AI hallucination court sanctions: the record since Mata v. Avianca

The short version

• An AI “hallucination” is a confident, made-up output. Here it means a case citation, quote or holding that the tool invented.
• Since Mata v. Avianca in 2023, courts in the United States, the United Kingdom and Canada have imposed AI hallucination court sanctions on the lawyers who filed them.
• Orders run from wasted costs and fines of a few thousand dollars to disciplinary referrals, file reviews and published warnings.
• Courts do not require bad faith. The United States Rule 11 duty reaches a careless filing, and a signer cannot hand the accuracy of a citation to AI, a junior lawyer or a client.
• The rule is the same across these courts: whoever signs owns every citation, and the duties that already bind a lawyer cover AI use.^1,2,3,4,5

This report reads each court’s own order. For the wider picture, see AI in dispute resolution; for the professional duties behind these orders, agentic AI in dispute resolution; and for how the law treats AI more broadly, AI in law.

What is an AI hallucination in a court filing?

A hallucination is a confident, well-formatted output that is false. In a filing it takes three main forms: a case that does not exist (a phantom or fabricated citation), a real case with a quote it never contained, or a real case given a holding it never reached.^3,4

The pattern recurs in every order below. Early cases were wholly invented opinions with fake quotes and citations. More recent orders describe false quotes and misstated holdings attributed to real cases, which are harder to spot. The duties the courts say this breaches are not new: the duty of competence, and the duty of candour to the tribunal, the obligation not to mislead the court. What the tools do, and why they produce this, is set out in AI legal tech.^1,3,4

What happened in Mata v. Avianca?

In Mata v. Avianca, Inc. a New York federal court fined two lawyers and their firm US$5,000 for filing a brief built on cases ChatGPT invented (22 June 2023). The brief cited at least six non-existent opinions with fake quotes and citations. What moved the court was less the first mistake than the doubling down: the lawyers stood by the fake cases after the court questioned whether they existed, and only began to correct the record once the court had ordered them to explain why they should not be sanctioned (an order to show cause). The court sanctioned them under Rule 11 and, in the alternative, under its inherent power, the court’s own authority to control its proceedings, and it declined to add a separate sanction under 28 U.S.C. § 1927.¹

The court ordered the US$5,000 penalty paid into the court registry within 14 days, and ordered the lawyers to send letters to their client and to each real judge falsely named as the author of a fake opinion. It said “existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings,” while adding that there is “nothing inherently improper about using a reliable artificial intelligence tool for assistance.”¹

Which courts have imposed AI hallucination court sanctions, and what did they order?

Courts in at least three countries, the United States, the United Kingdom and Canada, have acted, each under its own powers. The table below traces each named case to its own order.^1,2,3,4,5,6

Arbitrators are not exempt either. The arbitration-specific rulings on AI, including the Québec case ARIHQ c. Santé Québec, 2026 QCCS 1360, are collected in AI arbitration case law.⁷

How much are the penalties, and what else do courts order?

The money is usually modest, a few thousand dollars, because most of these orders aim to deter rather than to punish. The non-monetary terms often bite harder. Mata was US$5,000, joint and several, plus the letters. Coomer was US$3,000 on lead counsel and his firm and US$3,000 on co-counsel, which the court called the “least severe sanction adequate to deter and punish” here. Noland was US$10,000, plus publication “as a warning” and a referral to the State Bar.^1,4,5

These other terms are the common thread. In Zhang the lawyer faced personal costs, with the amount left to the Registrar, and had to review all her other court files for AI content within 30 days. In Ayinde the wasted costs, a costs order the lawyer pays personally, were £2,000 each, with referrals to two regulators. Across the orders the recurring menu is letters to the client and to the judges falsely named as authors, file reviews, disciplinary referral, and publication of the judgment, with striking the offending paper among the further powers the courts list but did not impose here.^1,2,3,5

There is no official count. A public database of AI hallucination cases listed more than 2,200 court and tribunal decisions worldwide as of 10 November 2026, and it is still growing. That figure is a tracker’s own tally, not an official statistic, and we checked every case named in this report against its own order.⁸

On who gets caught, the same database records both represented and self-represented filers, alongside high-profile matters such as Coomer v. Lindell, the My Pillow defamation case. The duty to check falls on anyone who signs, so a self-represented (pro se) litigant can be sanctioned as well, each case to be confirmed on its own order.^4,8

Do courts have to find that a lawyer acted in bad faith?

No. The duty reaches carelessness, not only dishonesty. In the United States, Rule 11 certifies that a signer read the filing, made a reasonable inquiry and that its legal contentions are warranted. In Coomer the court quoted Mata that “a fake opinion is not ‘existing law,’” and held that running a draft through AI or delegating the cite-checking to co-counsel does not satisfy the duty, which it called non-delegable.^1,4

In Canada, the threshold differs by the costs sought. Special costs need deliberate, reprehensible conduct, and in Zhang the court denied them, accepting that the lawyer did not intend to deceive. Ordinary personal costs needed only that the lawyer caused avoidable cost, and the court granted those. In England, contempt needs more: as the Divisional Court put it, “There must be knowledge that it is false, or a lack of an honest belief that it is true,” and “mere negligence as to the falsity of the material is insufficient.” That is why the court in Ayinde found the contempt threshold met in one case but did not pursue it, while wasted costs and regulator referrals still followed.^2,3

What should you do before you file?

You own the accuracy of every citation in anything you sign, and you cannot pass that duty to AI, a junior lawyer or your client. In England and Wales, the duty to check AI research against an authoritative source rests on a lawyer who uses AI “to conduct research themselves or rely on the work of others who have done so.” In the United States, the Rule 11 inquiry is non-delegable and continues while you keep advocating a point. In Canada, citing fake cases “is an abuse of process and is tantamount to making a false statement to the court.” The practical step the orders point to is the same one, and it is cheap next to the sanction: before you file, read every case the draft cites and confirm, against an authoritative source such as the official law reports or a reputable legal database, that it exists, says what it is cited for, and has not been overruled.^2,3,4

The direction of travel is open, not settled. In Ayinde the Divisional Court sent its judgment to the Bar Council, the Law Society and the Council of the Inns of Court, asking them to decide quickly what more they should do. That is an invitation, not a rule. The full professional-duty map, competence, candour and supervision, is in agentic AI in dispute resolution; the evidence-authentication cousin, deepfakes (convincing AI-faked images, audio or video) and other AI-generated exhibits, is in AI-generated evidence and deepfakes.³

What does this mean for anyone using AI in a dispute?

In our view, the orders since Mata v. Avianca point one way, whatever the jurisdiction. A named human stays responsible for what is filed. Courts in New York, Vancouver, London and Colorado put the duty on the person who signed, not on the tool. AI use that touches the record has to be open to checking, which is why these courts reach for verification against an authoritative source rather than a ban. And the output of any tool has to be verified before anyone relies on it. In the words of one order, “a fake opinion is not ‘existing law.’” The pace of AI is now ahead of the rulebooks, so the safest reading is that the existing duties of competence, candour, supervision and reasonable inquiry already govern AI use, and the newer citation rules only make that explicit. None of this is new law. It is the duty every filer already had, now applied to AI.^1,2,3,4