Generative artificial intelligence is increasingly being evaluated for medication information, clinical-pharmacy questions, treatment recommendations, deprescribing, documentation, and other pharmacotherapy tasks. Its apparent efficiency can obscure a medication-safety problem: fluent output may be difficult to audit, expensive to verify, insensitive to changing patient context, or persuasive even when wrong. This Current Opinion argues that the clinically relevant boundary is not whether a model can generate a plausible answer, but whether the surrounding use condition preserves evidence traceability, feasible verification, interpretable uncertainty, and meaningful human override in proportion to the consequence of error. Available literature supports useful task-level capabilities, retrieval-supported improvements, and emerging methods for uncertainty estimation, while also documenting fabricated references, omissions, variable calibration, context-sensitive failures, and adverse effects of incorrect AI advice on human decisions. We therefore propose an assistance–risk boundary architecture in which generative AI remains assistive only when its evidential basis can be inspected, checking effort is commensurate with workflow capacity, uncertainty and context limitations remain visible, and a qualified human can reject or escalate the output before medication action. This argument does not establish a validated risk threshold, autonomous-use criterion, or medication-safety benefit. Model version, retrieval corpus, clinical task, drug and patient characteristics, local guidance, workflow design, user expertise, and post-deployment change remain material boundaries. Prospective, task-specific evaluation is required before higher-consequence pharmacotherapy responsibilities can be justified.
