Humans are able to use knowledge of previous events to estimate the probability of future actions. Consequently, an unexpected event will elicit a prediction error as the prepared action has to be replaced by an unprepared option in a process known as "action reprogramming" (AR). Here we show that people with Parkinson's disease (PD) have a dopamine-sensitive deficit in AR that is proportional to the size of the prediction error. Participants performed a probabilistic reaction time (RT) task in the context of either a predictable or unpredictable environment. For an overall predictable sequence, PD patients, on and off dopamine medication, and healthy controls showed similar improvements in RT. However, in the context of a generally predictable sequence, PD patients off medication were impaired in reacting to unexpected events that elicit large prediction errors and require AR. Critically, this deficit in AR was modulated by the prediction error associated with the upcoming event. The prolongation of RT was not observed during an overall unpredictable sequence, in which relatively unexpected events evoke little prediction error and the requirement for AR should be minimal, given the context. The data are compatible with recent theoretical accounts suggesting that levels of dopamine encode the reliability, i.e., precision, of sensory information. In this scheme, PD patients off medication have low dopamine levels and may therefore be less confident about incoming sensory information and more reliant on top-down predictions. Consequently, when these internal predictions are incorrect, PD patients take longer to respond appropriately to unexpected sensory information.