“Is This AI?” — The First Signs This Wasn’t Going to Be Simple

Four days after sending the national facial-recognition FOI, the first genuinely interesting responses started arriving.

And they immediately showed how differently police forces could approach exactly the same request.

Dyfed-Powys Police responded first by sending us three previous FOI disclosures concerning facial recognition and asking whether those earlier answers dealt with our new request. More unusually, they told us that our current request had not yet been dealt with under the Freedom of Information Act and that, unless we replied within three working days, they would assume the previous material answered our questions and close the matter.

It didn’t.

The earlier disclosures were useful background, but our request was specifically about evidence of deletion, audit, technical assurance, incidents and governance. We replied explaining that significant parts of the new request remained unanswered and that we wanted the current FOI to proceed.

Then came West Midlands Police.

Their response took a very different approach.

West Midlands said the request was too broad to process without clarification. They questioned terminology such as “audit trail functionality”, “technical assurance” and “Appropriate Policy Document”, and argued that phrases such as “any recorded” or “equivalent record” made the request difficult to search.

They also raised something we had not expected to become part of the correspondence at all: generative AI. The force said it was seeing an increase in broad or excessive requests which might have been drafted using AI tools, and suggested that AI-generated terminology could create unnecessarily complicated requests. It advised us to simplify the wording and linked to Information Commissioner guidance on AI-assisted FOI requests.

Some of the criticism was worth taking seriously. Questions 2, 3 and 4 did not explicitly state whether they applied to LFR, RFR and OIFR, so we clarified that point.

But other objections were harder to understand. An Appropriate Policy Document, for example, is not terminology invented for this request. Nor was asking whether a system produces an audit record intended as a demand for obscure technical architecture.

The exchange nevertheless taught us something important very early in the project.

If this investigation was going to work, we needed to be prepared to distinguish between a request that was genuinely unclear and one that was merely inconveniently detailed.

So we refined where refinement was justified, preserved the substance of the questions, and carried on.

Four days into the project, we had already learned our first lesson:

asking the same question of every force does not mean every force will interpret it the same way.

And this was only the beginning.