Abstract

This paper introduces a structural account of agency, authority, and responsibility in artificial and distributed cognitive systems. It argues that prevailing debates in philosophy of artificial intelligence conflate behavioural performance with evaluative control, thereby obscuring the conditions under which normative attributions are warranted. To address this, the paper formalises evaluative structure in terms of two components: the source of evaluative standards and their revisability. This framework distinguishes systems that optimise within fixed criteria from those capable of transforming the criteria themselves. On this basis, the paper develops a strict condition for epistemic agency: a system is generative only if it participates in determining and revising the evaluative standards that govern its operation. Systems that lack such control are characterised as epistemic maintenance systems, regardless of their behavioural sophistication. The analysis shows that contemporary AI systems—including reinforcement learning from human feedback, preference-based alignment, and self-play architectures—operate under exogenous evaluative structures and therefore do not satisfy the conditions for epistemic agency or authority in the strong sense. The paper further argues that responsibility tracks control over evaluative structure rather than causal contribution or output generation. Apparent responsibility gaps arise from mislocating attribution at the level of system behaviour rather than at the level of evaluative authorship. This yields a corresponding account of authority: epistemic authority requires participation in the determination and revision of evaluative standards, not merely reliability or performance within them. The framework integrates and extends work in philosophy of action, social epistemology, and AI alignment. It provides a unified account of agency across human, artificial, and distributed systems, while clarifying the limits of behavioural criteria such as the Turing Test and the intentional stance. The result is a general constraint on normative attribution: claims about agency, authority, or responsibility are justified only where evaluative structure is made explicit and its locus of control identified.