Complaints to the UK housing ombudsman more than doubled after ChatGPT arrived, from 2,600 in 2022 to just over 7,000 last year. At the US Consumer Financial Protection Bureau, complaints over the same period rose fivefold. Brazilian court petitions and German parliamentary petitions show comparable jumps. Chris Schmitz, who has been tracking the pattern, calls it agentic flooding, and in a paper he will present next month at the AI Ethics and Society conference he works through 84 possible instances of it across 11 jurisdictions.
The curves have a common shape. Filings hold roughly steady until 2022, then start climbing and keep accelerating as AI tools spread. Schmitz is careful not to claim, for methodological reasons, that AI directly caused the increase, which is the right amount of caution for a correlation this clean and this convenient. In most of the cases the growth has not flattened, so the sensible expectation is several more years of it.
His explanation for the timing is capability plus reach. Getting usable output from ChatGPT 3.5 meant assembling a lot of context and phrasing the request precisely. Now it is often enough to paste in the text of a letter, or photograph it in the Claude app, and get a good enough answer in a single query. The work of turning a grievance into a formal submission fell from an afternoon to a minute, and the volume followed.
The nearest precedent is not a government one. Bug bounty programmes went through the same surge last year: inboxes filling with low-quality reports written by large language models, rarely containing a real security issue, each one still requiring a human to open it and check. That consumed real resources. The structural parallel for public agencies is exact and unpleasant — five times the claimants against the same budget.
Where the analogy breaks is on what is inside the envelope. Bug bounty triage was mostly reading junk. Schmitz's finding is that most of the new government filings come from people with legitimate claims. In the overwhelming majority of the cases he studied, the applicant was entitled to the service or payment and was applying for precisely that. Some of the new volume is plainly hostile, but the bulk of it is real claimants pursuing things they would previously have given up on.
Public policy has a name for what was stopping them: administrative burden. The form was the barrier, not the entitlement. An AI assistant that can read a denial letter and draft an appeal removes part of that burden, and every service Schmitz examined — from benefit applications to formal court appeals — runs through online systems an assistant can reach. The full dataset is posted separately.
This is the part worth sitting with, because the framing is doing a lot of work. Agentic flooding is a vivid phrase and, I think, a misleading one for what the data underneath it shows. A flood is something you defend against. What Schmitz describes is a system that was quietly under-serving eligible people for years, suddenly being asked to do the job it was designed for. From inside an agency the two are indistinguishable: both look like a queue that will not stop growing. The risk is that agencies read the metaphor rather than the finding, build filters against AI-assisted submissions, and catch exactly the claimants the programmes exist to serve.
The budget question is the one the research cannot answer. Schmitz's answer is redesign — rebuilding services so they are usable with AI assistance rather than in spite of it — and he is honest that this is large-scale work most organisations have barely begun. He treats the moment as grounds for optimism, and his broader argument is that describing in detail what good uses of AI look like is itself part of developing it safely. He points to the experience of using ChatGPT to complete a tax return as evidence that AI can genuinely walk a person through a procedure of this kind.
Optimism aside, the arithmetic is unforgiving on a timescale nobody controls. Redesigning a benefits system takes years; the filings are arriving now and, on his own reading of the curves, will keep accelerating.
These systems were built on an unstated assumption: that difficulty would ration demand, and that the gap between who is eligible and who applies would stay wide enough to fund the difference. That assumption has stopped holding, and nothing has been put in its place.