In two experiments we manipulated the prior probability of occurrence for two alternatives. After a first learning session, in a second session the cue to bias the decision was reversed. Our investigation shows that subjects are able to learn the reverse bias only when the bias of the first session is in line with their expected outcome. When, during the first session, the actual outcome of the bias is not in line with the expected outcome, there is an inhibition for the reversal bias learning in the second session. We investigate this phenomenon with computational models of choice showing that the inhibition of reversal is due to an increase in the rate at which subjects accumulate evidence for repeated, unexpected stimuli. We discuss a possible theoretical explanation that links this phenomenon to similar results found in the literature on reversal learning and to the effect of novelty on learning.