LLM safety filters are getting in the way of actual court work at Switzerland's highest tribunal.
The Swiss Federal Supreme Court already uses small, on-premises language models for translation and summarization across its four official languages — French, German, Italian, and English. That narrow, well-scoped deployment is exactly the kind of low-risk AI use that legal experts have said makes sense. The problem is that criminal case files routinely describe violent and sexual offenses in clinical detail, and current model guardrails treat those descriptions the same way they treat harmful requests — refusing to process them or burying outputs in disclaimers. Researchers from the court's environment introduced TF-RefusalBench, a multilingual benchmark built from public Swiss Supreme Court rulings, containing 5,200 prompts designed to measure where and how often this over-refusal happens.
The findings matter because they reframe a debate that has mostly focused on whether AI hallucinates legal facts. A model that refuses to translate a case file, or that wraps a summary in so many caveats that it becomes unusable, causes a different kind of harm — one that compounds across every employee who hits the same wall every day. The study also shows over-alignment varies by model, prompt language, and the language of the text being processed, meaning there is no single fix.
The researchers tested two remedies: careful prompting and "abliteration," a technique that removes refusal directions from a model's internals entirely. Prompting helped. Abliteration eliminated refusals with minimal hit to translation or summarization quality — though deploying a model with its safety directions surgically removed will raise its own questions for courts that have to justify their tooling to the public.