Current AI safety systems face a critical flaw: they often refuse legitimate prompts containing keywords related to dangerous topics. A civics tutor and a content filter might use the same underlying model but need opposite behaviors on election-related content—one should answer factual questions while the other should refuse manipulative requests. Yet most safety systems treat all election-related topics as monolithic dangers, blocking both educational and harmful content indiscriminately. Hugging Face researchers have published a paper proposing "boundary-aware safety," a more nuanced approach that addresses this over-refusal problem. Rather than refusing entire topics, the framework trains models to recognize and refuse only harmful subsets within a topic—such as political manipulation—while continuing to answer benign related questions. The researchers tested on political content, where they could clearly distinguish between factual political information that should be answered and manipulative persuasion that should be refused. The team identified three key failures in current safety training: coverage gaps where failed generation attempts are discarded, false refusals on benign prompts containing dangerous keywords, and inability to measure where safety boundaries should be drawn. Their solution involves retry mechanisms to recover failed examples, in-distribution training on benign-but-dangerous-looking prompts, and evaluation methods measuring boundary precision. Initial results show significant improvements in targeted refusal rates, though the team acknowledges trade-offs with over-refusal on certain benchmarks, highlighting the ongoing complexity of balancing safety with usability.