Modern theories of cognition place an emphasis on the (un)certainty of available information. This raises the question whether we trust more external sampled information or internal inference processes. The specific properties of visual processing around the blind spot region allow us to address this. Although there are no photoreceptors corresponding to the physiological blind spots, we experience visual content there as if it were veridical, when it is in fact only "filled in" based on the surroundings. We asked subjects to choose between a stimulus partially presented in the blind spot that elicits fill-in and another at the same eccentricity outside of the blind spot. Subjects displayed a systematic bias toward the blind spot stimulus, where the filled-in part could have actually concealed a non-target. Two control experiments confirmed this finding and demonstrate that this is not an effect of eccentricity, but a property of the filling in process. This intuitively puzzling effect finds a straightforward explanation within the context of predictive coding. The filled-in signals are produced by the brain's generative model based on spatial-context priors. In contrast to other locations, predictions at the blind spot cannot be compared to feed-forwards inputs and therefore no error signal is generated. As a consequence, the error measure for the inferred percepts reaches the lower bound and are estimated as more reliable than actually seen contents. This experiment gives credibility to the interpretation of bottom-up signals not as conveying independent information about the world, but information relating to deviations of internal expectations.