Designing Large Language Models That Refuse Unsafe Medication Questions Well

Large language models can produce fluent medication information while remaining vulnerable to factual error, missing context, unsafe personalization, adversarial manipulation, and misleading confidence. Refusal is therefore not merely a conversational limitation; when appropriately designed, it may function as a medication-safety control. This article develops a proposed Pharmacy Refusal-Safety Framework for determining when a model should answer, qualify, redirect, or refuse a medication question. The framework combines a multi-axial taxonomy of unsafe and unanswerable questions, a layered risk-interpretation architecture, an action-selection boundary, a non-abandonment response contract, and lifecycle governance. Questions are characterized according to potential harm, urgency, personalization, context sufficiency, epistemic answerability, premise validity, misuse potential, vulnerability, and professional-authority requirements. The selected action is constrained by the consequences of error rather than linguistic confidence alone. A safe refusal should state the relevant boundary, avoid covert individualized advice, preserve appropriate general information, identify missing context, direct the user toward a feasible next action, and provide urgent safety-net guidance when danger is plausible. Evaluation requires realistic safety-critical cases, expert adjudication, subgroup analysis, adversarial testing, user-comprehension assessment, workflow simulation, and prospective monitoring. The framework is an original non-empirical synthesis rather than a validated clinical algorithm or deployment standard. Its categories and transitions may behave differently across models, medicines, languages, users, jurisdictions, and care settings. Empirical validation is therefore required before clinical or pharmacy implementation.