Traditionally, inductive learning algorithms such as decision tree learners have employed attribute-value representations, which are essentially propositional. While learning in first-order logic has been studied for almost 20 years, this has mostly resulted in completely new learning algorithms rather than first-order upgrades of propositional learning algorithms. To re-establish the link between propositional and first-order learning, we have to focus on individual-centered representations. This short paper is devoted to the nature of first-order individual-centered representations for inductive learning. I discuss three possible perspectives: representing individuals as Herbrand interpretations, representing datasets as an individual-centered database, and representing individuals as terms.