In the attitude attribution paradigm, observers must estimate the true attitude of an author who was assigned to advocate a particular position. Observers’ tendency to attribute an attitude in line with the expressed position despite its having been assigned is called the correspondence bias. While there is strong evidence that such attributions are externally invalid, it is less clear whether they are internally consistent. This research develops a Bayesian model that specifies what attitude an observer should attribute, given assumptions about the prior attitude distribution, and perceptions of the degree of compliance shown in the essay and the strength of the situation. The model reproduces classical findings regarding chosen vs. assigned positions, prior attitude probability, and degree of compliance, and also fits newly collected data. The results suggest that future research should examine observers’ assumptions and perceptions, and focus less on the reasoning process itself.