Environmental information available to our senses is incomplete and to varying degrees ambiguous. It has to be disambiguated in order to construct stable and reliable percepts. Ambiguous figures are artificial examples where perception is maximally unstable and alternates between possible interpretations. Tiny low‐level changes can disambiguate an ambiguous figure and thus stabilize percepts. The present study compares ERPs evoked by ambiguous stimuli and disambiguated stimulus variants across three visual categories: geometry (Necker cube), motion (stroboscopic alternative motion stimulus, SAM) and semantics (Boring's old/young woman). We found that (a) disambiguated stimulus variants cause stable percepts and evoke two huge positive ERP excursions (Cohen's effect sizes 1–2), (b) the amplitudes of these ERP effects are inversely related to the degree of stimulus ambiguity, and (c) this pattern of results is consistent across all three tested visual categories. This generality across visual categories points to mechanisms at a very abstract (cognitive) level of processing. We discuss our results in the context of a high‐level Bayesian inference unit that evaluates the reliability of perceptual processing results, given a priori incomplete, ambiguous sensory information. The ERP components may reflect the outcome of this reliability estimation.