Where Does the Answer Come From? A Scoping Review of Evidence Tracing, Hallucination, and Safeguards in Generative Medicines Information

Generative artificial intelligence can produce medicines information rapidly, yet fluent answers can conceal uncertainty about where claims originated, whether cited sources exist, and whether retrieved evidence supports the wording presented. This scoping review mapped how generative medicines information systems attach evidence to claims, define and measure hallucination and citation error, implement retrieval and provenance, evaluate claim–source alignment, and apply human verification, conflict handling, and escalation safeguards. A protocol-led scoping review used the Population–Concept–Context framework and followed JBI and PRISMA-ScR principles. Peer-reviewed studies of generative systems producing medication or medicines information were eligible when they reported extractable evidence-tracing, citation, hallucination, retrieval, or safeguard methods. Records were screened, charted, appraised, and synthesized through evidence-tracing and safeguard taxonomies. The evidence base was methodologically heterogeneous and concentrated in retrospective, cross-sectional, or simulated evaluations. Systems attached evidence through model-generated references, search-linked citations, constrained document retrieval, or retrieval-augmented generation, but citation presence did not reliably establish bibliographic validity, relevance, or claim-level support. Hallucination definitions varied across fabricated references, incorrect citation elements, unsupported factual statements, and source–claim mismatch. Retrieval reduced some unsupported generation yet introduced failures involving document selection, temporal validity, partial entailment, and evidence conflict. Human review was frequently recommended, but escalation thresholds, reviewer workload, audit trails, and prospective workflow performance were rarely evaluated. Evidence tracing in generative medicines information remains fragmented. Trustworthy evaluation requires separate assessment of source existence, source quality, retrieval completeness, claim-level support, contradiction, and human adjudication. Current safeguards alone do not establish clinical safety or deployment readiness.