neurochemistry.
For the practising clinician the conceptual payoff is the reframe. A patient who "hears things" is not simply malfunctioning; the same inferential machinery that makes all of us efficient perceivers has its gain set too high. That is a more humane and more actionable story than a deficit narrative, and it points toward interventions – cognitive, pharmacological, or both – aimed at recalibrating the balance rather than suppressing a symptom.
Why a Subclinical Sample Matters Here
Studying schizotypy rather than diagnosed schizophrenia is a deliberate strength, not a compromise. It strips away the confounds that haunt patient research – antipsychotic medication, chronic illness, hospitalisation – and lets the mechanism be observed in its uncontaminated form. If the prior-weighting imbalance is already detectable along a personality dimension, it is more plausibly a cause than a consequence of illness, and it becomes a candidate marker of vulnerability that could be tracked long before a first episode.
From Salience to Speech
Much of the salience-attribution literature has focused on reward and visual surprise. This study extends the same predictive logic into language, the most socially loaded domain of human inference. Hearing what we expect to hear is the ordinary glue of conversation; the finding suggests that the social brain's drift toward over-interpretation and the perceptual brain's drift toward hallucination may be two readings of a single overweighted dial.
The mechanism that lets a healthy listener reconstruct speech in a noisy room is, when its gain is set too high, the same mechanism that manufactures a voice that was never there.
The sample sizes are modest, particularly in the neuroimaging study, and the glutamate association is correlational rather than causal. Schizotypy is a proxy for psychosis proneness and the findings need replication in clinical populations and across languages before any diagnostic or therapeutic use. Task-based hallucinations are an analogue of clinical hallucinations, not the symptom itself.