Invest1 publisher3 min readPublished
Daily filing caps miss the 34 AI-flooding cases where only the documents got longer
A new study traces 87 per cent of documented demand surges at government services to one capability, cheaply produced text, and its authors say 84 cases prove existence without measuring how widespread the flooding has become.
The Investor · Invest desk
What happened
- A study documenting 84 cases of AI-driven flooding across 11 jurisdictions, co-authored by Chris Schmitz with Lewis Hammond of the Cooperative AI Foundation and Alan Chan of GovAI, goes to the AAAI ethics conference on October 12-14.
- Complaints to the UK housing ombudsman went from 2,600 in 2022 to just over 7,000 in 2025, and CFPB complaints about banks and lenders ran five times higher across the same window.
- Of the 84 cases, 50 show a rise in the number of filings, 76 show single filings growing longer and more complex, and 42 show both patterns at once.
- A matrix of 13 factors puts the sharpest near-term risk in services that pay off when a claim succeeds and where complex paperwork has held demand down, such as tax returns, small claims and property-value appeals.
- In 14 of the 84 cases agencies reinstated fees or blocked IP ranges, and Australia has mooted bringing back fees for Freedom of Information requests.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Any institution that budgets complaint handling per complaint is planning against a flow whose cost of production has fallen to nearly nothing on the filer's side and not at all on the reader's side.
- exposure A fee or an in-person requirement introduced because of machine-drafted filings reaches every applicant in the same queue, including the ones who wrote their own.
- precedent If reinstated fees survive contact with the complaints, the first instrument an agency reaches for becomes a price, and every service that pays out on a successful claim is a candidate for one.
- contradiction The CFPB's fivefold rise reads like a measured trend, but the study only kept services where someone had already blamed AI, so nothing here tells you how many services were untouched.
A case needed a plausible route by which AI cut the cost of applying, evidence that demand had actually moved, and an explicit statement from officials or a credible third party naming AI as the cause [3]. The researchers began with 2,288 government services across twelve countries [2]. Eighty-four survived, which is 3.7 per cent [1]. In 58 of them a government official pointed at AI directly [13]. That is 69 per cent of the sample resting on the word of the office carrying the backlog [6].
One of the two series a bank or a regulator would actually reach for tells you more than the other. The UK housing ombudsman numbers work out at about 2.7 times over three years, near 39 per cent a year compounded [4]. The CFPB figure is a multiple with no base attached: five times as many complaints about banks and lenders, and no filing count published for either end of the window [5]. Anyone sizing a complaints desk against that has to supply the denominator themselves.
The Bank of England has cautioned that autonomous AI systems could pose a threat to financial stability [20]. The study reports that there are no agents yet that go to agency websites by themselves [11], so what these 84 cases document is people filing, with drafting that costs almost nothing.
Seventy-six minus 42 leaves 34 cases where filings got longer without getting more numerous [2], against 8 where they got more numerous without getting longer [3]. Daily submission caps are the normal defence, and they price volume [9]. One filing at a German social court ran past 4,000 pages [8], and it sits comfortably inside a one-a-day limit.
One case in six drew friction [5]. The alternatives the paper lists all cost money: more staff, the agency's own AI, or a redesigned intake [16]. Fees and IP blocks cost the agency nothing, and the paper warns they fall hardest on poorer and less digitally literate applicants [19].
My read is that a fee reinstated to manage a queue is a screen, and what it screens for is willingness to pay. The counter-thesis is short: an office with a backlog has every incentive to name AI, and this study admitted cases on exactly such a statement [3]. What would settle it is a service that publishes filing counts and page counts for the same years, because a flat length distribution against rising volume puts the whole thing back to ordinary demand growth. For now the documented concentration is judicial and legal services at 19 cases, ahead of regulatory complaints at 10 and welfare and social security at 9 [12].
What to watch
- Whether the AAAI presentation on October 12-14 publishes the absolute CFPB filing counts that the five-times multiple omits.
- Whether Australia actually reinstates Freedom of Information fees. A reinstatement would put the friction response on a national scale.
- Whether any of the 14 agencies that chose fees or IP blocks reports what happened to filing volumes afterwards.