Who Is Really Making The Decision? Agency Without A Subject In Algorithmic Decision Systems

Abstract

Decision systems in which a person selects from options that an algorithm has generated, ranked, and time-stamped strain the model of agency underwriting most of our moral vocabulary: one subject, one intention, one act. This article argues that agency in such systems is best understood not as a property some node possesses but as a structural property instantiated in the allocation of decision-constitutive powers across a system. The argument proceeds by conceptual analysis, supported by a documented case set drawn from the AI Incident Database, regulatory instruments, and published evaluations of deployed models. Distributed agency is defended as a supervenient, multiply realisable, relational property whose bearer is a structure rather than a subject: weaker than collective agency, and requiring no intentions be attributed to artefacts. Four indicators — informational control, option-space shaping, temporal structure, and traceability — make the distribution readable in concrete systems, and are shown to be distinct from adjacent frameworks concerning cognitive integration and meaningful human control. Where retrospective blame fails for structural reasons, responsibility should attach to the power to design and revise the distribution; this is non-equivalent to outcome-based developer liability, and its limits are identified for cases in which no party holds revision power. Coding ten cases yields four morphologies, one of which the framework did not anticipate and which sceptical treatments of the responsibility gap struggle to accommodate.