·3 min read
If you cannot explain it, do not ship it
A public service can use a model. It cannot hide the rule, the source or the official who owns the answer.
Photo: Camera dei Deputati on Unsplash
Government answers have a second audience: the person who will ask how you knew. A fluent paragraph that cannot show the circular or the statute is a political problem waiting for a journalist.
The UK’s generative AI framework for government treats this as operational, not decorative. Use the tools. Do not pretend they replace the official who signs the advice.
The Stanford AI Index keeps charting adoption. Adoption is not the hard part. The hard part is a minister’s office that can still explain the number in the briefing.
Grounding here means the live policy, the last circular, the dataset the department actually holds, not a model’s memory of a blog post about that department.
Named owners matter more in public service than in a startup demo. The system should know which team owns the bot, which records it may read, and what it must refuse.
Publish the limitation. If the assistant cannot see classified or incomplete files, say so. Silence reads as omniscience, and omniscience is the thing you cannot defend.
Explainability is not a heatmap. It is a citation a civil servant can stand behind in a committee.
The TruFyre Way
If you cannot explain the briefing, do not ship it. TruFyre keeps the pack, the citation and the accountable official on one path for FOI and audit.