AI agents overwhelm public services with a flood of requests

The ease with which tools like ChatGPT or Claude allow people to fill out forms and draft complaints is driving a sustained rise in requests to public bodies in several countries. In the United Kingdom, complaints to the housing ombudsman have more than doubled since ChatGPT arrived: from 2,600 in 2022 to just over 7,000 last year. In the United States, the Consumer Financial Protection Bureau (CFPB) recorded a 5-fold increase in complaints over the same period. The article notes similar jumps in court petitions in Brazil and parliamentary petitions in Germany.

Researcher Chris Schmitz is documenting this phenomenon, which he calls "agentic flooding." In a study to be presented next month at the AI Ethics and Society conference, he analyzes 84 separate cases of possible request flooding across 11 jurisdictions, and finds broad evidence that AI tools are changing the way people interact with public services. For methodological reasons, the study stops short of claiming that AI is the direct cause of the increase, but most of the 84 cases follow a similar pattern: submissions stayed relatively flat before 2022 and then grew at an ever-faster rate as AI technology spread. One telling detail is that, in most cases, that growth has shown no sign of slowing, suggesting it will keep rising in the coming years.

According to Schmitz, it is easy to see why AI's growing capability and availability drives greater use: people discover this is something that can be done and, little by little, it becomes easier and easier. It used to require gathering a lot of context and writing very precise instructions for ChatGPT 3.5; now it may be enough to paste in some text or take a photo of a letter with the Claude app to get a pretty good answer on the first try.

The article draws a parallel with what happened last year in bug bounty programs, whose inboxes were flooded with low-quality reports generated by language models. Those reports rarely contained real security flaws, but companies were still obliged to review them one by one, a considerable drain on resources. It is easy to imagine public services facing a similar problem if they have to handle five times as many applicants on the same budget.

Schmitz qualifies this, however: the situation is not identical. While bug bounty programs mostly received worthless submissions, most of the new requests to public services come, he says, from real people with legitimate claims. In his words, "the vast majority of cases we found are people who are entitled to claim something, claiming that thing." If those people did not file their requests before, it may be down to what the public policy world calls "administrative burden": the sheer effort of completing complex paperwork deterred many applicants who were entitled to benefits or compensation.

That is precisely the optimistic reading Schmitz offers: if AI can ease that administrative burden, it opens an opportunity to redesign many of these services in a way better suited to the AI era, rather than simply treating the phenomenon as artificially generated spam. As he puts it, "a big part of AI going well is being able to spell out what the good version of things looks like. And anyone who has used ChatGPT to do their tax return knows there is a good version of this, where you are being helped." For the researcher, this could be the moment to completely rethink how these administrative processes work, though he acknowledges it is a huge task that has barely begun to be tackled.