Artificial intelligence–supported recommendations can influence clinical decisions long before any institution formally describes an algorithm as a decision-maker. This creates a problem that cannot be resolved by predictive accuracy alone: a recommendation may remain formally advisory while acquiring practical authority through workflow design, clinician dependence, default effects, interpretive asymmetry, or difficulty of override. This theoretical article examines when that transition occurs and what it implies for responsibility, patient contestability, and justified reliance. Using Decision-Rights / Normative Responsibility Theory, the analysis distinguishes predictive performance from practical authority and treats authority as a distribution of operative decision rights among clinicians, institutions, patients, and AI-mediated processes. The argument develops in three steps. First, clinical authority is shown to emerge when AI-supported outputs materially structure which options are considered, how disagreement is handled, and what burden falls on clinicians who depart from the recommendation. Second, responsibility is tied to actors’ actual capacity to interpret, override, explain, escalate, and revise AI-mediated decisions, while patient contestability is defined as the effective opportunity to challenge both the recommendation and the institutional conditions under which it was relied upon. Third, justified reliance is conceptualized as conditional alignment among epistemic warrant, operative decision rights, answerable responsibility, and contestability. The resulting account rejects both accuracy-based authority and purely trust-based approaches. AI-supported recommendations may properly shape care, but reliance becomes normatively defensible only when the human and institutional actors expected to answer for the decision retain meaningful authority over it and patients retain a credible route to question its use.
