To what extent is it possible to manipulate beliefs by providing interpretations of unknown events? I characterize the feasible posteriors across signals when the agent is exposed to a set of models to interpret observable signals and adopts the model that best fits what is observed. Because each signal could trigger the adoption of a different model, posteriors across signal realizations might not average to the prior. The scope of persuasion is large, even for a persuader who does not control or know the signal the agent observes. I apply this framework to political polarization, finance, lobbying, and self-persuasion.