AI Regulation Tracker / Court ruling
Eleventh Circuit Refers Lawyer for Discipline Over Fake AI Citations
On July 10, 2026, the U.S. Court of Appeals for the Eleventh Circuit affirmed a dismissal and, in a published opinion, referred plaintiffs' counsel Anthony F. Sabatini for possible discipline after finding his briefs relied on fake, AI-generated citations. A referral is a first step, not a sanction, but the court has now routed hallucinated citations into the attorney-discipline machinery.
The underlying case was a Covid-era dispute. A group of commercial aviation employees objected to their employers' pandemic policies on masks, testing, and vaccination. The district court dismissed their third amended complaint for lack of personal jurisdiction and failure to state a claim, and on July 10, 2026 the Eleventh Circuit affirmed. That part is ordinary. The part that matters for the rest of us is at the end of the opinion.
What the court actually did
Judge Britt Grant, writing for the panel, said the court was being "more candid than usual" about how weak the claims were, because plaintiffs' counsel Anthony F. Sabatini "has not been candid with us." Sabatini, the opinion says, "filed multiple briefs replete with fake and hallucinated citations. Even after being warned."
Here is the sequence, straight from the opinion. His opening brief relied on at least eight nonexistent cases, including one it said the Eleventh Circuit had decided. After the defendants flagged the problem, Sabatini acknowledged in a late proposed reply brief that those citations were "erroneous or unverifiable" and asked to withdraw his reliance on eight listed cases. Then, in the court's words, "things go from bad to worse: the eight cases Sabatini 'withdrew' did not match a single one of the eight hallucinated cases in his opening brief. And not only were they not the right cases, all eight were also hallucinated."
The court did not treat this as a filing glitch. It treated it as a professional-responsibility failure. "By outsourcing his legal work to an AI algorithm," the panel wrote, "Sabatini violated his ethical duties to both his clients and this Court." It reached back to the duty of competent representation under Model Rule 1.1, and to the ABA's own guidance on generative AI, Formal Opinion 512 (2024), to make the point that "completely outsourcing one's legal work to artificial intelligence software is not competent."
Then the line that is going to get quoted for years: "Whatever the merits of artificial intelligence, it is no substitute for actual intelligence. Any 'use of AI requires caution and humility.'" The court borrowed that last phrase from Chief Justice Roberts's 2023 Year-End Report on the Federal Judiciary. It also stated the duty in plain terms: "When lawyers rely on AI tools, there is no way to get around the obligation to verify that the software got it right, and that what it got wrong will not end up in court filings."
The consequence is the headline. "By signing his name on briefs filled with hallucinated citations, not once but twice," the court wrote, Sabatini "violated his professional obligations, both to his clients and to this Court." So, "in a separate order, this Court, through the Chief Judge, will refer the matter to the Committee on Lawyer Qualifications and Conduct," citing 11th Circuit Rules, Addendum 8, Rule 2(C).
Two things to be precise about. First, a referral is a first step, not a verdict. The committee has not ruled, and nothing in the opinion suspends or disbars anyone. Second, the court was careful to say this level of misconduct is unusual. It noted that when this kind of problem surfaces, "the typical response is some version of an apology from the attorney, often with sanctions to follow," and that "it is rare to see the kind of blatant and repeated misconduct that Sabatini has committed in this case." Read that as a signal about direction, not as a claim that every AI slip ends in a discipline referral. What is unmistakable is that a federal appeals court has now formally routed AI-fabricated citations into the attorney-discipline machinery, in a published opinion, with the lawyer named.
The Maryland warning shot, three months earlier
This did not come from nowhere. On April 30, 2026, Judge Julie R. Rubin of the U.S. District Court for the District of Maryland issued a memorandum opinion and order in Johnson v. Nationstar Mortgage, LLC, Civil No. 1:25-cv-00855-JRR. The plaintiff, Darrin L. Johnson Jr., was self-represented. He had moved for a preliminary injunction to stop a foreclosure, and the court denied it on the merits.
