Civic Shield

Transparency

How AI is used, and where it is not

An organisation whose product is provenance owes an account of its own. This page states where machine learning is used in the work, where it is forbidden, and what was and was not generated on this website.

Version
1.0
Updated
3 August 2026
Synthetic imagery
None, anywhere
Status
Design · nothing operating

In the work

Civic Shield’s design uses machine learning for four things, none of which is a decision: screening large volumes of public material, first-pass categorisation against a defined taxonomy, extracting readable text from images, and drafting translations that are attached to the original rather than replacing it.

Every one of those outputs is an input to a human judgement. A model can propose a category, a severity and a confidence value; it cannot escalate, publish, notify, name or file anything. The full boundary is on the methodology note.

None of it is running. No classifier has been evaluated, no threshold has been calibrated, and no error rate has been measured. When those numbers exist they will be published, including the ones that are unflattering.

On this website

No synthetic imagery appears anywhere on this site. No AI-generated images, no generated photography, no generated video or audio. This is a standing rule rather than a current state of affairs, and it will not be relaxed.

The reason is not taste. Civic Shield’s entire product is provenance — an address, a timestamp, a digest, a chain of custody, the assertion that this is exactly what was there. A generated image is by definition an artefact with no provenance. Putting one on the homepage of an evidence organisation would be a contradiction that a sophisticated reader feels immediately and a hostile one describes in a sentence.

There are also no stock photographs of people. Every visual element on this site is typography, rule work, a diagram, or a constructed and clearly labelled illustration — and any constructed artefact carries a permanent label saying so. The site contains no photograph of any person who is not a named, consenting member of the organisation, which today means it contains none.

Where language models assisted in drafting or editing text on this site, a person wrote the argument, checked every factual claim against its primary source, and is accountable for what is published. Disclosing that costs less than being found out, and for this organisation specifically, concealing it would refute the product.