AI Is Killing Court Cases Across America | AI & Law

Self-represented Americans now make up nearly one in five federal civil cases. The chatbot in their hand is the tool. The thing that put them there alone is the system.

Donald Sauve is 69. He lives in Minnesota. Last year, he tried to sue his ex-wife, her divorce lawyer, and the state judge who’d ruled against him. He handwrote the complaint himself and filed it pro se. A federal court dismissed it within a month for lack of jurisdiction.

Three months later, he was back. This time, the filing was neatly typed. Fifty additional documents had been added, including something he called a “Case Law Synthesis,” a survey of precedents that supposedly supported his claim. ChatGPT and Claude had helped him write it. U.S. District Judge Patrick J. Schiltz dismissed the whole thing again in a 14-page ruling.

A few states over, in Las Vegas, a tenant named Nicole Silverberg was suing her former landlord pro se in federal court. She didn’t have a lawyer either. The cost was the problem. She turned to ChatGPT and Grok, and according to Reuters reporting and follow-up coverage in Inc., a federal judge called her out for citing cases that didn’t exist. “I didn’t know AI could be wrong at first,” Silverberg told Reuters. “It was like telling me Santa Claus was not real.”

These stories are easy to read as cautionary tales about AI. They’re not. They’re cautionary tales about a legal system most Americans can no longer afford to enter the front door of. AI is the tool people are reaching for. The question worth asking is why they had nothing else to reach for in the first place.

The Number That Matters

In March, two researchers at MIT and USC, Anand Shah and Joshua Levy, released a working paper that drew on administrative records from more than 4.5 million non-prisoner federal civil cases and 46 million PACER docket entries. Their finding: the share of federal civil cases filed pro se, meaning by someone representing themselves without a lawyer, jumped from a steady long-term average of 11 percent to 16.8 percent in fiscal year 2025. Their sample of recent filings showed AI-generated text in roughly 18 percent of complaints by early 2026, up from essentially zero before ChatGPT was released. Shah’s reaction to Fast Company captured the discovery: “We were just floored.”

Pro se cases now generate 158 percent more docket entries in their first 180 days than they did before AI was widely available. The system is being asked to read, docket, and respond to filings at a volume it wasn’t designed for, often from people who don’t know what a procedural motion is.

You could blame the chatbots. Most coverage has. But over 90 percent of all U.S. civil cases are handled in state courts, which means the federal numbers almost certainly understate what’s happening. And the trend has a more obvious cause than the technology that made it visible.

What People Are Walking Away From

In 2022, the Legal Services Corporation, the federally funded nonprofit that supports civil legal aid programs, published the most recent edition of its national Justice Gap Study. The headline: low-income Americans receive inadequate or no legal help for 92 percent of their substantial civil legal problems. That was up from 86 percent in 2017. The gap isn’t closing. It’s widening.

Of those who didn’t seek help when they had a problem, 46 percent cited cost. More than half said they didn’t know whether they could find and afford a lawyer if they needed one. LSC-funded organizations have to turn away roughly half of all eligible requests they receive due to limited resources.

California isn’t a refuge from this. The State Bar’s 2019 Justice Gap Study, surveying Californians across all income levels, found that 55 percent had experienced at least one civil legal problem in the past year, and nearly 70 percent of them received no legal assistance. 

As Stanford law professors David Engstrom and Lucy Ricca put it last year, the state has more than 250,000 lawyers and a justice system its own residents can’t access: “Imagine if the medical system allowed only doctors to provide all of the needed medical services, including drawing blood, stitching a finger, or doing an ultrasound. The system would collapse. This is what we are seeing in law.”

In 2022, the California Legislature shut down the State Bar’s working groups on regulatory reform and paraprofessional licensing, which had been studying ways to let non-lawyers handle simpler legal tasks. AI didn’t create the access vacuum in California. The legislature foreclosed the most promising alternative path to closing it, and then a free chatbot showed up.

The New Numbers Are Worse Than the Old Ones

In 2024, Stanford researchers were reporting that general-purpose AI hallucinated on 58 to 88 percent of legal questions. That was bad, but it was bad in a useful way. When AI is obviously wrong more than half the time, you learn to double-check.

By October 2025, benchmark testing from Vals AI found ChatGPT scoring 80 percent on a structured legal research test, beating the 71 percent baseline of human lawyers given the same questions. The hallucination problem didn’t disappear. It got more dangerous because the wrong 20 percent is now wrapped in confident, fluent, plausible prose. People stopped double-checking.

