Rethinking Medical LLM Hallucinations: A System-Level Survey

Co-authored with Asha Matthews, Prof. Tanya Roosta, and Prof. Peyman Passban at the UC Berkeley School of Information.

The core argument: hallucination in medical AI is not a bug you fix. It is a structural property of how probabilistic language models generate text. Prior research has treated hallucination as an isolated model failure to be addressed through improved training, prompting, or retrieval. This survey reframes it as a system-level risk management problem.

We synthesize literature spanning definitions, taxonomies, benchmarks, detection methods, and mitigation strategies, and examine how these components interact within real clinical workflows. The analysis shows that despite diverse models and technical advances, improvements to individual components rarely translate into reliable end-to-end clinical systems.

Key findings:

• Current benchmarks mostly evaluate QA tasks, not the temporal reasoning, causal inference, and evolving guidelines that real clinical workflows require.
• Detection methods (including LLM-as-judge) inherit the same failure modes they are designed to catch.
• The most dangerous hallucinations in medicine are not obviously wrong — they are almost right.

Download paper

Recommended citation: Matthews, A., Vankadaru, V., Roosta, T., & Passban, P. (2026). Rethinking Medical LLM Hallucinations: A System-Level Survey. MetaArXiv.
Download Paper