How many submissions would you expect on a typical development application to a local council? In the City of Sydney, until May 2026, the record was 800, set by a particularly contentious business development. That record was broken after a go-kart track’s lease came under threat from competing demands on the car park it uses. The operators appealed to social media, and their community, aided by AI, made 3,682 submissions, or, as the operator gleefully pointed out, 345,000 words that would take council staff 45 days to read — almost all in favour of the application.
It’s a silly story, but the order-of-magnitude jump is the same hockey-stick curve now showing up in consultations and government services worldwide, as large language models collapse the cost of writing a legally fluent submission. Governments around the world are not prepared to manage or draw genuine signal from the noise of the extraordinary volume of submissions.
There’s a name for this trend: agentic flooding.
In their paper, “Characterizing Agentic Flooding of Government Services”, researchers at the Hertie School, the Cooperative AI Foundation, and GovAI document 84 cases across eleven jurisdictions: German social courts reporting a 55 percent year-on-year caseload jump tied to AI-generated claims, Australia weighing the return of Freedom-of-Information fees after a wave of AI-drafted requests, self-represented litigants multiplying in U.S. federal courts. In nearly nine of ten cases the mechanism is simpler than an autonomous agent filing on its own: a model drafts, a person hits send. The services most exposed pair financial stakes with gatekept complexity — court claims, tax filings, benefits appeals. Australia’s Fair Work Commission recently reported a 40 percent jump in cases between 2023–24 and 2024–25, pointing to generative AI as a contributing factor.
That leaves overwhelmed governments at a fork in the road. Made well, easier submission is a genuine democratic good; it lowers the barrier to civic processes that have always been harder to navigate than they should be. But when volume outstrips what any office can read, the choice is binary: build the capacity to actually listen at this new scale, or add friction (fees, narrower channels, caps on submissions) to make the volume manageable again.
Governments often reach for friction because it’s fast and familiar. But friction doesn’t distinguish a coordinated bot campaign from a constituent who finally found the words to say what they meant — it closes the door on the people public participation is supposed to hear from.
At RxC we come down consistently on the side of building capacity, increasing access, amplifying voice, and helping governments listen better. It is the same instinct behind Jennifer Pahlka’s Recoding America and the new bipartisan Recoding America Fund: if you want institutions to serve people, give them the ability to do the work rather than ration access to it. This was the spirit behind our submission to the U.S. House Administration Committee’s Subcommittee on Modernization and Innovation, in response to its recent call for ideas on the future of congressional modernization.
Our submission opens with a House-specific data point: requests to the Office of Legislative Counsel rose 72 percent in this Congress’s first sixty days, and its attorneys now report spending longer cleaning up AI-drafted text from members’ staffers than drafting from scratch. If that pressure is already reshaping how bills get drafted, it’s coming next for the constituent mailbox — the primary channel through which people reach their representatives.
Instead of adding friction, we propose member offices build capacity: a bounded, opt-in pilot in five to ten volunteer offices, recruited across both parties, using open-source synthesis tools already running elsewhere. These tools are robust against agentic flooding because their bridging-based algorithms look for positions that draw support from people who otherwise disagree. Volume buys nothing — ten thousand letters saying the same thing register as one view, held by one cluster in the system. What moves the needle is not how many messages back a position, but how many different clusters do. This keeps minority positions from being drowned out, and lets an office see when a view finds unexpected agreement across usual lines of division. To game this kind of system, one would have to fake not quantity but diversity — inventing many distinct constituents who disagree on everything but your point — which is what provenance checks are there to catch; and if a flood arrives as a genuine variety of perspectives, so much the better. That’s just more input, which is exactly what these systems are built to harness.
The good news: none of this is experimental. It’s tested, and just needs championing, adapting, and implementing — building on the vTaiwan experiments that made Taiwan the reference point for this kind of listening, Team Mirai’s recent work in Japan, the leading examples from Bowling Green, Kentucky and Engaged California, and RxC’s own work with MPs’ offices in Australia. Tools like Mirai Gikai, part of Team Mirai’s suite of feedback and sensemaking software, are proven at scale; what’s missing is embedding that capability into the day-to-day machinery of government.
As Kirsten Gullickson of the Congressional Data Task Force put it in July 2026: “Congress, of course, is a legislative, lawmaking institution; but in practice, it is also an information institution. Every bill, amendment, report, hearing, vote, and publication becomes part of an ecosystem of data that must be managed, preserved, and made accessible over time.” Constituent correspondence belongs to that ecosystem too, and right now it’s the part with the least capacity built around it — even though trust in these institutions depends on treating it the same way: ideas from the community, AI-aided or not, still need human engagement, parsing, and consolidation before they expand democratic participation rather than erode citizens’ faith in their own capacity to contribute to it.
That’s the throughline connecting agentic flooding to the rest of the House Modernization work we’ve been tracking: this is a capacity problem, not an access problem, and the time to build the solution is now. United States Legislative Markup (USLM) has made the U.S. Code and enrolled bills genuinely machine-readable, and the unified lobbying-disclosure portal, lda.gov, just launched. But the Comparative Print Suite and the Committee Activity Portal both stop short of the public, and the loudest current energy in House Modernization — the staff-facing AI assistants drawing press coverage this year — is aimed at helping insiders navigate Congress, not at helping Congress hear the public back.
Our submission asks for something narrower and more testable than a policy fight over that pattern: run the pilot, measure staff hours saved and synthesis accuracy against a documented checkpoint, and let the evidence decide what comes next. Agentic flooding isn’t a distant risk for legislatures to study from the outside — it’s already reshaping how the House drafts its own bills, and it already turned a go-kart track’s car park dispute into a 345,000-word test of one council’s capacity to listen. The question this modernization call poses, and the one our submission tries to answer concretely, is whether Congress builds the capacity to listen at machine scale — for diversity rather than volume — before the volume forces its hand, with consequences for civic participation everywhere.
Thanks to Marci Harris, Galen Hines-Pierce and Max Ghenis for conversations that sharpened this submission.