Accounting for why discrimination between different perceptual contents is not always accompanied by conscious detection of that content remains a challenge for predictive processing theories of perception. Here, we test a hypothesis that detection is supported by a distinct inference within generative models of perceptual content. We develop a novel visual perception paradigm that probes such inferences by manipulating both expectations about stimulus content (stimulus identity) and detection of content (stimulus presence). In line with model simulations, we show that both content and detection expectations influence RTs on a categorization task. By combining a no-report version of our task with functional neuroimaging, we reveal that violations of expectations (prediction errors [PEs]) about perceptual content and detection are supported by visual cortex and pFC in qualitatively different ways: Within visual cortex, activity patterns diverge only on trials with a content PE, but within these trials, further divergence is seen for detection PEs. In contrast, within pFC, activity patterns diverge only on trials with a detection PE, but within these trials, further divergence is seen for content PEs. These results suggest rich encoding of both content and detection PEs and highlight a distributed neural basis for inference on content and detection of content in the human brain.