When making two consecutive perceptual decisions about a single visual stimulus, is an observer's second decision biased by the first decision? We tested a probabilistic model of conditioned perception (Stocker & Simoncelli, 2008) that hypothesizes that such biases emerge from the observer's goal to remain self-consistent between decisions. The stimulus comprised 52 Gabor patches (spatial frequency: 8 cycles/deg, uniform orientation) placed on a grid with jittered locations and presented foveally for 250 ms. Observers first performed two orientation discrimination tasks in isolation. Task 1: The observer compared the stimulus orientation to a reference orientation indicated by a pair of lines. Task 2: Two Gabor stimuli were presented consecutively. The observer compared the second stimulus orientation to the first one. To test the model, we estimated noise parameters for the model from these discrimination data (individually for each observer) and computed the model's predictions for performing the two tasks sequentially. In the sequential task, the observer first compared the stimulus orientation to the reference and then compared the orientation of a second stimulus to the first. The model observer, when making the second decision in the sequence, discards the sensory evidence for any orientation inconsistent with the first decision, and is thus more likely to respond "counter-clockwise" in the second task after responding "clockwise" in the first task, and vice versa. This predicts both a repulsive shift and a change of slope in the psychometric curve. An alternative model, which optimally uses all the available evidence, predicts a smaller shift and no change in slope. The observers' second decision was consistently biased by the first one. The measured psychometric curves were well predicted by the conditioned perception model and were distinct from the predictions of the alternative model. Past decisions thus condition the use of sensory information in subsequent perceptual tasks.