Categorizing objects is a central issue in several disciplines of the Social Sciences and Humanities (SSH). This is why we can think at first sight that machine learning is a great asset in an age of an ever-growing access to information. However, it turns out that many pittfalls along the road usually prevent us to leverage very powerful tools: lack of *meaningful* data, good definition or cost of a manual annotation, etc. What is more, the very task of categorization can be ill-defined and lead to a considerable waste of time. I will explain what I mean by the "dilemma of categorization" and give several illustrations that come from previous collaborative, multidisciplinary projects.