Study documents 84 suspected cases of LLM text flooding government services

A newly posted preprint argues that government services are already absorbing a surge of AI-generated applications, appeals and comments, and that agencies have few good ways to respond without hurting the people those services exist to help. The paper, "Characterizing Agentic Flooding of Government Services," went up on arXiv in August 2026 and reached the Hacker News front page, drawing attention from practitioners. Its core claim is that tools which make polished writing nearly free are colliding with public institutions sized for a far lower volume of hand-crafted submissions.

A new name for an emerging problem

The researchers lead with the upside. AI agents can genuinely improve access to government, helping people apply for benefits, parse dense policy language and lodge their views on proposals. The trouble is what happens when that convenience scales. Once contacting an agency costs almost nothing in time and effort, demand can rise past what the receiving end can process. The paper labels these surges "agentic flooding," shortened to "flooding."

The evidence so far

To gauge how real the effect already is, the authors assembled a dataset of 84 potential flooding cases spread across 11 jurisdictions. The word "potential" matters: this is a collection of suspected instances, not an audited census. Even so, on the strength of that evidence the researchers posit that flooding is probably occurring widely today, with the bulk of it attributed to large language models producing text at near-zero cost rather than to more elaborate autonomous systems.

Where the risk concentrates

The paper's second contribution is a risk matrix for scoring how exposed a given service is. Applying it, the authors conclude that near-term danger sits with services that combine a financial payoff with administrative complexity. The logic is easy to follow: a program that distributes money gives people strong reason to submit, and a complicated application or appeal process rewards whoever can produce the most complete-looking paperwork, which is precisely what an LLM does cheaply.

The response dilemma

The third contribution maps what governments can actually do. Judging by precedent, the authors expect the available responses would be enough to halt most flooding cases. The catch is which responses are fastest to deploy. Measures that add friction, with fees cited as the leading example, can be rolled out quickly, but friction lands hardest on applicants with the least time and money. A government that grabs the quickest lever risks trading away equitable access, inverting the accessibility gains that AI was supposed to bring in the first place. The paper closes by recommending near-term actions meant to blunt flooding without triggering that trade-off, though its abstract does not spell out what those actions are.

Why it matters

This is an early attempt to name and measure a second-order effect of consumer AI: not what the models can do, but what happens to institutions once everyone can use them at once. For agencies, the risk matrix offers a practical screening tool for deciding which intake processes to shore up first. For policymakers, it reframes the debate: the question is not whether to respond to machine-written volume, but whether to respond with friction or with slower, structural changes to capacity. The paper's reception on Hacker News suggests the people who build and operate these systems already recognize the pattern, even before regulators do.