Mis-Grounding AI Agency: Heteronomous Assemblages and the Metaphysics of Maintenance

Debates about AI agency typically ask whether individual systems possess the right intrinsic properties to count as agents. This paper argues that this framing rests on a systematic mis-grounding: our classificatory practices draw the grounding arrows from models to safety and agency facts when the more precise arrows run from heteronomous assemblages: networks of models, human labour, infrastructures, and institutions through model profiles to those facts. Drawing on grounding theory (Fine, Rosen, Schaffer) and social ontology (Epstein, Haslanger, Searle), I develop three nested claims. The Grounding Thesis holds that facts of the form “S is safe, aligned, or trustworthy in C” are grounded in the organisation of the heteronomous assemblage A(S, C) rather than in model-intrinsic properties. The Misgrounding Thesis holds that "compulsory autonomy," a metaphysical ideology that treats individual models as ontologically prior, inverts this structure, encoding the inversion in classificatory practices through which AI systems are individuated and governed. The Maintenance-Grounded Agency Thesis holds that the same model can cease to instantiate F(S, C) when its maintenance pattern degrades, not because the base stops causing certain behaviours but because it no longer satisfies the conditions that make “safe in C” true. The framework is defended against the realist challenge through pre-symptomatic cases and the Searlean account of status functions, and is shown to resolve puzzles about flickering agency, responsibility attribution, and temporal degradation that individualist approaches cannot adequately handle.