Contemporary discourses accompanying the deployment of machine learning tend to fit within a meta-narrative of automation. Whether heralded as inevitable or criticized as reductive, this process of automation tends to be considered as concurrent to technology’s extension in general. By drawing on the social and technical history of machine learning, I would like to suggest that play offers at least one other way of framing machine learning’s development, one that could help foster other expectations and evaluations of machine learning performances. The genealogy of machine learning I will provide draws on certain historical tropes to problematize contemporary debates. I will distinguish two normative paradigms: automation and play. Each one expresses differing, albeit not incompatible values, expectations and objectives when evaluating machine behaviors and their interactions with human ones. Whereas automation pushes us to consider machines as ideally working by themselves, play requires a social and affective engagement whose outcome is always partially unpredictable. It helps account for the relative open-endedness, intractability and recursiveness of machine learning systems embedded in social practices. More specifically, I underline why play provides a sweet spot for thinking about performances such as those we are increasingly seeing in machine learning systems, that combine both rule-following behaviors and forms of improvisation upon those rules. Play includes automaticity as a level of behavior among others. The more general claim I am making is that play gives us a window onto a different history of machine learning and for imagining other forms of social interaction with and through technology.