Many people are excited by functional neuroimaging as a new tool for cognitive science; many others are sceptical. In this opinion article, I describe a 'forward inference' that one can make from patterns of brain activity to distinguish between cognitive theories. I give an example of forward inferences in research on recognition memory, and outline some statistical criteria for a 'qualitative difference' in brain activity. Forward inferences resemble the dissociation logic long-used in behavioural studies of healthy and brain-damaged people, although I argue that dissociations in neuroimaging data can go beyond behavioural dissociations. Nonetheless, forward inferences are only as good as the cognitive theories to which they pertain, and are most valuable in conjunction with other types of inference.