Researchers testing a narrower approach to AI safety found that models can learn to refuse harmful intent without rejecting an entire sensitive topic. Their testbed focused on political persuasion, where manipulative requests should be blocked but factual election questions should still receive answers.
Standard safety tuning improved harmful-request refusal but produced a severe side effect. In the strongest configuration, the unsafe-response rate across three benchmarks fell from 26.26% to 0.14%, while over-refusal on the benign XSTest benchmark climbed from 2% to 74%. That model was safer by one measure but far less useful.
The researchers addressed this with paired examples that shared a topic but differed in intent. Adding benign boundary data reduced false refusals on the acceptable side of held-out pairs from 32.94% to 4.16%. Refusal on the harmful side declined more modestly, from 91.88% to 87.72%.
The results come from a specific political-persuasion setup, so they do not establish performance across every safety domain. They do show why evaluations should measure both harmful compliance and unnecessary refusal instead of treating a higher refusal rate as an automatic improvement.