Study finds 84 cases of AI-driven flooding at government agencies worldwide

A study documenting 84 cases of “agentic flooding” across 11 jurisdictions connects the surge to cheap AI-generated text.

87% of the surges trace to one capability

Complaints to the UK’s housing ombudsman more than doubled after the arrival of ChatGPT, going from 2,600 in 2022 to just over 7,000 in 2025.

The US Consumer Financial Protection Bureau (CFPB), which manages complaints about banks and lenders, had five times as many complaints through the same window.

Petitions to the Brazilian judiciary and the German parliament mounted in a similar fashion.

The paper, co-authored by Chris Schmitz with Lewis Hammond of the Cooperative AI Foundation and Alan Chan of GovAI, will be presented next month at the AAAI Conference on AI, Ethics, and Society, October 12-14.

Also, the number of individual filings is growing quickly, and this is what the researchers stress about the most.

Quantitative flooding, more requests, and qualitative flooding, a single request growing longer and more complex, are not mutually exclusive. Of the 84 cases, 50 show quantitative flooding, 76 show qualitative flooding, and 42 show both.

One filing deposited with a German social court ran past 4,000 pages.

Normally daily submissions are capped to counter flooding. That doesn’t help when one person puts forward one huge document.

The same capability explains 87% of the surges in the sample. Language models can produce text cheaply and at scale.

There are no agents yet that go to agency websites by themselves.

14 of the 84 cases drew fees or IP blocks

The researchers commenced with 2,288 government services across twelve countries but retained only 84 that passed three tests simultaneously.

A case needed a plausible way for AI to have reduced the cost of applying, evidence that demand had actually shifted, and an explicit statement from officials or a credible third party that AI had been the cause.

Judicial and legal services crowned the list with 19, followed by regulatory complaints at 10 and welfare and social security at 9. In 58 of the 84 cases, a government official pointed the finger directly at AI.

Based on a matrix of 13 factors, the researchers ascertained that the sharpest near-term risk is in services that pay off when a claim succeeds and where complex paperwork has long held demand down, such as tax returns, small claims and property-value appeals. The paper is cautious about its limits.

Eighty-four cases do not prove causation or that services in general are affected. It’s a proof of existence, not a measurement.

Governments can dampen demand with fees, rate limits, or in-person requirements. Or add capacity with more staff, their own AI, or a redesigned intake.

In 14 of the 84 cases, agencies reached for friction, reinstating fees or blocking IP ranges. Australia has mooted bringing back fees for Freedom of Information requests.

But the catch, the paper warns, is that these fixes disproportionately impact poorer and less digitally literate applicants, closing the same door AI was opening.

The Bank of England has cautioned that autonomous artificial intelligence systems could pose a threat to financial stability, and early consumer agents have already gone wrong in the wild.

The smartest crypto minds already read our newsletter. Want in? Join them.