We compared the effects of guessing versus choosing an upcoming task. In a task-switching paradigm with four tasks, two groups of participants were asked to either guess or choose which task will be presented next under otherwise identical conditions. The upcoming task corresponded to participants' guesses or choices in 75 % of the trials. However, only participants in the Choosing condition were correctly informed about this, whereas participants in the Guessing condition were told that tasks were determined at random. In the Guessing condition, we replicated previous findings of a pronounced reduction of switch costs in case of incorrect guesses. This switch cost reduction was considerably less pronounced with denied choices in the Choosing condition. We suggest that in the Choosing condition, the signaling of prediction errors associated with denied choices is attenuated because a certain proportion of denied choices is consistent with the overall representation of the situation as conveyed by task instructions. In the Guessing condition, in contrast, the mismatch of guessed and actual task is resolved solely on the level of individual trials by strengthening the representation of the actual task.