Normative Sponsorship Under Opacity: When Institutions May Own Machine-Mediated Commitments

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Institutions can become committed to machine-mediated claims even when no one can reproduce the model’s internal computation. The resulting problem is not solved by demanding complete technical transparency, nor by placing a nominal human “in the loop.” This paper develops Normative Sponsorship: the conditions under which an answerable institutional bearer may warrantably adopt machine-mediated output as its own position. Sponsorship is distinguished from deployment responsibility, causal involvement, approval, and post hoc liability assignment. A five-part anti-scapegoating constraint requires epistemic access, authority, review bandwidth, independent rejection capacity, and repair power. The paper then states an opacity threshold. Model-internal states may remain closed, but at the grain at which the institutional commitment can competently be contested, the relevant inputs, operative premises, governing rule or target criterion, and defeat conditions must be independently reproducible. Rejection must be unpenalized and institutionally effective. This yields a further distinction between opacity, where the basis cannot be independently reconstructed, and capture, where the basis is knowable but organizational design prevents disagreement from binding the outcome. The account supports collective sponsorship, requires reason provenance, and treats sponsorability as use-relative rather than a fixed property of a model. A broader overview appears in Empty Occupancy: Answerability and the Ownership of Machine-Mediated Institutional Commitments, Cott (2026); the present paper isolates the sponsorship argument and its opacity threshold.