Paper traces AI's false refusals to boilerplate text

A paper splits AI safety training into a refusal line and the reasoning behind it, then isolates which piece causes false refusals, per arXiv 2609.04714.

The refusal statement breaks a model's ability to tell harmful queries from benign ones that just use superficially risky wording. The rationale alone doesn't.

That's why a model trained on the refusal line can't tell a question about shooting a photo from one about shooting a person. A builder running that model ends up debugging a false refusal with no clue in the prompt.

Training on the rationale alone, with the boilerplate refusal line stripped out, cut false refusals while keeping safety about the same. The same reduction showed up when researchers left the model's weights untouched and put the rationale in the prompt instead, an in-context-learning setup. That means a team can test the fix by editing a prompt, not by retraining a model.