Photo by Brett Jordan on Unsplash
- As of August 17, 2026, the anchor case remains Mata v. Avianca Airlines: in May 2023, attorneys Steven Schwartz and Peter LoDuca were sanctioned $5,000 by Judge P. Kevin Castel after filing briefs built on AI-generated citations that did not exist.
- Studies of AI legal tools show hallucinated cases in 15-30% of queries, depending on the model and how complex the question is. Run that against a 20-citation brief and you get three to six invented authorities.
- At least 12 documented instances of lawyers sanctioned or disciplined over AI-fabricated citations had accumulated as of mid-2024, with penalties ranging from $5,000 fines up to bar disciplinary referrals.
- Courts in Texas, Missouri and the federal system now require certification that AI-generated content was checked. The duty is on the lawyer — and the delay, the refiling and the lost credibility land on the client.
What's on the Table
It is a Monday morning in August 2026, and a paralegal is doing something that did not exist as a job task four years ago: pulling every case cited in a draft brief and confirming, one by one, that the case is real. Not that it says what the brief claims. That it exists. According to Google News, which surfaced SmartCompany's reporting on the growing pile of AI-assisted filings hitting court dockets, that verification step has quietly become the most valuable half-hour in the building.
The reason traces back to a single filing. In May 2023, in Mata v. Avianca Airlines, lawyers Steven Schwartz and Peter LoDuca submitted material containing at least six fabricated cases — names like "Varghese v. China Southern Airlines" and "Shaboon v. Egyptair," complete with citations and quoted reasoning. None were real. Judge P. Kevin Castel sanctioned the two attorneys $5,000. What made the episode instructive was not the size of the fine but the texture of the fakes: they read exactly like law. Correct reporter format. Plausible circuit. Judicial tone.
In plain terms, a large language model is a very good guesser about what a sentence should look like next. It is not a database of decided cases. As legal technology specialists have repeated since 2023, "generative AI is not a search engine" — and that single distinction explains nearly every sanction that followed.
The Math Nobody Runs: What a 15% Error Rate Does to One Brief
Here is the part the coverage tends to skip. A hallucination rate sounds like a technical footnote until you apply it to a document.
Take an ordinary motion with 20 cited authorities — unremarkable for a summary judgment brief. At the bottom of the documented 15-30% range, that is three invented cases (20 × 0.15). At the top, six (20 × 0.30). Six is not a hypothetical number: it is the count of fabricated cases in the Avianca filings. A shorter 10-citation motion fares better in absolute terms but no better in kind — one or two ghosts at the low end, three at the high end.
Chart: Fabricated citations a brief would contain if every authority came from an unverified AI legal tool, applying the documented 15-30% hallucination range (studies cited in reporting current as of August 17, 2026) to briefs of 10 and 20 citations. Arithmetic is ours; the underlying rate range is from the research.
One more piece of arithmetic, and it is ours rather than the court's: the $5,000 Avianca sanction spread across six fabricated cases works out to roughly $833 per invented citation. That is the cheap part of the bill. The expensive part is the refiling, the credibility hit with the judge who now reads every future submission from that firm with a raised eyebrow, and — for the client — a case that lost months on a detour it never asked for.
A careful skeptic pushes back here, and fairly: those hallucination rates come from general-purpose chatbots, and the serious legal software vendors ground their tools in real case databases. True. But grounding narrows the failure mode rather than closing it — a tool can retrieve a genuine case and still mischaracterize its holding, which is a subtler error than an invented case name and much harder for a rushed reviewer to catch. The verification question does not disappear when the citation turns out to be real.
The Rules That Already Govern This
There is no need for a new AI statute to make this actionable. The duty of competence and the duty of candor to the court already cover it: a lawyer who files unverified AI-generated research is, under existing professional conduct rules, making a representation they have not checked. Bar associations have been blunt that responsibility does not transfer to the software — the technology cannot be blamed for the error.
Layered on top, courts in Texas, Missouri and across the federal system have issued standing orders requiring counsel to certify that any AI-generated content has been verified for accuracy. Florida and California have both published formal ethics opinions on generative AI in practice, and both emphasize verification. A court would likely look first at whether that certification was signed and whether anyone actually did the work behind it.
Which Fits Your Situation: Three Questions Before You Sign
If you are hiring a lawyer, you are not in a position to audit their research process — but you are in a position to ask three specific things, and the quality of the answers tells you a great deal. This is the same verification instinct Ai Agents applied to vendor AI-security claims: the question is never whether AI was used, it is what was independently checked.
Not "do you use AI" — the useful question is who verifies citations and at what stage. Major firms have moved to mandatory AI training and verification protocols for every attorney using generative tools. A firm that can describe its process in two sentences has one. A firm that gets defensive probably does not.
Your lawyer should know this without looking it up, because the certification requirement in Texas, Missouri and various federal courts is theirs to sign. If your matter is in one of those jurisdictions, the answer should be immediate.
Pick a single case cited in a filing you receive and search the name in a free public database. You are not evaluating the legal argument. You are confirming the case exists. It takes about ninety seconds and it is the exact step that was skipped in Avianca.
Where This Goes Next
The tempting read is that AI is bad for law. Our analysis points somewhere narrower: the failures cluster almost entirely at the drafting-and-citing stage, where a model is asked to produce authority it cannot possess. In contract review — summarizing clauses in a document the tool has actually been given — legal technology has a much cleaner record, because the source text is sitting right there. The same is true of law firm automation aimed at intake, scheduling and document assembly. On balance, the more likely outcome is not retreat from AI legal tools but a hardening split: heavy adoption where the tool reads a document you supplied, and mandatory human verification where it claims to know something about the world.
The at-least-12 documented sanction cases logged by mid-2024 are best understood as an early, cheap warning rather than the peak. Judges are now primed to look, and the second offense at the same firm will not draw a $5,000 fine.
Frequently Asked Questions
Can I sue my lawyer if they filed fake AI-generated case citations in my case?
Sanctions in these matters have run from $5,000 fines to referrals for bar disciplinary proceedings, and disciplinary complaints to the state bar are a separate track from any civil claim. Whether a filing error rises to malpractice depends on the harm it caused in your specific matter and your jurisdiction's standard — that is a question for an independent attorney reviewing the file, not a general rule.
How do I check whether a case citation my lawyer used is real?
Search the case name and citation in a free public source such as a court's own opinion database or a public legal research site. If nothing returns, that is a flag worth raising — the fabricated citations in the Avianca filings looked entirely legitimate on the page and only failed when someone tried to pull the actual opinion.
Are law firms required to tell clients when they use AI legal tools?
Disclosure obligations vary by jurisdiction. Several courts, including in Texas and Missouri and parts of the federal system, require lawyers to certify to the court that AI-generated content has been verified, and Florida and California have issued formal ethics opinions addressing generative AI use. Client-facing disclosure is a separate question governed by your state's rules and your engagement letter — worth reading before you sign.
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute legal advice. No independent testing of any legal software or AI product was conducted for this piece. Rules on AI use in court filings vary by jurisdiction; consult counsel licensed in yours for guidance on your specific matter. Research based on publicly available sources current as of August 17, 2026.