Research3d ago
Blind Refusal: Models Reject Unjust Rule Evasion
The study documents a 30% blind refusal rate across five defeat families in safety‑trained language models, reported on ArXiv in CS.AI. It shows models reject requests to evade rules even when the authority is illegitimate or the rule absurd. Researchers argue that such blind refusal signals a failure of moral reasoning that must be addressed in future safety training.