From AI Systems to Decision Events: Reframing the Unit of Accountability in AI Governance

This paper identifies the Governance Object Problem in AI governance: when AI systems mediate human judgment, delegation, and responsibility, current governance frameworks designate the AI system as the governance object while the AI-mediated decision/action event — the point at which AI output becomes human or institutional action — remains without a stable governance designation. The paper asks one bounded research question: how does system-centred AI governance leave the AI-mediated decision/action event under-specified as a unit of accountability? Three mechanisms are identified — Tool-Covered Agency, Accountability Object Collapse, and Authority-Action Decoupling — and their common consequence is described as the Event-Level Governance Gap. Empirical illustration is drawn from an audit of N=46 UK public-sector AI transparency records (ATRS corpus). Content-provenance technologies (SynthID, C2PA) are examined as boundary cases that record artefact lineage but not decision-event accountability. The paper draws on governance theory (Rhodes 1996, 1997; Bovens 2007) to anchor the claim that accountability requires a human actor with conduct capable of explanation and judgment — a condition that system-level governance alone does not satisfy. Related Zenodo corpus: https://doi.org/10.5281/zenodo.17992916 (Root Index)ATRS evidence pack: https://doi.org/10.5281/zenodo.20181288 This is a working paper prepared for possible journal review. It is not a peer-reviewed publication. A reading of representative papers across AIES proceedings (2018–2025) is used as indicative field-mapping only, not as systematic bibliometric evidence. Author note: AI writing assistance was used in drafting and revision. Research design, conceptual framing, theoretical positioning, source selection, and all final intellectual judgments are the author's own. The author reviewed, selected, revised, and approved all text.