# MARS-Gov beats prior bias detectors on Dutch government texts

> The new MARS-Gov system reaches 0.880 F1, cutting false interventions to 2.5% for Dutch bureaucratic documents.

Oossa · 2026-10-09 · https://oossa.com/en/mars-gov-beats-prior-bias-detectors-on-dutch-government-texts

Researchers introduced MARS-Gov, a multi‑agent tool that spots biased language in Dutch government documents, rewrites it, and checks the rewrite. It adds a dynamic “10th juror” to handle groups not covered by fixed taxonomies. On the DGDB benchmark, it achieved a 0.880 F1 score, 20.2 points higher than the best zero‑shot large language model detector and 6.8 points above the top supervised Dutch encoder.

## The facts

- Published: Fri Oct 09 2026
- F1 score: 0.880, false interventions: 2.5%

## Why it matters

Dutch officials could rely on fewer false bias flags, saving time while still catching harmful language.

## Sources & references

1. [The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection](https://arxiv.org/abs/2610.11136) – arXiv, 2026-10-09

Last updated: 2026-10-09