The first wave of stories about lawyers filing AI-invented case citations was treated as a curiosity. It is no longer a curiosity. It is a body of case law with patterns in it.
The most complete public record is the AI Hallucination Cases database maintained by researcher Damien Charlotin, which tracks court decisions worldwide where AI-generated fabricated content was identified. As of 1 September 2026 it holds 2,006 cases, of which 1,376 are from courts in the United States and 630 from the rest of the world.
Two thousand cases is enough to stop treating this as a story about careless individuals and start treating it as a workflow problem.
What the recent sanctions actually look like
The penalties have settled into a recognisable band. From August 2026 alone, the database records a $8,000 sanction plus a continuing legal education requirement in a Northern District of Georgia matter, a $5,000 fine with a contempt finding and removal from the case before the New Mexico Supreme Court, a $1,500 monetary sanction with an adverse costs order and a bar referral from a Florida appellate court, and a formal public reprimand in the District of Arizona.
Read together, three things stand out.
First, the money is not the punishment. A $5,000 fine is survivable. A published opinion with your name on it, a bar referral, and removal from the matter are not, and those now appear routinely alongside the fine.
Second, courts have stopped accepting inexperience as mitigation. The early opinions in 2023 read as bewildered. The 2026 opinions read as impatient.
Third, the exposure is not limited to represented parties. Pro se litigants are being sanctioned too, and appellate courts are referring the matter onward.
Why competent lawyers keep filing fabricated citations
The instinct is to say that anyone who files an unread citation deserves what they get. That is true and it is not useful, because it does not explain why it keeps happening to people who are not careless.
Four mechanisms show up repeatedly.
The citation looks right. Fabricated citations are not obviously wrong. They have plausible party names, a plausible reporter, a plausible year, and often a real court. They are designed by the underlying system to look exactly like the thing they are imitating.
The verification step is the one under time pressure. Nobody skips checking a case because they think checking is optional. They skip it because the filing is due at five and the drafting took longer than expected.
Delegation hides the origin. A significant share of reported cases involve work produced by someone junior, or by contract support, and signed by someone senior who did not know a generative tool had been used. The signature carries the responsibility regardless.
General-purpose tools get used for legal research. A chatbot with no connection to a case law database will produce a citation-shaped answer because that is what it was asked for. Research tools grounded in licensed content behave differently, which is a large part of why grounded legal research platforms exist as a separate product category.
The controls that actually prevent this
The advice to “check your citations” is correct and has visibly not worked. What works is making the check structural rather than aspirational.
- Separate the drafting tool from the verification tool. Every authority cited in a filing is pulled up independently in a primary law database before the document goes out. Not summarised, pulled up. The person doing the check should see the case.
- Make the check someone’s named job. A cite-check that belongs to everyone belongs to nobody. On any filing, one named person signs off that every authority was opened and read.
- Ask the disclosure question at intake of the work, not at the end. “Was any generative tool used in producing this?” asked when work comes back from a junior or a contractor surfaces the risk while there is still time.
- Check the court’s standing order. A growing number of judges now require certification about AI use in filings. This varies by judge, not by jurisdiction, and it is checked at the filing stage.
- Log it. If a tool was used, record which one and for what. When a court asks, the firms that answer well are the ones that can answer specifically.
The supervision point
The ethics analysis is not novel. ABA Formal Opinion 512, issued in July 2024, grounds the duty in competence under Model Rule 1.1, which requires understanding the benefits and risks of the technology being used. A lawyer who does not know that a general-purpose model will invent citations does not meet that standard, and the opinion has been available long enough that not knowing is no longer a defence.
The governance gap remains the real problem. Survey work in 2026 found that 54 per cent of legal professionals had received no training on responsible AI use, and only 9 per cent worked somewhere with a written policy that was actually enforced. Adoption of these tools is now near universal. Supervision of them is not.
That gap is the whole story. Two thousand cases is what it looks like from the outside.
If your firm does not yet have a written position, see our practical structure for a law firm AI policy.
Where the legal industry reads first.
Enjoyed this article? Get the biggest legal industry updates, deals, appointments, insights and expert interviews in your inbox, free.
No spam. Unsubscribe anytime.
