A public database now tracks a growing set of court decisions worldwide in which judges have found (or strongly implied) that a lawyer or party filed a brief resting on hallucinated AI content, and the count keeps climbing. That is not a niche compliance footnote. It is a live docket any litigator can search by tool, jurisdiction, and outcome, and other courts are already citing it.
Read the opinions in that docket and a pattern emerges. Courts are not punishing the use of generative AI; they are punishing something narrower and more embarrassing, namely what the lawyer did (or failed to do) between the model's output and the moment of filing. The distinction matters, because it tells you exactly where the sanctions risk lives.
Fabricated Authorities That Slipped Past the Signature
The archetype is still Mata v. Avianca, the 2023 case that produced the first serious federal sanction for a brief built on nonexistent cases. The doctrinal basis was hardly novel. Rule 11 of the Federal Rules of Civil Procedure has long required that a signed filing be grounded in law that actually exists, and that is the whole point. Courts did not need a new rule to punish AI hallucinations; they already had one.
Everything since has followed the same logic. Fake case names, invented docket numbers, confident quotations that appear nowhere in the cited opinion, none of it is treated as a technology problem. It is treated as a candor problem, and the sanction attaches to the human who filed it.
Real Cases, Made-Up Quotes
A subtler failure mode has emerged, and it may be more dangerous than a wholly fabricated citation. The case exists, the volume and page number resolve, and the parties are real.
But the quoted language does not appear in the opinion, or it is stitched together from unrelated passages, or it says the opposite of what the court actually held.
These slip through more often than the pure fabrications because a bluebook check does not catch them. A clerk pulls the case, sees the reporter cite is valid, and moves on.
The misquote survives until opposing counsel, or the judge, actually reads the underlying decision. Then the sanction motion writes itself.
The Cover-Up Draws the Bigger Penalty
In June 2026, the Ninth Circuit sanctioned two immigration attorneys $2,500 each and suspended them from practice before the court for six months. The briefs contained citations to opinions that did not exist and fabricated quotes lifted into real cases.
What pushed the discipline harder was not the AI use. The panel was explicit that using generative AI is not itself sanctionable. The sequence that followed is what did the damage. The lawyers first characterized the errors as typographical. When pressed at oral argument, one denied AI use three times before conceding it was possible, and finally disclosed that an unlicensed drafter had produced the brief with no attorney verifying the authorities.
The court noted plainly that lighter sanctions would have been on the table if counsel had come clean the first time. Similar messages have shown up in other circuits' 2026 orders addressing the same fact pattern.
Repeat Offenders Draw Escalating Discipline
The last category is the one that will drive the harshest penalties going forward: lawyers sanctioned once, warned on the record, and then caught doing it again. Multiple 2026 orders have specifically referenced prior warnings, from the same court, from the state bar, from a supervising judge, as justification for stepping up from a monetary sanction to suspension or referral for discipline.
Numerous lawyers have landed in trouble for leaning on lawyers have landed in trouble for leaning on AI, and courts are less patient the second time. The pattern in these repeat cases is not defiance; it is inattention. The lawyer told the court they would verify going forward, adopted no actual procedure, and got caught by the same failure mode within months.
The Docket Points to a Short List of Controllable Habits
Read across the sanctioned opinions and a short list of controllable behaviors keeps appearing. None of them require sophisticated tooling.
- Pull every cited case yourself. Not the summary the model produced, not the headnote, the opinion itself. If a citation cannot be retrieved from an authoritative database, it does not go in the brief.
- Match quotes to the source text. A valid reporter cite is not a verification of the language attributed to it. Copy the quoted sentence into a search of the opinion itself.
- Disclose promptly when something breaks. The Ninth Circuit's opinion is explicit that early candor invites lighter sanctions. Silently swapping a bad citation for a good one, once you know the first was fabricated, is its own violation.
- Write down your firm's AI procedure. Many state bar associations have issued guidance instructing lawyers to understand the tools they use and verify all output. A written protocol is what a court will ask for after the fact.
The Signature Still Belongs to the Lawyer
Generative AI can genuinely compress the work of legal research. The docket does not contradict that. It reinforces a duty the profession has long recognized: the signature on the brief is a personal attestation, and no model in the loop transfers that duty to the tool.
The lawyers who understand this have kept using AI. The lawyers who did not are in the database.