AI Regulation Tracker  /  Court ruling

The Tenth Circuit put its AI citation standard in a published opinion, then declined to sanction

Lichfield v. Kubler, No. 25-4135, filed July 27, 2026 and marked PUBLISH. A brief citing an authority that does not exist failed the reasonable-inquiry certification. The lawyer paid nothing. The reasoning is now in the reported volumes anyway.

By Anthony Guerriero  ·  Published July 28, 2026  ·  Last updated July 28, 2026  ·  Standard news article

What does a filer in the Tenth Circuit have to do differently now?

Verify that every case in the brief exists, and that every quotation and proposition attributed to it exists, before the brief goes out. That was always the honest reading of the rule. What changed on July 27 is that the court wrote it down in an opinion marked PUBLISH, in terms that leave no room for treating a citation error as a clerical slip.

The second duty runs to what happens after the error surfaces. Opposing counsel flagged the bad citations in the Response Brief. The Reply Brief conceded "several citation errors," said they "have been corrected," attached no corrected brief, and argued instead that the Response Brief had mischaracterized controlling authority. The panel refused the pivot: "It is the court that decides which issues are 'more significant.'"

A Notice of Errata did arrive, filed by the authoring attorney the day after oral argument. The court gave partial credit and took most of it back: "But it should not have taken questioning at argument to elicit such a response."

What exactly did the court say Rule 46.5 requires?

The operative passage is short and worth having verbatim. The court began with the certification, that counsel has undertaken a reasonable inquiry to ensure that "the issues presented are warranted by existing law or by a nonfrivolous argument for extending, modifying, or reversing existing law," citing 10th Cir. R. 46.5(B)(2). Then:

It should go without saying that a bare-minimum reasonable inquiry will ensure legal authorities cited in papers submitted to the court actually exist, and the quotations or propositions attributed to them also exist. The Opening Brief in this case therefore failed to fully comply with Rule 46.5.

Lichfield v. Kubler, No. 25-4135, slip op. at 44

The next sentence supplies the consequence: "The filing of this non-compliant brief is conduct sanctionable at our discretion. 10th Cir. R. 46.5(C)." The opinion then lists what the court may do on its own motion, without any party asking: dismissal of the appeal, monetary payments, initiation of disciplinary proceedings, and payment of the opposing party's reasonable expenses.

Existence of authority is folded into the certification a lawyer already signs. No new rule was written. An old one was applied to a new failure mode.

Why did the court decline to sanction, and what did it say about next time?

Because of volume, measured against its own recent decision. In Amarsingh v. Frontier Airlines, Inc., No. 24-1391, 2026 WL 352016 (10th Cir. Feb. 9, 2026), a brief contained seven cases that do not exist and two that exist but do not contain the quoted material. The court ordered a $1,000 payment toward opposing counsel's fees and costs and referred the matter to the relevant disciplinary body. In Lichfield the count was one nonexistent authority and two mischaracterized cases. "Here, the errors are not as egregious," the panel wrote, and the brief "had significantly fewer errors than in Amarsingh."

That is the whole of the mitigation, and it is thin. Fewer fabricated citations than the last person is not a defence to plan around. The court paired the mercy with a warning in the same paragraph:

But, on the other hand, generative artificial intelligence is no longer a novel tool. Attorneys who choose to use it must be willing to also apply "actual intelligence in its execution."

Lichfield v. Kubler, No. 25-4135, slip op. at 45 to 46

The closing lines do the same work. Submitting a brief with fabricated or hallucinated citations "is a divergence from the standards expected of attorneys appearing before this court. Any such error should be corrected explicitly and without prompting." Then, immediately after declining to sanction: "Counsel is responsible for ensuring briefs and representations made to the court are warranted by the law and supported by the record, no matter the technology used to create them."

One qualification the court made itself, which most coverage will drop. The panel never found that generative AI produced the errors. It wrote that it assumed as much: "Although we assume the use of generative artificial intelligence led to the misrepresentations of legal authority." The duty does not depend on the answer, which is the point of that last clause.

Why does the published marking matter more than the outcome?

Because of what the court says about the alternative. A footnote attached to the Amarsingh discussion reads: "We cite unpublished decisions for their persuasive value only and do not treat them as binding precedent. 10th Cir. R. 32.1(A)." Amarsingh is unpublished. Lichfield is marked PUBLISH on its face. The circuit's most developed statement on fabricated citations has moved out of a category the court expressly declines to treat as binding.

Two things this page will not say, because the opinion does not. It does not describe itself as the circuit's first published word on generative-AI citations, and it says nothing about the precedential reach of this section. The only remark on precedential status is that footnote, and the footnote is about unpublished decisions. What is on the document is the word PUBLISH, the date, and the reasoning.

How does this compare with what other courts did to the same conduct?

The same fact pattern drew four different responses. The useful axis is not who was harshest. It is which lever each court reached for.

The contrast with Illinois is why this page exists alongside our coverage of that decision. The Illinois First District priced the conduct at $1,500 per false citation, deliberately above past practice, on the theory that a number changes behaviour. The Tenth Circuit charged nothing the day before and moved the standard into a published opinion instead. Deterrence by tariff on one side, deterrence by precedent on the other. The Illinois number tells you what an error costs. The Tenth Circuit text tells you what an error is.

Lichfield sits apart from the four-state split over self-represented filers. Amarsingh involved a barred attorney litigating pro se. Lichfield involved counsel of record on a represented appeal, and Rule 46.5 attaches to the attorney presenting the brief.

Abroad there is a fourth lever. A Brazilian federal court treated an unchecked AI-assisted pleading as juridically non-existent, so the filing had no effect and default followed.

What should a firm change in its brief-checking process?

Three things, and none of them are about which model you use.

First, cite verification has to be a separate, evidenced step with a name attached, performed against the reporter or the database record rather than against the draft. A step that leaves no trace is hard to describe to a panel as a reasonable inquiry.

Second, build the self-correction path before you need it. The costliest moment in this opinion was not the bad citation. It was the reply brief that acknowledged errors, filed no corrected brief, and attacked the other side. When a citation problem surfaces, file the correction immediately, on your own initiative, with no argument attached.

Third, do not let the outcome here become your internal precedent. Counsel escaped a sanction on a volume comparison. The court said in the same breath that generative AI "is no longer a novel tool," which is how a panel closes the window on a good-faith surprise defence.

This reaches past litigators. Any professional whose work product carries cited authority, an accountant citing a revenue ruling, a compliance officer citing a regulation, faces the same failure mode and needs the same control.

What this decision does not do

It sanctions no one. It does not ban generative AI in briefs and does not require disclosure of AI use, unlike the standing orders some district judges have adopted. It creates no new rule. The underlying appeal was a defamation case against a filmmaker and Netflix, affirmed on other grounds, and had nothing to do with AI. And the panel does not find that generative AI caused the errors. It assumes it.

One thing this page declines to do. The opinion's discussion of the citation errors refers only to "counsel" and to "the authoring attorney," so that is as far as this goes.

Frequently asked questions

Did the Tenth Circuit sanction anyone in Lichfield v. Kubler?

What does 10th Cir. R. 46.5 require of an attorney filing a brief?

What sanctions can the Tenth Circuit impose for a brief that does not comply with Rule 46.5?

How does this differ from the Illinois appellate decision issued the next day?

Last verified: July 28, 2026