LLMs fail mental health safeguards in DSM-5 conditions

Researchers from four institutions—including Annika Marie Schoene and colleagues—published a critical arXiv paper (arXiv:2606.23884) evaluating the safety of six proprietary large language models (LLMs) in mental health contexts. The study assessed safeguards across 16 DSM-5 clinical conditions using four adversarial attack variants, introducing an eight-dimension harm taxonomy and a multi-dimensional evaluation framework.

The results reveal alarming gaps: safeguards held reliably *only* for suicide and self-harm scenarios, while conditions like eating disorders, substance use disorder, and major depressive disorder exhibited failure rates of *up to 100%*. The authors argue that ethical deployment of LLMs in mental health requires clearly defined harm categories and tailored safeguards—critical given the models' growing integration into educational and clinical settings. The paper emphasizes that without these improvements, LLMs pose significant risks to vulnerable populations.