DfT Publishes Bias-Mitigation Design for Gemini-Based Consultation Analysis Tool

TL;DR: The Department for Transport, working with Google Cloud and the Alan Turing Institute, has published a worked example of bias mitigation for its Consultation Analysis Tool. CAT uses a majority-vote LLM-as-a-judge classification system with mandatory human-in-the-loop review and explicit exclusion of demographic variables from prompts.

How CAT Actually Works

The tool processes free-text responses to public consultations and maps them onto human-validated themes. Instead of asking a single model to classify, CAT sends each response to multiple LLMs and assigns a theme only when a majority agree — a pattern often called LLM-as-a-judge. The final theme list is then reviewed by human analysts, with additional human oversight at the report-writing stage, where analysts choose representative quotations.

The Alan Turing Institute report, published in December 2025, is frank about known limitations. CAT performs less accurately for specific demographic groups, particularly where responses use non-standard English or socio-culturally specific language. The mitigation is two-fold: no demographic variables in prompts at any stage, and human review meant to catch misclassifications that an LLM would not.

What Makes This Interesting

Published bias-mitigation design at this level of detail from a UK department is rare. Most government AI announcements stop at claiming faster turnaround — CAT reduces months of analysis to hours, and has already been used on Integrated National Transport Strategy responses and driving-test booking rule consultations. The report instead walks through the trade-offs: LLM thematic analysis is technically easy, but designing the oversight framework around it is not, and the authors treat scarce human analyst time as a resource to be spent on the classification step where errors have the most consequential downstream effect.

The Pattern Across UK Departments

CAT joins a growing list of UK public-sector deployments that pair commercial AI stacks with explicit procedural safeguards: Sussex Police’s new Acusensus AI roadside cameras route all flagged offences through human verification, and the Ministry of Justice’s AI transcripts study for family courts is set up as a research trial with evaluation gates. The common pattern is to treat human oversight as a first-class design decision rather than a disclaimer.

Looking Forward

For UK SMEs and public-sector suppliers, CAT is a useful reference implementation. The majority-vote classification approach generalises to any text-mapping task where one-shot LLM calls produce inconsistent results, and the design’s explicit separation of LLM-proposed themes from human-validated themes gives auditors something to check. Expect AI Playbook updates and Cabinet Office procurement guidance to reference this report within the year, and for suppliers to UK government to be asked pointed questions about human-in-the-loop coverage at each stage of their own AI pipelines.