Paranoid delusions or unfounded beliefs that others intend to deliberately cause harm are a frequent and burdensome symptom in early psychosis, but their emergence and consolidation still remains opaque. Recent theories suggest that overly precise prediction errors lead to an unstable model of the world providing a breeding ground for delusions. Here, we employ a Bayesian approach to test for such an unstable model of the world and investigate the computational mechanisms underlying emerging paranoia. We modelled behaviour of 18 first-episode psychosis patients (FEP), 19 individuals at clinical high risk for psychosis (CHR-P), and 19 healthy controls (HC) during an advice-taking task designed to probe learning about others' changing intentions. We formulated competing hypotheses comparing the standard Hierarchical Gaussian Filter (HGF), a Bayesian belief updating scheme, with a mean-reverting HGF to model an altered perception of volatility. There was a significant group-by-volatility interaction on advice-taking suggesting that CHR-P and FEP displayed reduced adaptability to environmental volatility. Model comparison favored the standard HGF in HC, but the mean-reverting HGF in CHR-P and FEP in line with perceiving increased volatility, although model attributions in CHR-P were heterogeneous. We observed correlations between perceiving increased volatility and positive symptoms generally as well as with frequency of paranoid delusions specifically. Our results suggest that FEP are characterised by a different computational mechanism - perceiving the environment as increasingly volatile - in line with Bayesian accounts of psychosis. This approach may prove useful to investigate heterogeneity in CHR-P and identify vulnerability for transition to psychosis.