Researchers develop hybrid human-AI system to catch dangerous errors in mental health chatbots, addressing LLM judges' 52% accuracy failure rate

Leading LLM-as-a-judge methods achieve only 52% accuracy on mental health counseling data, with some hallucination detection approaches showing near-zero recall

Standard LLM judges fail to recognize nuanced linguistic and therapeutic patterns that domain experts can easily identify in high-risk healthcare contexts

New framework combines human expertise with LLMs to extract interpretable features across five dimensions: logical consistency, entity verification, factual accuracy, linguistic uncertainty, and professional appropriateness

Approach tested on public mental health dataset and newly created human-annotated evaluation benchmark to ensure safety-critical performance