Damien Charlotin, a researcher at HEC Paris who tracks AI-hallucination court decisions worldwide, has documented over 1,400 cases in roughly three years, the vast majority of them in the United States. The single most underused statistic in this whole debate is the one from his database: three in five of those cases came from pro se litigants. Not lawyers showing off. People who couldn’t afford lawyers.

In Georgia, U.S. District Judge Marc Treadwell has ordered self-represented plaintiffs to disclose any AI use and attest to the accuracy of their submissions. Other federal judges across the country have issued similar standing orders. 

Bloomberg Law has documented at least 24 pro se litigants hit with monetary sanctions for AI-related court filings since the second half of 2023, with more than half of those fines coming after December 2025. A pro se ex-husband in Illinois was sanctioned over $3,000 in April after citing nine entirely nonexistent cases in his appellate brief. The court called it “a pervasive, nascent, and very serious issue afflicting courts.”

The defense the legal system is offering, that pro se filers need to verify everything before submitting, assumes those filers have the time, training, and database access to do what trained associates do. They don’t. That’s why they’re pro se in the first place.

The Counter-Story Nobody Tells

Not every pro se AI story ends in sanctions. In March, a federal magistrate judge in Colorado, Maritza Dominguez Braswell, ruled that a pro se employment discrimination plaintiff who had used AI to prepare his case against a major defense contractor was entitled to the same work-product protections an attorney would have received. 

The litigant, Archie Morgan, didn’t have to hand over the AI-generated drafts to the other side. Judge Braswell drew on the Supreme Court’s reasoning in Carpenter v. United States to hold that routing information through a third-party system doesn’t strip it of all privacy.

It’s the cleanest pro-access ruling on AI in the federal courts so far. It also requires Morgan to disclose what tool he used, and bars the use of consumer ChatGPT, Claude, or Gemini for confidential information going forward. The protection is real. The conditions are real. Both can be true.

This is the part nobody covers when they’re writing about AI as a courtroom villain: it has, in measurable ways, given some self-represented people a foothold they wouldn’t have had otherwise. The story isn’t AI good or AI bad. The story is that AI is filling a vacuum, and the vacuum is the actual problem.

Where California Is Actually Going

In January 2025, L.A. County’s tenant right-to-counsel ordinance took effect, the first of its kind in Southern California. The City of Los Angeles followed in April. The data that motivated the policy was stark: before passage, only 3 percent of L.A. tenants in eviction court had legal counsel, compared to 88 percent of landlords. That gap is closing for eviction defendants in those jurisdictions. It is not closing for anyone else, in any other kind of civil case, anywhere else in the state.

The argument for expanding the right to counsel is the argument against treating the AI problem as a technology problem. People aren’t downloading ChatGPT to be reckless. They’re downloading it because the alternative is showing up in court with nothing.

If You Have to Use AI, Use It This Way

This is the practical part, and it has to be practical because the policy answer is years away. If you are dealing with a legal problem in California and you’ve decided AI is your best available option, here is the bare minimum.

Verify every case citation. Search the case name and citation directly in Google Scholar Case Law and on Justia. If the case doesn’t appear in either, it almost certainly doesn’t exist. According to Charlotin’s database, fabricated citations are the single most common AI failure mode in court filings, and they’re the easiest to catch if you check.

Verify the statute. California codes are updated constantly. Pull every code section the AI cited and confirm it against leginfo.legislature.ca.gov, the official state legislative repository. Chatbots are often months or years behind on legislative changes.

Stop at the courthouse before you file. Every California superior court has a self-help center. They cannot give legal advice, but they can confirm whether a document is procedurally correct, which is a different question. Start at selfhelp.courts.ca.gov to find yours.

Be honest with the court about AI use. Disclosure orders are spreading, sanctions for hiding AI use are getting larger, and lying about it is a faster path to losing your case than the underlying mistakes.

None of this is enough. A pro se eviction defendant verifying citations on Google Scholar at 11 p.m. is doing the work that a trained associate with a paid database would do in a quarter of the time. The patch is not the same as a fix. But until the fix arrives, the patch is what people have. And the patch works better than nothing, which is what too many Californians are walking into court with today.

The chatbot is not the villain in this story. It’s the warning light. It is a symptom of a justice system that has long left most Americans to navigate legal challenges on their own. AI has simply made that reality harder to ignore.