Functional magnetic resonance imaging (fMRI) has quickly become the most prominent tool in cognitive neuroscience. In this article, I outline some of the limits on the kinds of inferences that can be supported by fMRI, focusing particularly on reverse inference, in which the engagement of specific mental processes is inferred from patterns of brain activation. Although this form of inference is weak, newly developed methods from the field of machine learning offer the potential to formalize and strengthen reverse inferences. I conclude by discussing the increasing presence of fMRI results in the popular media and the ethical implications of the increasing predictive power of fMRI.