Task activation studies using functional magnetic resonance imaging (fMRI) are designed to detect which regions of the brain activate in response to particular stimuli, tasks, or thoughts. The approach to assessment of brain activation in fMRI has largely followed what is commonly known as the massively univariate approach: first estimate a parameter independently at each voxel representing activation level, and then second threshold the resulting map of estimates or p-values to provide a subset of voxels that are considered to be active. The idea is that any parameters above threshold are to be considered as representing activated brain voxels. In this talk we will explore Bayesian approaches to determining activation that attempt to improve on this two-stage process, either by combining the two steps, or by looking for alternatives to standard thresholding. We will review both existing approaches from ourselves and others, and discuss potential new avenues for improvements using the Bayesian paradigm.