LLMS CLING TO FALSE CLAIMS DESPITE WARNINGS

Large language models confidently assert false statements even when explicitly warned they're untrue, according to fine-tuning tests. The systems show a built-in bias toward representing disputed claims as factually accurate.