The BaseResearcharXiv
LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context
According to the abstract, Anastasiia Kucherenko, Francois Brouchoud, Dimitri Percia David and Andrei Kucharavy tested four small open-weight models on benign legal prompts. They report that stating a legal role, such as defense lawyer, raised refusal rates by 2 to 20 times over the baseline.
Why it matters
The same thing, for different readers.
What it means for you
Telling a small on-device AI model that you are a lawyer made it refuse ordinary legal tasks more often, researchers report.
For lawmakers
Legal professionals pushed toward on-premises models for confidentiality may get uneven access to help, the authors say, which could introduce bias.
For researchers
Authority-style prefixes systematically raised refusals in four small models, while a known role-play jailbreak prefix had mixed effects.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Four modern small open-weight LLMs evaluated on benign legal-domain prompts with and without authority-style prefixes, plus a role-play jailbreak prefix.
- What it found
- Per the authors, authority-style prefixes increase refusal rates by 2 to 20 times over the no-prefix baseline; the jailbreak prefix has mixed effects.
- What it does not show
- Four small models and benign prompts only; the authors say further investigation is essential.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- How is AI changing childhood, health and daily life? · Are AI companies putting stronger safeguards into practice?
- People
- Anastasiia Kucherenko (Wrote it) · Francois Brouchoud (Wrote it) · Dimitri Percia David (Wrote it) · Andrei Kucharavy (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 20, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-19-legal-context-overrefusal-small-llms