Barus et al. published an article describing five cases of adverse drug reactions — recorded in the French National Pharmacovigilance Database — in patients who followed treatment recommendations proposed by AI.
The authors write that “Medication errors may arise from several intrinsic limitations of LLMs, including their probabilistic architecture, the occurrence of hallucinations, and the lack of clinical context available to the model. In addition, incomplete or poorly specified user queries may prevent LLMs from generating appropriate recommendations”. They list some limitations of their study and conclude: “These reports should not be interpreted as evidence that LLM use is inherently unsafe, but rather as illustrations of situations in which AI-generated advice may contribute to medication-related harm. (…) While LLMs may provide useful health information, both patients and healthcare professionals should remain cautious when interpreting medication-related recommendations generated by these systems.”
Barus R et al. When generative AI goes off the rails and leads to adverse drug reactions. Therapies. 2026. https://doi.org/10.1016/j.therap.2026.07.001.
santepublique-com.fr
HIFA profile: Bernard Seytre is a Consultant at BNSCommunication in France. Professional interests: Health communication and education. seytre AT bnscom.fr