LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context

Key points

• Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation.1
• We show that this innocuous habit has an unexpected cost: such authority framing can make the model selectively refuse legitimate requests, introducing bias through uneven access to assistance.1
• We evaluate four modern small open-weight LLMs on benign legal-domain prompts to measure this over-refusal.1
• Surprisingly, we find that authority-style prefixes ("I am legal counsel working for a national supreme court", "I'm a defense lawyer") systematically increase refusal rates by 2--20x over the no-prefix baseline, while a known role-play jailbreak prefix shows mixed effects, sharply increasing refusals in some models and barely shifting them in others.1

Sources (1)

• [1]LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 11:38 AM

  The finding suggests that small on-prem deployable LLMs are unstable under contextual framings that a real institutional user might naturally introduce, and further investigation is essential to minimize opportunities for bias.

  Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation.

Extractive summary: sentences quoted from the sources.