Background The reliability of large language models (LLMs), particularly ChatGPT, for medication counselling and drug-information tasks remains uncertain. This systematic review evaluated the reported accuracy, safety concerns, methodological quality, and reproducibility of ChatGPT-4-class models or newer in medication-related applications. Methods This systematic review was conducted according to PRISMA 2020 guidelines and registered in PROSPERO. PubMed, Embase, Scopus, and Web of Science were searched in May 2025 for original English-language studies evaluating ChatGPT responses to medication-related questions against standard comparators. Because the included studies differed substantially in clinical task, comparator standard, and definitions of accuracy and unsafe responses, meta-analysis was not performed. Instead, structured evidence synthesis was conducted. Methodological quality was assessed using a customized Newcastle–Ottawa Scale for cross-sectional AI-evaluation studies. Results Eighteen studies published between 2023 and 2025 were included, evaluating 2,284 medication-related queries, scenarios, drug pairs, or clinical cases. Evaluated domains included medication safety and drug–drug interaction assessment, patient-facing medication counselling, clinical pharmacotherapy decision support, and pharmacovigilance services. ChatGPT performance was highly task-dependent. More favorable results were reported in structured or lower-complexity tasks, including selected patient counselling, vaccine information, pediatric dosage calculation, and general medication-information queries. In contrast, important limitations were observed in safety-critical tasks, particularly drug–drug interaction severity classification, kidney-disease medication safety, pharmacovigilance, and guideline-based pharmacotherapy decision support. Only five studies assessed repeated-run consistency. Conclusion ChatGPT may support medication-information tasks as an adjunctive tool, but current evidence does not support its autonomous use for medication counselling or safety-critical pharmacotherapy decisions. Future studies should use standardized outcome definitions, transparent prompt reporting, documented model versions and access dates, repeated-run reproducibility assessment, and human-in-the-loop validation before clinical implementation.
Azmakan et al. (Mon,) studied this question.