Safety classifiers for large language models typically map prompts directly to labels, but a new paper argues that this misses a crucial intermediate step: the user's intent. The authors propose that modeling intent explicitly—as a signal between the prompt and the final safety label—can lead to more robust classification, especially on hard cases where the surface text is ambiguous.

To support this idea, they introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with intent information. Using this dataset, they demonstrate that intent-aware training improves safety classification across multiple training regimes, not just in one narrow setting. This suggests that incorporating intent is a generally useful design choice rather than a fix for a single failure mode.

The paper is a single source, so there are no conflicting findings to compare. Its main contribution is both a new resource and a clear argument for rethinking how safety classifiers are structured. The results indicate that paying attention to why a prompt was written, not just what it says, can make safety systems more reliable.