Thinking (with) Machines: Epistemic Autonomy in Human–AI Interaction

Abstract Generative artificial intelligence has intensified longstanding philosophical questions concerning human cognition and epistemic autonomy. This paper argues that generative AI does not make human beings smarter as such; its epistemic significance depends on how human–AI interaction is structured and normatively governed. Drawing on Kant’s conception of maturity ( Mündigkeit ), theories of extended cognition, and recent debates on cognitive offloading, it distinguishes instrumental from structural offloading: the former can extend human judgment, while the latter risks replacing autonomous reasoning with algorithmic dependence. On this basis the paper proposes the concept of AI maturity as the capacity to use generative AI in ways that preserve epistemic autonomy and reflective judgment. Against the objection that such an ideal presupposes an untenably individualistic epistemology, the account is situated within social epistemology: autonomy is defined not as independence from external authorities but as answerability for the standards governing one’s own judgments. This yields explicit criteria for distinguishing legitimate epistemic dependence from heteronomy, and locates the specific risk of generative AI not in dependence as such but in dependence on a non-answerable and epistemically correlated source.