The Agency Error in AI Governance: Coherent Output, Constraint, and the Misclassification of Artificial Systems

AI systems do not act. They generate coherent outputs under constraint. The persistence of agency attribution in artificial intelligence governance arises not from the presence of intention in these systems, but from a structural feature of their design: sufficiently constrained systems operating over structured data reliably produce the observable markers—coherence, responsiveness, apparent goal-directedness—that humans use to infer agency in other contexts. The inference is not irrational. It is misplaced. This paper identifies that misplacement as the central, undertheorized problem in AI governance. Across philosophy of mind, legal scholarship, and science and technology studies, AI systems are treated as candidates for agency—as full agents, quasi-agents, or distributed actants—despite the absence of the conditions agency requires. This paper argues that these frameworks share a category error: the conflation of output coherence with underlying subjectivity or purposive action. Drawing on philosophy of mind, legal theory, and science and technology studies, it develops a non-agentic model of AI systems as constraint-bound output generators and demonstrates how prevailing frameworks—the intentional stance, legal doctrines of responsibility, and distributed agency models—fail on their own terms when applied to systems of this kind. The governance consequences are direct. Misattributed agency produces unstable liability regimes, because responsibility cannot be correctly located when the wrong entity is treated as its source. It produces conceptual drift in policy language, as terms such as “AI decision-making” and “AI autonomy” migrate from heuristic to descriptive and embed false ontological assumptions in regulatory design. And it produces misplaced regulatory targets, directing institutional attention toward the system rather than toward the human decisions—in design, deployment, and use—where governance can actually take hold. This paper proposes a framework grounded in operational classification rather than metaphoric attribution. The reframing does not alter who bears responsibility—designers, deployers, and users remain accountable—but whether responsibility can be deflected. The agency frame grants AI systems enough ontological status to absorb causal attribution without bearing legal responsibility, creating a structural gap in accountability. The non-agentic model closes that gap: a constraint-bound output generator cannot function as a deflection point, and the defense—“the algorithm decided,” “the system determined”—becomes categorically unavailable. Every outcome traces to a human decision. That is what correct classification makes possible.