As machine learning has proliferated, so have concerns about fairness and bias. Much of this work, however, relies on two implicit assumptions: that fairness means treating groups the same, and that this definition applies uniformly across different forms of oppression, such as racism, sexism, and ableism. In this talk, I will challenge both assumptions, showing how they can produce harmful consequences. I argue that fairness sometimes requires recognizing differences between groups as well as developing identity-specific approaches that do not generalize across different forms of oppression. I will then apply these principles to the case of chatbot personalization. Ultimately, I show how prevailing notions of fairness risk going both too far and not far enough.