Hallucinated dosing: An emerging medication safety risk in generative AI

Large language models (LLMs) are increasingly being integrated into health-system pharmacy and clinical workflows, where they support documentation, summarize patient cases, and respond to medication-related questions from clinicians and patients. While their outputs are fluent, confident, and often indistinguishable from human-written guidance, an emerging and underrecognized medication safety risk has become apparent: the generation of fabricated or incorrect medication doses that appear authoritative but lack grounding in pharmacological principles, clinical guidelines, or patient-specific data. Hallucinations, defined as generated content that is syntactically plausible but factually unsupported, are a well-documented limitation of generative language models.1 In medication-related tasks, hallucinations may manifest as incorrect doses, intervals, or administration instructions. The term hallucinated dosing has been introduced to describe medication recommendations generated by artificial intelligence (AI) systems without a valid evidentiary basis and expressed with unwarranted certainty. Although dosing errors from clinical decision support systems have previously been described, hallucinated dosing represents a qualitatively distinct failure mode that arises from probabilistic text generation rather than rule-based inference.