Notes ·
AI Is Exposing How Much Government Capacity Depends on Making Services Hard to Use
The phrase "agentic flooding" sounds immediately negative.
I'm not convinced all of it is.
The paper looks at what happens when AI dramatically lowers the amount of work required for ordinary people to interact with government bureaucracies.
Apply for a benefit.
Appeal a denial.
File a complaint.
Submit a public comment.
Request government records.
Challenge an assessment.
Navigate a court procedure.
A lot of these systems are difficult enough that simply understanding what you are entitled to, figuring out the correct procedure and writing the required response can take hours.
An LLM can make a significant portion of that work cheap.
The result appears to be that people are submitting more requests, and the requests themselves are becoming substantially more complex.
The authors collected 84 potential cases across 11 jurisdictions where government officials or credible third parties connected a change in demand with the public's use of AI.
They are careful about what that means.
The dataset does not establish that AI caused all of these increases, and it cannot tell us how common this phenomenon is across government as a whole.
But there is enough evidence to suggest something real is happening.
And we aren't even talking about highly autonomous agents yet.
In 87% of the cases, the enabling technology was essentially an LLM helping someone produce sophisticated text cheaply. The person was still navigating the website, filing the paperwork and handling the rest of the process.
That alone is apparently enough to create noticeable pressure.
The paper found quantitative flooding, meaning more submissions, in 60% of the cases.
It found qualitative flooding, meaning more complicated submissions, in 90%.
Half had both.
Some of the examples are remarkable.
German social courts reportedly attributed a 55% year-over-year increase in cases during 2025 partly to AI-generated claims.
Australian officials have discussed bringing back fees for Freedom of Information requests after receiving increasing numbers of AI-assisted submissions.
The researchers also point to growth in lengthy filings from self-represented litigants in U.S. federal courts.
In one of the cases they studied, letters had grown to more than 4,000 pages.
There is obviously room for abuse here.
AI can generate terrible legal arguments just as efficiently as good ones. It can make fraudulent claims easier to produce. Someone deliberately trying to overwhelm a system can automate work that previously required substantial human effort.
But most of the cases in the dataset are not described as adversarial.
They are ordinary people using AI to interact with government.
That distinction matters.
We may have been using difficulty as a rate limiter
The most interesting sentence in the paper for me is essentially that some of the most vulnerable government services have historically been protected by friction rather than design.
That is a remarkable thing to consider.
A person may legally be entitled to appeal a decision.
But appealing requires understanding a complicated rule, finding the correct form, gathering evidence, writing a coherent argument and meeting a deadline.
Some percentage of people will simply stop.
The bureaucracy doesn't have to process those appeals.
The difficulty of navigating the system became part of its capacity planning.
AI changes the economics of that interaction.
If something that took four hours now takes 20 minutes, more people are going to do it.
If an AI can monitor a case, remember deadlines, interpret correspondence and help draft another appeal after a rejection, some people who previously would have given up will continue.
That looks like "flooding" from inside the agency.
From the citizen's side, it might simply look like finally being able to exercise a right they already had.
This is where the terminology makes me uncomfortable.
A legitimate appeal does not become illegitimate because software made it easier to write.
A Freedom of Information request doesn't become abuse merely because someone used an LLM to formulate it.
Someone discovering a government benefit they qualify for is not attacking the system by applying for it.
If making the process easier overwhelms the service, perhaps the service was depending on people failing to navigate it.
The obvious government response is also the worst one
The paper divides possible responses into two broad categories.
Governments can increase their capacity.
Or they can suppress demand.
Suppressing demand is much easier.
Charge a fee.
Add identity checks.
Introduce rate limits.
Require someone to appear in person.
Restrict digital submissions.
Block automation.
Make the process harder again.
The authors found governments had already introduced some kind of friction in 17% of the cases they studied.
That might solve the immediate capacity problem.
It also recreates exactly the administrative burden AI was removing.
And the people most affected by that friction are likely to be the people who already had the hardest time navigating government systems.
Someone with money can pay a filing fee.
Someone with a flexible job can attend an appointment during business hours.
Someone who can afford an attorney can have the attorney deal with the bureaucracy.
Everyone else gets the CAPTCHA.
That doesn't seem like a good outcome.
The other answer is to fix the systems
The more difficult response is increasing government capacity and redesigning services so they can handle the demand that actually exists.
Structured data instead of arbitrary documents.
Better digital identity.
Automated processing of routine cases.
Pre-populating information the government already has.
Clearer eligibility rules.
Better APIs and interfaces.
AI assisting the government with intake and processing just as AI is assisting the public with submitting requests.
The researchers found that governments are already experimenting with some of this. In 25% of their cases, agencies introduced AI tools in response to increased demand.
That may be where this gets particularly interesting.
AI helps citizens generate requests.
The resulting volume exceeds the bureaucracy's ability to process them.
The bureaucracy deploys AI to process the requests.
Eventually both sides of an administrative process may have software doing most of the mechanical work.
At that point, perhaps we should reconsider why the process requires thousands of words passing back and forth in the first place.
This is happening before agents really arrive
That may be the part I find most notable.
The paper calls this agentic flooding, but most of what they found isn't really autonomous agents yet.
It is people using LLMs as assistants.
Imagine when the agent can monitor every government program someone may qualify for.
It notices a change in eligibility.
Collects the required documents.
Prepares the application.
Submits it.
Tracks the response.
Recognizes a denial.
Reads the applicable regulation.
Drafts the appeal.
Submits that.
Tracks the deadline.
And keeps going until the administrative process is actually finished.
The psychological burden disappears along with much of the clerical burden.
That could generate vastly more interaction with government than simply making ChatGPT write a letter.
I think that could be incredibly valuable.
Government services are supposed to exist for the public.
If someone is legally entitled to something, making the process confusing enough that they never claim it should not be considered a successful scalability strategy.
AI may expose a lot of bureaucratic systems that have quietly depended on exactly that.
The danger is that governments respond by rebuilding the friction faster than they rebuild the capacity.
I hope we do the opposite.
If AI makes it possible for more people to actually use the rights, appeals and services that already exist, the answer shouldn't be to make those things difficult again.
It should be to build government systems capable of handling the public actually using them.
Read "Characterizing Agentic Flooding of Government Services."