The application of machine learning methods to neuroimaging data has fundamentally altered the field of cognitive neuroscience. Future progress in understanding brain function using these methods will require addressing a number of key methodological and interpretive challenges. Because these challenges often remain unseen and metaphorically “haunt” our efforts to use these methods to understand the brain, we refer to them as “ghosts”. In this paper, we describe three such ghosts, situate them within a more general framework from philosophy of science, and then describe steps to address them. The first ghost arises from difficulties in determining what information machine learning classifiers use for decoding. The second ghost arises from the interplay of experimental design and the structure of information in the brain – that is, our methods embody implicit assumptions about information processing in the brain, and it is often difficult to determine if those assumptions are satisfied. The third ghost emerges from our limited ability to distinguish information that is merely decodable from the brain from information that is represented and used by the brain. Each of the three ghosts place limits on the interpretability of decoding research in cognitive neuroscience. There are no easy solutions, but facing these issues squarely will provide a clearer path to understanding the nature of representation and computation in the human brain.