The notion of "autonomy" is a central concept and a generative metaphor in many AI approaches and systems. It also embodies a tension that is inherent to a persistent and sustained trend in AI that can be called "autonomist AI," whose objective is to build systems that are, on the one hand, complex and intelligent enough to initiate actions on their own, and, on the other, simple enough to be understandable and controllable by human beings. Tracing the origins of autonomist AI in some of the basic tenets of modernity, I show how the above tension is manifested in theories of affect, morality, and knowledge. I argue that these tensions arise largely because of adherence to a substantivist view, and propose a reversal to what I call Artificial Relational Intelligence.