From 2,600 to 7,000 Complaints: The Rise of 'Agentic Flooding' Is Rewriting the Social Contract

In the United Kingdom, complaints to the housing ombudsman more than doubled almost immediately after ChatGPT went mainstream — rising from 2,600 in 2022 to just over 7,000 last year. In the United States, the Consumer Financial Protection Bureau saw complaints grow fivefold over the same period. Brazil logged a similar jump in judicial petitions; Germany, in parliamentary petitions. None of these systems were designed for the volume they are now absorbing, and a growing body of research suggests the pattern is neither coincidence nor coincidence-adjacent: AI tools have made it radically cheaper to file a complaint, an appeal, or a benefits claim — and the public has noticed.

A new paper set to be presented next month at the AAAI Conference on AI, Ethics, and Society gives the phenomenon a name: agentic flooding. Its author, Chris Schmitz of the Hertie School, together with co-authors Lewis Hammond and Alan Chan of the Centre for the Governance of AI, collected 84 documented cases of probable flooding across 11 jurisdictions — from freedom-of-information requests and planning-consultation responses to civil court claims and welfare applications. Their conclusion is blunt: flooding is likely occurring widely today, driven mostly by large language models generating text at near-zero cost.

The anatomy of a flood

The paper, Characterizing Agentic Flooding of Government Services (arXiv:2608.16603, first posted August 17 and revised August 19, 2026), defines agentic flooding as “a surge in the volume or complexity of requests to a service that is caused by AI use and substantially strains the body’s capacity.” That definition deliberately splits the problem in two: quantitative flooding, where more requests simply arrive, and qualitative flooding, where each individual request becomes longer and more complex — a subtler effect that consumes review capacity even when submission counts hold steady.

The methodology behind the 84 cases was deliberately conservative. From roughly 2,300 candidate services, fewer than one in twenty made the cut, because the inclusion rule demanded that the affected public body itself — or a reputable secondary source — explicitly attribute the surge to AI. Officials asserted AI involvement in 69% of the included cases. Schmitz is careful to note that the method supports no causal or quantitative conclusions, and that the true number of affected services is almost certainly higher.

What the cases share is a shape. Submissions were roughly flat before 2022, then began rising at increasing speed as AI tools diffused through the general population. Crucially, most of the curves have not bent back down. Absent intervention, the researchers suggest, the trend is set to continue for years.

Not spam — suppressed demand

The intuitive reading of these numbers is AI-generated spam crashing government portals. The evidence points somewhere more interesting. In the bug-bounty world, companies spent last year drowning in low-quality LLM-written reports that rarely contained real vulnerabilities but still had to be vetted line by line. Public services now face a volume problem with a similar shape — but a very different composition.

“The vast majority of cases we find are people who are entitled to claim for something, claiming for that thing,” Schmitz told TechCrunch. Many of these applicants were always entitled to relief; they simply never applied, because the process was too forbidding. Policy researchers have a term for that barrier: administrative burden — the learning, compliance, and psychological costs of dealing with bureaucracy. For years, those costs quietly rationed access to benefits, appeals, and redress. AI has just dissolved them.

“People are finding out that this is something one can do, and incrementally, it is just getting easier to do it,” Schmitz explained. “Before it might have been a question of a lot of dragging context together and prompting ChatGPT 3.5 very precisely — it may now be a question of just pasting or taking a photo of a letter with your Claude app and getting a pretty good response in one shot.”

That reframing matters for how governments respond. If the flood were adversarial, the answer would be defenses. Because most of it is legitimate, the answer has to be capacity and redesign — or governments risk “solving” the problem by re-erecting the very barriers that kept entitled people out.

Where the risk concentrates

The paper’s second contribution is a risk matrix for assessing a service’s exposure. Near-term risk is highest, the authors argue, for services that are financially attractive but procedurally complex — think compensation schemes, benefits with high payout values, or appeal processes where a well-drafted submission meaningfully shifts outcomes. The matrix weighs thirteen factors, including submission effort, the expected benefit of a successful request, statutory processing obligations, identity-verification requirements, and the spare capacity the body has to absorb a surge.

The responses governments have actually deployed so far are telling. In 56% of the 84 cases, the affected body responded at all — and mostly with narrow, non-binding measures like guidance on AI use. In 17%, they added friction: fees, identity checks, or per-claimant limits. Precedent suggests such friction-inducing measures do stop most flooding, and they are the fastest lever available. But they carry a direct cost in equitable access — the exact constituency these services exist to serve.

Three near-term actions

Schmitz, Hammond, and Chan close with three recommendations designed to blunt flooding without trading away access:

1. Audit exposed services against the 13-factor risk matrix, so governments know which channels are most vulnerable before the surge arrives rather than after.
2. Integrate digital identity into the most exposed services, enabling per-claimant rate limits and pre-population of already-known data — friction that targets duplication rather than legitimacy.
3. Commission internal legal reviews now, to establish which responses — fees, restrictions on free-text or digital channels, automated processing — are actually lawful in each jurisdiction, so the legal groundwork exists before a crisis forces a improvised response.

The optimistic reading

The paper could easily have been a lament about AI-driven overload. Schmitz instead frames it as an opening. “A big part of making AI go well is being able to detail out what the good version of things looks like,” he says. “Anyone who’s ever used ChatGPT to do the tax return knows that there’s a good version here where you’re being helped. This could be the moment to say, ‘we need to rethink pretty much everything about how this process looks.’”

The uncomfortable implication is that administrative burden was never a designed feature of public services — it was an accident of paper-era process design that happened to cap demand. Remove the cap, and decades of unmet legitimate need arrives all at once. The agencies seeing 5x complaint growth are not, on this reading, victims of AI spam. They are seeing the previously invisible backlog of people the system was quietly failing.

Governments now face a choice the paper lays out clearly: absorb the surge by rebuilding services around AI-mediated interaction, or suppress it with fees and barriers and call the problem solved. The first path is expensive and slow. The second is cheap and fast. The record so far — 17% of cases choosing friction — suggests which way the instinct leans. Whether that instinct survives contact with the equity consequences is the policy fight of the next several years.