Predictive medication-safety systems are commonly evaluated as tools that identify patients, orders, or situations at elevated risk. Once embedded in clinical work, however, prediction becomes part of the system generating the outcomes on which later safety judgments depend. Recommendations can redistribute clinician attention, alerts can alter prescribing behavior, suppression and prioritization can change which hazards remain visible, and successful interventions can modify the labels and data streams used for subsequent monitoring or retraining. These effects complicate the interpretation of apparent improvement or deterioration after deployment. A model may retain acceptable technical performance while the surrounding clinical system develops new workload, reliance, or residual-risk problems; conversely, apparent statistical drift may partly reflect successful intervention. This Current Opinion examines predictive medication safety as a sociotechnical intervention rather than an isolated classification task. It distinguishes automation bias from appropriate reliance, alert fatigue from alert migration, model drift from intervention-induced feedback, and visible alert reduction from reduction in total medication risk. The resulting position is conditional: predictive systems can improve defined medication-safety outcomes, but their safety cannot be inferred from discrimination, calibration, alert counts, or override rates alone. Evaluation must include how the deployed system changes behavior, information production, error visibility, and the residual distribution of medication risk.
