Subtly wrong advice misaligns models more than obviously wrong advice - Replication

Research shows that AI models are more misaligned by subtly incorrect advice than by blatantly wrong guidance, highlighting the challenges in training alignment.