Trust Is Not a Single Outcome in Pharmacy AI: A Realist Review of Acceptance, Reliance, and Refusal

Trust is frequently treated as a desirable implementation outcome for artificial intelligence (AI) in health care. In pharmacy practice, however, acceptance, observed reliance, calibrated verification, overreliance, refusal, disengagement, and workarounds are distinct outcomes that may arise through different mechanisms and have different safety implications. To explain what forms of trust-related behaviour occur around pharmacy-relevant AI, for whom, under which technical, clinical, professional, organizational, and governance conditions, and why. A RAMESES-aligned realist review used an initial programme theory, purposive iterative searching, relevance-and-rigour selection, context-mechanism-outcome extraction, design-appropriate appraisal, contradictory-case analysis, and abductive and retroductive reasoning. The search identified 54 retrieval records. Four duplicates were removed, leaving 50 unique records: 41 evidence records and nine methodology or reporting sources. Five evidence records were excluded during title and abstract screening. Of 36 reports sought, one was not retrieved; 35 full texts were assessed, nine were excluded, and 26 reports representing 26 studies entered the realist synthesis. No quantitative synthesis was conducted.

The evidence indicated that acceptance could occur without observed reliance. Appropriate reliance appeared conditional on task fit, credible validation, uncertainty communication, training, workflow integration, and retained professional control. Incorrect or authoritative AI advice could produce overreliance, whereas safety, privacy, accountability, equity, relational, or autonomy concerns could support justified refusal. Poor integration and responsibility ambiguity could contribute to disengagement or workarounds. Increasing trust is not an adequate implementation objective. Pharmacy AI should instead be evaluated for calibrated decision use, including verification, override, escalation, refusal, and non-use. The proposed programme theory is explanatory and requires prospective validation.