In June 2023, a New York lawyer named Steven Schwartz became the cautionary tale of the AI era. He'd asked ChatGPT to find cases supporting his client's argument in Mata v. Avianca, and the chatbot obliged with six confident, well-formatted federal decisions. None of them existed. Judge P. Kevin Castel sanctioned Schwartz, his colleague Peter LoDuca, and their firm $5,000, and the legal world treated it as a freak accident — a warning that would surely scare everyone straight.
It did not. Three years later, the problem is not a handful of embarrassing anecdotes. It's a documented, daily, worldwide phenomenon.
The number that keeps climbing
Legal researcher Damien Charlotin maintains the most comprehensive public database of AI hallucination cases in law — court decisions where a judge found that someone submitted AI-fabricated content. As of July 5, 2026, it catalogs 1,725 cases, and it grows almost every day. On a single day in March 2026, U.S. courts issued 17 decisions flagging suspected AI hallucinations.
The geographic spread tells you this isn't a quirk of one court system:
| Jurisdiction | Cases |
|---|---|
| United States | 1,187 |
| Canada | 190 |
| Australia | 96 |
| United Kingdom | 59 |
| Israel | 54 |
The breakdown by who is doing it is the part that surprises people. Self-represented litigants account for 1,016 cases — more than lawyers. But lawyers are not far behind at 667, and the database now even records 24 cases involving judges whose rulings leaned on fabricated citations. This is not a problem confined to people who can't afford counsel.
What "hallucination" actually looks like
The database sorts the failures into categories, and the shape of the problem is consistent. Of the 1,725 cases, 1,447 involve fabricated material — most often case law that simply does not exist. Another 710 involve misrepresented sources (real cases twisted to say something they don't), and 466 involve false quotations attributed to real decisions.
The first "decision" included "gibberish" legal analysis and a "nonsensical" and internally inconsistent procedural history before it "abruptly" ended without a conclusion.
That was Judge Castel describing one of Schwartz's fake cases back in 2023. What's striking is how little the failure mode has changed. Large language models are pattern generators, not databases. Asked for a citation, they produce something that looks like a citation — a plausible name, a plausible reporter number, a plausible quote — with no mechanism to check whether the thing they invented is real. Schwartz even asked ChatGPT whether his cases were genuine; it assured him they "can be found in reputable legal databases such as LexisNexis and Westlaw." They could not.
The penalties are getting serious
For a while, courts were lenient — a warning here, a $5,000 fine there. That patience is running out.
In an opinion dated December 12, 2025, U.S. Magistrate Judge Mark D. Clarke of the District of Oregon called one case "a notorious outlier in both degree and volume" and imposed $110,000 in fines and fees against two attorneys — the largest AI-hallucination penalty ever handed down by an Oregon federal judge. In Couvrette v. Wisnovsky, a $12 million family dispute over an Oregon winery, the plaintiffs' lawyers filed three briefs containing 15 references to nonexistent cases and eight fabricated quotations.
The aggravating detail wasn't the AI use itself — it was the cover-up. After opposing counsel flagged the fakes, lead attorney Stephen Brigandi quietly stripped the false citations and refiled, without verifying the arguments that remained. Clarke ordered Brigandi to pay $80,000 in attorney fees plus roughly $15,000 in fines, ordered co-counsel Tim Murphy to pay $14,000, and dismissed the case with prejudice. As the judge put it: "If there was ever an 'appropriate case' to grant terminating sanctions for the misuse of artificial intelligence, this is it."
The courts are institutionalizing distrust
The judiciary's response has moved from case-by-case scolding to structural rules. By early 2026, at least 25 federal district courts had adopted standing orders requiring attorneys to certify whether AI was used to prepare a filing and to confirm that a human reviewed every citation. Charlotin has even built an automated reference checker, PelAIkan, trained on the database to catch fabricated citations before they reach a judge.
The irony is sharp. The same technology being sold to lawyers as a research accelerator is generating a parallel body of law about how not to use it — a self-documenting archive of professional negligence.
The Bottom Line
The lesson of Mata v. Avianca was never "don't use AI." It was "verify what it tells you," and 1,725 cases later that lesson clearly hasn't landed. AI is genuinely useful for drafting, summarizing, and brainstorming — but it does not know what is true, and it will manufacture a citation as readily as it writes a sentence. Every fabricated case in Charlotin's database traces back to the same missing step: a human who trusted the output without opening a real reporter to check it. Until that step becomes reflexive, the counter will keep climbing — and the fines will keep growing.
This piece discusses professional and legal consequences of AI misuse; it is not legal advice.
