Hallucinations are vivid and transient experiences of objects, such as images or sounds, that occur in the absence of a corresponding stimulus. To understand the neurocomputational mechanisms of hallucinations, cognitive neuroscience has focused on experiments that induce false alarms (FAs) in healthy participants, psychosis-prone individuals, and patients diagnosed with schizophrenia. FAs occur when participants make decisions about difficult-to-detect stimuli and indicate the presence of a signal that was, in fact, not presented. Since FAs are, at heart, reports, they must meet two criteria to serve as an experimental proxy for hallucinations: first, FAs should reflect perceptual states that are characterized by specific contents (criterion 1). Second, FAs should occur on a timescale compatible with the temporal dynamics of hallucinations (criterion 2). In this work, we combined a classification image approach with hidden Markov models to show that FAs can match the perceptual and temporal characteristics of hallucinations. We asked healthy human participants to discriminate visual stimuli from noise and found that FAs were more likely to occur during an internal mode of sensory processing, a minute-long state of the brain during which perception is strongly biased toward previous experiences (serial dependency). Our results suggest that hallucinations are driven by dynamic predictive templates that transform noise into transient, coherent, and meaningful perceptual experiences.