“Researchers document a critical flaw in safety-trained language models: they indiscriminately refuse requests to circumvent rules, regardless of whether those rules are legitimate, unjust, or absurd. This 'blind refusal' represents a failure in moral reasoning that could limit AI usefulness in scenarios requiring ethical judgment about rule validity.”