Normative competence and the two faces of AI agency

Are contemporary AI systems agents? This paper argues that the question conceals a systematic ambiguity between two concepts of agency: a functional concept, on which an agent is any system that pursues goals with sufficient sophistication, and a normative concept, on which an agent is a being that acts for reasons it takes itself to be answerable to. Drawing on the philosophy of action, I argue that the normative concept is the one operative in our practices of responsibility, trust, and moral legibility. The strongest philosophical defenses of AI agency leave the normative concept unsatisfied, even when they succeed in establishing genuine cognitive capacities. The ambiguity has practical consequences: when systems satisfying only the functional concept are deployed in roles constituted by normative practices, the results include responsibility gaps, the systematic performance of answerability without its substance, and a redistribution of normative burdens that benefits deployers at the expense of users.