This is a time of both anxiety and excitement for the translation sector. Global connectedness and dramatic improvements in artificial intelligence mean that prototypical professional translation, or translation as we knew it, is likely to become marginal to society in the short to medium term. It no doubt will be replaced, however, by a range of activities characterized by varying degrees of professionalism, flexible multilingualism and human-computer collaboration that will make translation broadly conceived even more central. This opens up exciting perspectives for empirical research on translation, but arguably calls for reflection on our established research paradigms, from the very basic issue of defining and delimiting our object of study, through to the range of methods we use to investigate it. To make this discussion more concrete, I will refer to two very different scenarios: one concerning machine- and learner-translated texts, the other translation in multilingual news production. Both case studies point to the centrality of parallel corpora as a research infrastructure, but at the same time signal the need for more flexibility in how we conceive of well-established translation and corpus categories and assumptions.