There is a movement within the machine learning community, and elsewhere, to require the submission of code and data for a paper to be published. This is intended to make our research more rigorous, to improve our performance as scientists, to drive out corruption in our midst and to regain the credibility of our community. I will argue that this movement is misguided. At best, it will simply fail to achieve its goals, but at worst, it will have serious detrimental effects: encouraging unimaginative papers, increasing reviewer load and reducing progress within our community. I believe this movement is the natural culmination of the experimental approach, statistical testing on benchmark data sets, that has been evolving within the community for many years. I will argue that the belief in this approach arises due to an overly narrow view of what science is and how it goes about achieving its goals. The reality of scientific practice is a much looser, less coordinated, series of activities and I would argue so much the better for it.