A number most people haven't seen yet
There is a public database, maintained by legal researcher Damien Charlotin, that does one quiet, useful thing: it tracks every court case where someone got caught submitting AI-hallucinated legal citations. As of early July 2026 it held 1,668 cases. It updates daily, so by the time you read this the number is higher.
That total alone is easy to wave away, because a lot of it is self-represented people who leaned on a chatbot. The part that should stop you is narrower. In 653 of those cases, the responsible party was a practicing lawyer, not a layperson. The United States accounts for 1,163 of the total. This is not a fringe story about people misusing free tools. It is happening to licensed attorneys, in real filings, in front of real judges.
Sources for every number here are linked at the bottom. This is a data piece, not a sales pitch, and I would rather you check the primary records than take my word for it.
The penalties stopped being trivial
Early on, the sanctions were embarrassing but survivable, small fines and a stern order. That has changed. Penalties now run into five figures per attorney, and they have moved past money.
The widely cited founding case is Mata v. Avianca (2023), where a $5,000 sanction was imposed after a lawyer filed a brief full of cases that did not exist. Since then it has escalated. Attorneys have been fined $15,000 each in a federal appeals matter. There have been bar referrals, filings struck from the record, and the first suspensions tied specifically to AI-fabricated citations. In Withers v. City of Aberdeen (June 2026), a federal judge in Mississippi suspended two lead attorneys from practicing in that district for two years after both sides filed hallucinated citations.
A two-year suspension is not an embarrassment you apologize for and move past. It is a hole in a career.
Why this keeps happening to careful people
The instinct is to assume these are careless lawyers cutting corners. Some are. But the more uncomfortable finding is that the tools people trust to be safe are not as safe as they sound.
Stanford's RegLab tested the purpose-built legal research tools, the paid ones marketed to law firms as reliable. General-purpose chatbots hallucinated on legal queries somewhere between 69% and 88% of the time, which most lawyers already suspect. But the dedicated legal tools, from LexisNexis and Thomson Reuters, still hallucinated between 17% and 33% of the time. Westlaw's AI-Assisted Research came in around 33%. These are the products sold specifically to prevent this problem, and one in three to one in six answers still contained a fabrication or a misgrounded citation.
That is the real trap. A lawyer using a tool that advertises reliability, on a deadline, reasonably assumes the citation it produced is real. Most of the time it is. The times it is not are the times that end up in the tracker.
The people most exposed are the ones with the least backup
The ABA's 2025 Legal Technology Survey found that AI adoption among lawyers has climbed to about 30% overall, but among solo practitioners it sits at 18%, well below larger firms. The stated reason for the hesitation is consistent: accuracy and reliability are the number one barrier to using AI at all.
That caution is rational. A solo or two-person firm does not have a research librarian, a second associate to double-check citations, or a paid citator subscription running in the background. When the AI invents a case, there is often no one between the draft and the filing. The lawyers most worried about accuracy are also the ones with the fewest safety nets, which is exactly why this hits them hardest when it hits.
What actually reduces the risk
None of this means avoid AI. It means treat its output the way you would treat a bright new intern's memo: useful, fast, and not filed until someone verifies it. A few habits that hold up, drawn from what the sanctioned cases have in common:
Verify every citation before it goes in a filing. Not a sample, every one. The cases in the tracker are overwhelmingly ones where a real-looking citation was never checked against the actual record.
Do not trust a single tool's own reliability claim. The Stanford data is the argument here: even the tools built for this hallucinate. Cross-check against a source the tool did not generate.
Keep the manual step. A citator check, or at minimum confirming the case exists and the quote appears in the actual opinion, is the specific step that catches the failure mode driving these sanctions. There are free ways to do the existence-and-quote check, for example searching the opinion directly on a public database like CourtListener, before you pay for anything.
The pattern in every sanctioned case is the same: the verification step got skipped. The fix is not a smarter AI. It is refusing to file anything the AI produced until a human confirms it is real.
Where this leaves the rest of us
This is a legal story, but the lesson generalizes to anyone using AI for work that matters. The tools are genuinely useful and genuinely fallible in the same breath, and the failures are quiet. They look exactly like the successes until someone checks.
If you want the underlying data, Damien Charlotin's tracker is public and worth a look. Full credit to that dataset, which is the source for the case counts and sanction figures above. I write about testing and verifying AI output at CLSkills, and if that is your world, you can get on the list here for the free 75-page guide and future write-ups like this one. No pitch attached to this piece either way, the data stands on its own.
Sources: Damien Charlotin's AI Hallucination Cases Database (case counts and sanction figures); Stanford RegLab, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools"; ABA 2025 Legal Technology Survey coverage, LawSites.