The authors argue that perception is Bayesian inference based on accumulation of noisy evidence and that, in masked priming, the perceptual system is tricked into treating the prime and the target as a single object. Of the 2 algorithms considered for formalizing how the evidence sampled from a prime and target is combined, only 1 was shown to be consistent with the existing data from the visual word recognition literature. This algorithm was incorporated into the Bayesian Reader model (D. Norris, 2006), and its predictions were confirmed in 3 experiments. The experiments showed that the pattern of masked priming is not a fixed function of the relations between the prime and the target but can be changed radically by changing the task from lexical decision to a same-different judgment. Implications of the Bayesian framework of masked priming for unconscious cognition and visual masking are discussed.