AI is creating a strange new problem for the courts: it is making legal paperwork cheap enough that more people can file lawsuits without lawyers, but not necessarily clear enough — or accurate enough — for courts to process efficiently.
Recent research points to a sharp rise in self-represented "pro se" filings after ChatGPT-style tools became mainstream. Citing MIT and USC research, The Decoder reports that U.S. federal civil cases filed without a lawyer nearly doubled from the pre-AI average, with pro se filings reaching 16.8% of federal civil cases in fiscal year 2025. AI-generated text is now appearing in a meaningful share of complaints, and docket activity from self-represented plaintiffs has surged alongside it.
The bottleneck moved, but it did not disappear
That is the core tension. AI can widen access to justice — it can help someone who was previously priced out of the legal system put a complaint in front of a judge. That is genuinely valuable.
But it can also flood institutions with polished-looking filings that are overlong, legally incoherent, factually hallucinated, or all three. Courts still have to read every one of them, classify them, respond to them, and rule on them. Clerks and judges are not optional steps in that process.
The bottleneck shifted from "can someone draft the document?" to "can the system absorb the output?" Those are very different problems, and only one of them got easier.
Volume is not the same as quality
404 Media has covered specific cases where AI-assisted pro se filings cited nonexistent precedents, used fabricated case numbers, or included statutes that do not apply to the jurisdiction. Judges have had to issue sanctions and warnings. Some courts are now crafting standing orders specifically addressing AI use in filings.
This pattern — AI making it cheap to produce formal-looking output, while the receiving institution bears the cost of evaluating it — is not unique to courts. It is a preview of what happens across any system where document production was previously rate-limited by skill and effort.
Grant applications, appeals, compliance packets, insurance claims, academic submissions: wherever there is a formal intake process that requires human review, AI is about to test how much volume that process can absorb before it changes its rules or breaks down.
Access to justice and document overload are both real
It would be a mistake to frame this as purely negative. Many people who file pro se do so because they have a real legal claim and no realistic path to an attorney. AI making that option more accessible is a genuine improvement in legal equity for a segment of the population that has historically been locked out.
The problem is that the same tools are also being used by people filing nuisance suits, by those who misunderstand their legal situation, and by bad actors looking to use legal process as harassment. Courts cannot easily distinguish between these categories at intake. They have to process the whole stack.
The access-to-justice benefit and the paperwork-overload cost are arriving at the same time, through the same channel.
Why this matters for SunMarc
This is a clean illustration of a broader product design lesson: useful AI tools should not just generate documents. They need structure, source grounding, scope limits, review flows, and friction at the right moments.
The court-docket problem is what you get when a generation tool has no guardrails about accuracy, relevance, or institutional fit. The tool produces output confidently. The output looks professional. But the downstream cost — the clerk who has to file it, the judge who has to read it, the opposing party who has to respond — is entirely invisible to the person who hit "generate."
The winning products in the next few years will not be the ones that make it easiest to produce more output. They will be the ones that help users produce better outcomes. That means building in source grounding so citations are real, scope constraints so documents match the actual context, and review steps so a human has to confirm before something consequential goes out the door.
Friction is often framed as a failure of UX. In contexts where output quality matters — legal, medical, financial, compliance — friction is the feature. The product that slows you down at the right moment is the one that does not embarrass you or harm someone else.
The takeaway
AI lowered the cost of producing formal work faster than institutions could adapt to receiving more of it. Courts are the most visible current example, but the underlying pattern will show up everywhere.
For anyone building AI-assisted products: the question is not just "can we generate this?" It is "what happens downstream when we generate a lot of it?" The systems on the receiving end of your output — human reviewers, clerks, judges, compliance teams, counterparties — have finite capacity. Designing as if they do not is how you accidentally build a problem rather than a product.