Hallucinations are percepts experienced in the absence of external sensory input. Patients with hallucinations show both behavioral and neural alterations in statistical learning in simple sensory tasks. Predictive processing accounts of hallucinations suggest that these alterations bias patients’ perception toward their sensory expectations. However, 1) whether this framework holds when considering real-world hallucinations with complex content and 2) which upstream neural mechanisms cause these alterations in learning are hotly debated. To address these questions, we used fMRI to record neural activity while patients with schizophrenia and healthy controls were presented with a dynamic auditory environment. In a subset of participants, we also collected positron emission tomography (PET) and magnetic resonance spectroscopy (MRS) scans to measure differences in dopamine function and cortical glutamate, respectively. We find that patients experience auditory verbal hallucinations (AVHs) intermittently in the scanner (i.e., report hearing voices in objective silence), and that the rate of AVHs varies as a function of the statistics of the task environment. Importantly, this effect is stronger in patients with high hallucination severity suggesting that they are more sensitive to changes in sensory statistics. Building on neural models of predictive coding, we then constructed a computational learning model that updates trial-by-trial stimulus expectations via sensory prediction errors (sPEs). We find that the scaling of stimulus prediction errors in the auditory cortex negatively correlates with both hallucination severity and dopamine function in the associative striatum, but not cortical glutamate or dopamine in other brain regions. Overall, our results are consistent with a model of hallucinations whereby excess nigrostriatal dopamine drives changes in sensory learning and the probability of experiencing false percepts.