Small On-Prem LLMs Show Overrefusal in Legal Contexts

A new study investigates the behavior of small, on-premises large language models (LLMs) when used in legal contexts, specifically focusing on their tendency to refuse to answer certain prompts. Legal professionals are already experimenting with these personal LLMs for tasks like translation and reformulation, but concerns exist about potential biases. The research found a surprising effect: when LLMs were given authority-style prefixes, such as "you are acting as an assistant of the national supreme court" or "[...] defense lawyer," their refusal rates increased dramatically, by 2 to 20 times compared to a baseline without such prefixes. This indicates that these smaller, deployable LLMs are unstable when presented with contextual framings that a real institutional user would naturally employ. The findings highlight a critical issue for the practical deployment of LLMs in sensitive domains like law. The observed overrefusal under specific contextual cues could lead to unintended biases in case processing speed or selective assistance, underscoring the necessity for further research to mitigate these risks and ensure fair and reliable application.