Who decides when the AI disagrees? A defensibility-threshold theory of accountable AI adjudication

Research on algorithmic advice explains whether people trust, accept, or weight an AI recommendation, but says little about what an accountable professional does when their own judgment and that recommendation point to different actions. We theorize this situation as accountable AI adjudication, in which the professional settles the disagreement through deference, override, verification, or escalation, ending in a final allocation of decision authority. We argue that the choice follows a defensibility standard expressed as two thresholds on the professional’s belief that the AI recommendation is correct. Below the lower threshold the professional overrides, above the upper threshold they defer, and between the two the case is unresolved. Error-cost asymmetry and the joint condition of accountability and evidentiary auditability move the thresholds, while the cost and availability of verification and escalation decide how unresolved cases are settled. The same AI, equally accurate and equally trusted, can therefore be deferred to in one case and rechecked in the next. Developed for accredited testing laboratories, the theory yields nine falsifiable propositions, separates defensibility from trust, and links reliance research to organizational theory on authority. It implies that raising AI accuracy or trust will not by itself improve decisions, because evidentiary records, decision rules, and escalation design also determine whether AI advice helps.