Predictive processing is a domain-general account of the brain as a hierarchical, dynamically evolving Bayesian predictive model: of a volatile, stochastic world whose nested hidden states interactively and probabilistically generate sensory inputs. The brain infers the hidden causes of ambiguous, noisy neural activity at each level by predicting that level’s most probable state a priori, and integrating this with its estimate of the most likely state given the observed data. Discrepancies between expectations and observations – prediction errors – influence the posterior estimate in proportion to their precision, which itself must be inferred (from the estimated reliability of the data, relative to the confidence in the prior). A subtle, pervasive perturbation to the precision-weighting of prediction error signalling is theorized to explain the formation of psychotic symptoms. However, how predictive processing alterations at different levels manifest, across different phases of illness, and how these relate to phenomenological changes (such as the form of delusions’ relationship to evidence, confidence, and motivation) is not yet well understood.

Hypothetical predictive processing alterations, suggested by simulation and pharmacological models of the development of psychotic illness, were tested in participants with at-risk mental state (ARMS), first episode psychosis (FEP), or healthy cognition (HC) (Chapter 2). The use of low-level perceptual priors in visual discrimination (Chapter 3); the use of performance (sub)optimality to monitor meta-level confidence (Chapter 4) and neural signatures of confidence-modulated learning rate and choice temperature (Chapter 5); and the effect of beliefs about an external cue’s source on bias and sensitivity to it (Chapter 6) were investigated. A separate behavioural study investigated the perceptual inference and hierarchical reinforcement learning tasks (Chapters 3 and 4) in treatment-resistant chronic schizophrenia (CSZ) (Chapter 7). “Weak priors” accounts of trait-like vulnerability, previously challenged by empirical findings that ARMS participants made better use of priors to improve the sensitivity of visual discrimination than did controls, were tested on a modified version of that visual inference task whose demands more specifically probed lower-level priors. Having thereby prevented higher-level gist-based priors from being useful to performance on the visual discrimination required, I found that priors were underweighted (relative to controls) in ARMS and especially in FEP (Chapter 3), and also in CSZ (Chapter 7), consistent with the theoretical role of low-level priors in trait vulnerability to psychosis.

Chapter 4’s task and hierarchical reinforcement learning model had previously been used to demonstrate reduced confidence-modulation of learning and choice under ketamine. Surprisingly, this was not replicated in ARMS (a clinical model of the same prodromal construct for which ketamine is a pharmacological one): ARMS and HC did not differ from one another behaviourally or computationally (Chapter 4). FEP participants’ behavioural impairments superficially resembled those demonstrated under ketamine, but what characterized FEP at the computational level was their elevated baseline choice temperature relative to controls. This introduced “agentic uncertainty” into the (sub)optimality teaching signal used to monitor confidence, such that it less reliably related to the validity of lower-level beliefs – yet this epistemically-degraded metacognitive estimate, in FEP, modulated learning and behaviour to the same extent as observed in the other groups: with viciously circular ramifications for FEP participants’ ability to robustly maintain an adaptive model in the face of environmental stochasticity. Functional neural correlates of behaviour on this task, though broadly replicated in my larger sample, were not similarly affected in patients as they had been under ketamine. Instead of ketamine’s broad disengagement of the large dorsoparietal choice temperature network, in both FEP and ARMS there was suggestive evidence of a more circumspect disruption to (right) anterior insula, whose negative relationship with learning rate was replicated (Chapter 5). Interestingly, the behavioural case-control study of this task in Chapter 7 found reduced confidence-modulation of choice temperature in CSZ – in the absence of any performance impairment (though choices were on average riskier in CSZ than matched controls). The overall pattern across these studies resonates with classic and contemporary accounts of the adaptations acquired over time in the progression from acute psychosis to chronic schizophrenia. Irresolvable, uninformative uncertainty resulting from disorganization of response selection (in acute psychosis) gradually comes to be assigned less weight (in accordance with its newly degraded epistemic value). Learning to reduce the influence of metacognitive (un)certainty over the degree to which prepotent lower-level inferences constrain behaviour may permit some habitual engagement with salient aspects of non-delusional reality and weaken the motivational grip of firm delusional convictions.

Predictive processing’s domain-general, hierarchical, and dynamic Bayesian understanding of the brain as an evolving model of its environment (including the reliability of its own estimates thereof) fares well in explaining delusions’ psychological form and its temporal evolution. However, an adequate theory must also explain delusions’ content. The domain-specificity with which delusions usually centre around social themes has been the basis of challenges to predictive processing’s explanatory feasibility. The particularly taxing computational challenges inherent in social inferences may partially explain their especial vulnerability (Chapter 1), but this Bayesian account faces further problems in reckoning with the bizarre content of many more severe delusions (which may be less social, but whose implausible propositions are still thematically specific: typically involving fundamentally altered relations between mind, body, and world).

Chapter 6 manipulated a cue’s source (social/non-social), independently from its predictive validity. Controls were more biased towards the cue, and FEP participants were less biased towards it, when it was perceived to be social. In HC, sensitivity to the cue’s value was blunted in the Social condition – perhaps reflecting healthy developmental priors tuned to ecological statistics of the environment. Patients’ sensitivity was not affected by Condition. Though rational within the task, this suggests ARMS and FEP are both associated with a tendency to overestimate how reliably others’ mental states can be inferred from physical, observable ones. Chapter 4’s findings suggest that in FEP, this may extend also to one’s own mental states. Delusions’ domain-specific content may reflect disruptions to the domain-general statistical learning mechanisms by which agentic causal processes can be distinguished from mechanical or natural law-like ones. Failure to thus differentiate social from non-social sources may be computationally, as well as phenomenologically, core to understanding the content of delusions across the psychosis spectrum.