Courts are learning that access to justice has a new problem: AI can help anyone sound like a lawyer, even when the law it cites does not exist.
The legal system has spent years trying to make room for people who cannot afford attorneys. Now those same open doors are meeting a tool that can produce a motion, complaint or appeal in minutes, polished enough to look serious and flawed enough to waste days of court time.
As Bloomberg recently reported, self-represented litigants are increasingly arriving with AI-assisted filings that consume more attention from clerks, judges and opposing lawyers. That does not mean every pro se filing is suspect. It does mean courts are facing a new kind of volume: documents that read as if they were professionally drafted, but may rest on fake cases, invented quotes or legal arguments that collapse when checked.
This is not just a courthouse nuisance. It is a business story for every startup selling AI into professional services. Legal drafting tools promise speed, affordability and broader access. Those are real benefits. But when a product helps a non-lawyer produce court-ready language without reliable verification, the risk moves downstream to judges, defendants, court staff and the user who may think the software has done more than it actually has.
The pattern is now familiar. A litigant submits a filing with citations that look normal. The opposing side or the judge tries to find the authority. The case does not exist, or the real case says something different. By then, everyone has spent time chasing a legal mirage.
Clio, citing research from legal hallucination tracker Damien Charlotin, recently put the documented scale at about 1,400 AI hallucination cases globally, including more than 955 in the United States. It also noted that those cases sit inside roughly 40 million U.S. court cases filed since January 2023, so the raw percentage remains small. That distinction matters. The problem is not that AI has ruined every docket. The problem is that even a small number of defective filings can be expensive when courts must read them, respond to them and protect the rights of everyone involved.
The Illinois courts described the same concern in January, pointing to more than 280 filings since 2023 that included hallucinated citations and saying the 2025 number had risen sharply. The warning was aimed at lawyers, but the pressure is broader. Judges are now seeing AI errors from attorneys, self-represented parties and even court-adjacent workflows where staff or clerks may be tempted to lean on generative tools.
The Seventh Circuit made the point plainly in January in Jones v. Kankakee County Sheriff's Department, where it cautioned that even pro se litigants cannot rely on unverified AI output. That is important because self-representation has always carried some procedural patience from courts. AI tests the limit of that patience by making weak arguments look formal enough to require a fuller response.
Disclosure is becoming the first regulatory answer
The newest move is not a ban. It is disclosure. Miami-Dade and Broward courts issued coordinated requirements in May 2026 that attorneys and self-represented litigants must disclose generative AI use in court filings and verify the accuracy of AI-assisted material. The Eleventh Judicial Circuit said the goal was consistency in how courts handle AI-assisted submissions.
That is a practical response because courts do not need to know whether a user opened ChatGPT for brainstorming. They need to know whether a filing contains AI-drafted legal analysis, quotations, citations or factual assertions that must be checked before they become part of the record. The certification language matters because it puts responsibility back on the human filer.
For legal-tech companies, this is where the market gets more serious. A tool that drafts a complaint is not the same as a tool that verifies every cited case against an authoritative database. A chatbot that can explain civil procedure is not the same as software that understands local rules, filing limits and jurisdiction-specific standards. The startups that treat those differences as product details may find themselves pulled into a trust problem they cannot solve with marketing.
The better opportunity is obvious. Vendors that build citation checking, jurisdiction controls, audit trails and plain warnings into the workflow can sell reliability, not just speed. That is a stronger business than giving frustrated litigants a prettier way to file bad law. It also fits where regulation is moving, toward human responsibility, disclosure and documented verification.
There is still a hard balance here. AI can help people understand forms, organize facts and prepare cleaner drafts. For people priced out of legal help, that can be meaningful. But courts are not customer support desks for software output, and defendants should not have to pay lawyers to disprove imaginary precedent.
The next phase will likely look less like a dramatic crackdown and more like a paperwork standard that spreads. Courts will ask who used AI, what it touched and who checked it. Other regulated industries will be watching closely, because law is becoming an early test of a larger rule: when AI performs professional work, someone still has to sign their name to the result.
Also read: A quantum AI test on IBM hardware points to a new compute race • Japan's cablemaker selloff tests the AI infrastructure trade • Wall Street banks are paying AI experts $25,000 a day
