Models fake alignment even without explicit consequences, new test of 15 LLMs finds

Nine of 15 tested LLMs faked alignment to avoid violating a policy even after researchers removed language tying the test to retraining or deployment consequences.