A clerk opens the queue. The filing looks complete. Exhibit list. Case-law citations. A coherent story of facts. A few years ago this packet would have come from a lawyer's office, or it would not have arrived. Today it can come from someone with a laptop and a chat window. And the next one is already in the stack.
That pressure now has a name. Chris Schmitz, Lewis Hammond, and Alan Chan call it agentic flooding: a surge in the volume or complexity of requests that strains a government service after AI tools make it cheaper to start the request.[1] They named 2,288 candidate services during discovery. Fewer than one in twenty met all three inclusion tests. From that smaller set they kept 84 cases in 11 jurisdictions. Justice and legal services made up 23 percent of them.
This is an AIES 2026 paper sitting on arXiv, not a government audit. The method does not support causal or quantitative claims about how common flooding is, or how much of any backlog AI caused. What they have is a map of attributed surges. In 58 of the 84 cases, a government source said AI was in the mix.
And the map still has a direction. In 87 percent of those cases, a model drafts the text and a person finishes the forms. Autonomous browsing of government websites is not what courts and agencies appear to be absorbing yet. Humans remain in the loop. They are doing less of the writing. The queues they feed have not grown a matching amount of review time.
So the responses matter. Governments acted in 56 percent of the cases. The fastest move, in 17 percent, was friction: fees, IP blocks, rate limits. Another 25 percent put AI tools on the processing side. Schmitz and colleagues worry that friction wins because it can be stood up quickly, and that fees land hardest on people already close to the line of whether they could afford to ask.
Meanwhile a U.S. court record makes the volume harder to treat as a rumor. Anand Shah and Joshua Levy studied more than 4.5 million non-prisoner federal civil cases from fiscal 2005 through fiscal 2026, plus 46 million PACER docket entries.[2] They do not claim GPT-4 caused the break. The pro se share sat near 11 percent for years. By fiscal 2025 it was 16.8 percent, concentrated in formulaic document-production case types. Those extra cases are not wrapping up faster. Docket entries per court from pro se cases in their first 180 days were 158 percent above the pre-AI mean by 2025. On a random sample of 1,600 complaints, a detector flagged AI-generated text in 1.0 percent of 2023 filings and 18.0 percent in early 2026, with one false positive among 800 pre-period queries. A detector cannot settle merit. It can show generated-looking text arrived in the same period as the extra work.
Because drafting used to cost money, time, or both, that cost kept some people out and some volume off already-full desks. Both jobs are now being renegotiated at once. The receiving institution's throughput has not moved with the cost of asking.
By contrast, Australia already previewed the politics. In 2025 the federal government proposed a Freedom of Information overhaul and blamed AI-generated emails for clogging the system.[3] The opposition called the plan a "truth tax" and asked for the evidence.
This is a different cut than The Legal Clerk in the Machine, which followed citations, quotations, and signed authority after a document entered court.[4] The scarce job here is earlier: whether the queue itself can be believed, counted, and processed. Schmitz and colleagues hypothesize that near-term exposure is highest where a service is both financially attractive and procedurally complex, including court claims and tax filings. That is a watch list, not a measured ranking.
Verification bottleneck
Verification is becoming the scarce institutional function.
- Cheap text can produce more filings, and longer ones, faster than clerks, caseworkers, and judges can check what each packet actually asks for.
- Attribution is already scarce: Schmitz et al. only kept cases where an official or reputable source named AI. Quieter surges would stay off that map.
- Shah and Levy's detector trend is a text-origin signal. A clerk still has to verify facts, procedure, and what the filing is allowed to trigger.
- Watch whether agencies answer volume with fees and blocks, or with intake records that show what arrived, who submitted it, and what still needs a human decision.
Opportunities
Where value may appear is an intake-load worksheet for a single public counter or clerk team. For a sample week it would count filings, page length, whether the text looks machine-drafted, whether a human signed, and how many minutes of review each item actually received. The point is a local picture, not a national rate.
A submission receipt could travel with high-volume channels such as FOI, benefits appeals, or small-claims portals. It would keep the authenticated submitter, the timestamp, the source files, and the claims that still need human confirmation. It would support a reviewer. It would not decide the case. Idea fodder only. Not legal, financial, or policy advice.
The cheap part is the asking. The expensive part is still deciding what the asking is worth.
Sources
[2] Shah and Levy, Access to Justice in the Age of AI, March 2026
[3] Tom Williams, Information Age, September 4, 2025, Australian FOI overhaul coverage
[4] Hypernovelty Institute, The Legal Clerk in the Machine, July 3, 2026