This is the draft of a paper which treats machine translation as an experiment. It examines some types of negative results produced by this experiment. The paper starts with an analysis of machine translation systems. The components of these systems embody sophisticated natural language processing methods. They could cope in principle with obstacles raised by meaning and the knowledge of the world. The paper concludes that the main obstacle is the inability to read the mind of the author of a text. This should not be a problem if the human mind had a linguistic structure. The most plausible conclusion seems to be that the mind's structure is not linguistic.