The human auditory system rapidly encodes auditory regularities. Evidence comes from the oddball paradigm, in which frequent (standard) sounds are occasionally replaced with a rare (deviant) sound. Deviants relative to standards typically elicit signs of prediction error (e.g., MMN and P3a). It is, however, less clear whether deviants, which also bear predictive information but are encountered less often than standards, might inform auditory prediction. To investigate this, naïve participants listened to sound sequences constructed according to a new, modified version of the oddball paradigm: two kinds of deviants differing in their probability of repetition, yield the sound actually following a deviant either conditionally likely or unlikely. As this sound is either the same deviant (repetition) or a standard (no repetition), it is either unlikely or likely with respect to the global stimulus probability at the same time. In an active deviant detection task, we replicated previous behavioural findings, demonstrating that predictive information carried by deviants (conditional probability) is extracted when behaviourally relevant. Our analyses further reveal that respective effects at the level of response times increase with time. However, in a passive listening setting both MMN and P3a were confined to violations of rules based on global probability, while not being sensitive to conditional probability. Though some sensitivity to conditional probability had been observed in a previous study, these effects were tiny compared to those of global probability. Thus, the auditory system seems to mainly rely on rules that are encountered frequently (standard regularity), at least during passive listening.