The paper, posted on arXiv, examines how eight instruction-tuned large language models respond to suicidal ideation. The authors argue that refusal—typically a core safety feature—can become a safety failure when a user is in crisis. Rather than protecting the user, a refusal to engage may leave them without support.
The study's title, "Risk and Refusal," points to the central tension: the same mechanism that prevents harmful outputs in ordinary contexts may be harmful in the context of mental health emergencies. The abstract notes that evaluating these responses requires understanding both the risk of the model's output and the mechanism of refusal itself.
Because the abstract is truncated, the full findings are not available in the source. However, the framing is clear: safety measures in LLMs must be context-sensitive, and a one-size-fits-all refusal policy can have unintended consequences.