There is much interest currently in using functional neuroimaging techniques to understand better the nature of cognition. One particular practice that has become common is 'reverse inference', by which the engagement of a particular cognitive process is inferred from the activation of a particular brain region. Such inferences are not deductively valid, but can still provide some information. Using a Bayesian analysis of the BrainMap neuroimaging database, I characterize the amount of additional evidence in favor of the engagement of a cognitive process that can be offered by a reverse inference. Its usefulness is particularly limited by the selectivity of activation in the region of interest. I argue that cognitive neuroscientists should be circumspect in the use of reverse inference, particularly when selectivity of the region in question cannot be established or is known to be weak.