Swiss Federal Supreme Court's In-House AI — Anonymisation Model at 92% Precision and Recall, Plus ChatTF, Under Federal Audit

The Swiss Federal Supreme Court (Bundesgericht, Lausanne) runs AI in-house for data-protection reasons. Since 2021 it has used an AI model for quality assurance when anonymising judgments before publication; because commercial cloud computing was ruled out for data-protection reasons, the court procured two specialised GPUs and trained a legal Swiss RoBERTa-type entity-extraction model on its own premises. The court's head of IT reported recall of 92.26% and precision of 92.57% for the model, with the AI proposing additional anonymisation suggestions to human staff.
In 2025 the court introduced ChatTF, an internal alternative to commercial LLMs. Open-source models (Llama 3.3, Apertus) run without prompt anonymisation, while prompts sent to commercial models (ChatGPT, Gemini, others) are anonymised before sending and de-anonymised on return. Users rate answers with a thumbs-up so the system can be specialised to the court's needs, and the IT team advises using search engines for specific facts and LLMs for drafting, summarising and translating. Further use cases include explainable semantic search in case law.
The Swiss Federal Audit Office opened an audit of this use (sfao-25732, published 4 May 2026) asking whether the court works more efficiently thanks to AI, whether use complies with trustworthiness principles and whether synergies with other federal courts exist. The audit page gives the court's context (7,000–8,000 appeals a year, 40 judges, 170+ clerks, 2024 expenditure almost CHF 115 million, just under 7% on IT services); the detailed findings were not available to the curator.
Evidence caveat: the precision/recall figures are self-reported by the court in a conference presentation and no efficiency or time-saving measurement has been published. The practice is credible mainly as a sovereignty-first, human-in-the-loop design with independent audit scrutiny.

Read the full analysis: https://www.ch-open.ch/wp-content/uploads/2026/05/02-Daniel-Brunner.pdf