Rethinking Medical LLM Hallucinations: A System-Level Survey

With Asha Matthews, Prof. Tanya Roosta, and Peyman Passban.

Most prior work treats hallucination as an isolated model failure to be fixed through training, prompting, or retrieval. This survey treats hallucination as a structural property of how probabilistic language models generate text. In medicine, the errors that matter most are the ones that are nearly correct, because they pass surface checks. We synthesize work on definitions, taxonomies, benchmarks, detection methods, and mitigation strategies, and look at how those pieces interact inside real clinical workflows. Improvements to individual components rarely translate into reliable end-to-end systems.

The survey also named a gap that became the next project: almost no interpretability work had been done on medical LLMs specifically. Readable but Not Controllable is the empirical follow-up.

Cite as: Matthews, A., Vankadaru, V., Roosta, T., & Passban, P. (2026). Rethinking Medical LLM Hallucinations: A System-Level Survey. Artificial Intelligence in Medicine, in press.