Knowing When to Disagree with AI: Override Selectivity as a Strategic Decision Capability

Human oversight of artificial intelligence (AI) is often evaluated by how much people rely on algorithms. That framing misses a more difficult strategic question: whether disagreement occurs when it should. We develop override selectivity as the capacity to allocate human resistance to cases in which the human has an observed comparative advantage while permitting AI correction when AI has the advantage. The behavioral diagnostic is the Override Selectivity Index (OSI), the difference between override rates in these two conflict states. We ground the upstream representation boundary in Discovery Plane Theory (DPT) and the downstream consequence boundary in Experience Architecture (EA), while keeping both formal frameworks distinct from the behavioral measure. We reanalyse two independent human–AI experiments comprising 5,810 real trial-level decisions from 381 participants. Loan decisions show anti-selective disagreement (OSI = −.095, participant-bootstrap 95% CI [−.152, −.037]), whereas chess decisions show positive selectivity (OSI = .292, 95% CI [.228, .350]). Participant-clustered models reproduce this divergence. We then use FiFAR as a synthetic-expert learning-to-defer benchmark and validate the Managerial Strategic Override Dataset (MSOD) v1.0 as a prospective causal instrument with 10 strategic cases, eight factorial conditions, 80 unique stimuli, and zero AI-criterion assignment mismatches. MSOD contains no collected manager responses, so no behavioral claims are drawn from it. The combined evidence suggests that the strategic value of human oversight lies not in retaining a veto, but in learning when that veto is informationally justified.