Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Safety training can teach a model to reject political manipulation while still answering factual political questions, but it must measure both behaviors. In a research team's account, training examples pair prompts about the same topic with different intentions, one to refuse and one to answer. Adding permissible examples near that boundary reduced mistaken refusals on held-out benign prompts from 32.94% to 4.16%, while refusal of harmful prompts fell from 91.88% to 87.72%. The result shows a measurable trade-off: a large reduction in blocking legitimate questions alongside a smaller reduction in rejecting harmful ones.