The ICO flagged a real problem with AI-generated FOI requests. We're built to avoid it.

In May 2026 the Information Commissioner's Office published guidance for public authorities on handling AI-generated Freedom of Information requests. We've been on both sides of the problem it describes: receiving these requests inside a council, and now writing requests professionally. The guidance describes exactly what disciplined request-writing is built to avoid.

What the ICO actually said

The guidance, published 6 May 2026, responds to a real shift in volume: monitored public authorities received 94,526 FOI requests in 2025, up from 83,041 in 2024, the highest number in a single year since monitoring began in 2005. Alongside the volume, the ICO flags a pattern in how AI-generated requests read: longer, broader and less focused than before, harder to pin down to what's actually being asked, containing factual inaccuracies, and in some cases citing case law or ICO decisions that don't exist.

The ICO is clear that this doesn't make a request invalid. FOIA is technology-neutral: using AI to help write a request doesn't, in itself, disqualify it, and a request stays valid unless it fails the basic requirements of section 8 (in writing, a name, an address for correspondence, and a description of the information wanted). What the guidance tells authorities to do instead is use the tools they already have: ask for clarification, apply the section 12 cost limit where a request is genuinely too broad to answer within it, treat truly vexatious volumes as vexatious, and meet their section 16 duty to advise and assist rather than reject a request outright because it reads oddly.

Why this happens

An AI tool asked to write an FOI request will do exactly that: generate something that sounds plausible, uses the right vocabulary, and cites the Act. What it doesn't know is which authority actually holds the answer, which statutory duty the request sits under, or what a council's FOI team can realistically search for without triggering a cost-limit refusal. That's not really a flaw in AI. It's a flaw in asking a general-purpose tool to stand in for someone who has actually done the job.

What we do differently

Every request we send is written by someone who was on the other side of this exact process: reading incoming FOI requests, deciding what a council actually held, and drafting the replies, across thirty years in UK local government including as a service head. That shapes the actual requests, not just the marketing copy: naming the specific statutory duty a request sits under, as in our piece on the two-tier deflection, targeting the tier of government that actually holds the answer, and keeping each question narrow enough to survive a section 12 challenge instead of getting refused for being too broad.

This isn't an argument against AI

We use AI too, for classifying hundreds of replies and tracking statutory deadlines across a live campaign, exactly the kind of repetitive, structured work it's good at. The line sits somewhere specific: a person who has actually done the job decides what to ask and reads what comes back. AI helps with the parts that don't need judgement. It doesn't write the request.