But the court added a section it titled "Admonition Concerning Citation of Hallucinated Authority." Reviewing Johnson's filings, the court could not locate a case he had cited, Jefferson v. U.S. Bank, 2020 WL 5500414 (D. Md. Sept. 11, 2020). When the citation was run through Westlaw, it "yields no results." The court concluded that "it appears the citation may be a 'hallucinated' case citation derived from the use of a generative artificial intelligence ('AI') tool (e.g., ChatGPT, Gemini, etc.)."
Judge Rubin then walked through Federal Rule of Civil Procedure 11 and made a point that self-represented litigants sometimes assume does not apply to them. "While pro se pleadings and filings are granted a degree of indulgence not extended to lawyers," the order says, "pro se litigants must still abide by Rule 11, and they may be sanctioned for its violation." Quoting recent decisions from within the district, the court underlined the through-line: "Parties bear responsibility for the filings they sign, regardless of actual authorship."
The court held its fire, but not quietly. "In this discrete instance, the court will not issue a Rule 11 show cause order as to Plaintiff's citation to a hallucinated case," Judge Rubin wrote. "Nevertheless, Plaintiff is strongly cautioned: should he engage in similar conduct in the future (i.e., citation to hallucinated authorities or the like), the court may exercise its discretion to issue an order to show cause why he should not face sanctions under Rule 11."
Put the two together. In Maryland, a self-represented filer with one fabricated citation gets a documented warning and a clear statement that the next time carries sanctions exposure. In the Eleventh Circuit, a licensed attorney with sixteen fabricated citations across two rounds of briefing gets named in a published opinion and referred for discipline. Same underlying problem, different point on the same escalating curve.
The pattern is now the story
Neither court claimed to be breaking new ground, and that is the point. The Eleventh Circuit said outright, "We are far from the first court to see lawyers uncritically rely on artificial intelligence software and submit briefs citing nonexistent cases." The Maryland order cited a run of recent decisions in the Fourth Circuit and its district courts dealing with fictitious, AI-assisted citations, and noted that courts have issued show cause orders on their own initiative when filers cite made-up cases.
What has shifted is the posture. The early wave of AI-hallucination stories read like isolated embarrassments: a lawyer caught out, an apology, maybe a fine. These two rulings reframe it as a professional-duty problem with an enforcement track: verification is an obligation you cannot delegate to the software, and the failure to verify is what draws the sanction. The Eleventh Circuit even flagged the technical reason the risk is real, citing peer-reviewed work on how AI models are prone to producing confident, agreeable outputs that "echo users' opinions and beliefs, even when those views are incorrect." That is a court telling the bar it understands why the tool fails, and holding the human responsible anyway.
What changes for lawyers
Nothing in either ruling requires you to disclose that you used AI, and neither one bans the tools. If you were waiting for a rule that says "announce your AI use," this is not it. What both rulings do is make explicit a duty that was always there and give it consequences.
The duty is verification, and it is personal. You certify every filing you sign. Rule 11 already made you responsible for a reasonable inquiry into the law and facts in anything you put in front of a federal court, and the professional conduct rules already required competence. AI does not change either one. The Eleventh Circuit said the quiet part out loud: signing your name to a brief you did not check is the violation, and "the AI wrote it" is not a defense. It is, if anything, an aggravator, because it shows you outsourced the judgment you were retained to provide.
The exposure is broader than a single bad brief. Discipline referrals, Rule 11 sanctions, and the reputational hit of being named in a published opinion all now sit on the table for the same conduct. And the standard reaches everyone who files, not just big-firm litigators. Solo practitioners, self-represented parties, and anyone using a consumer AI chatbot to draft a motion are all inside the same perimeter.
What to do now
Treat any AI-assisted output as a draft that is wrong until you have personally proven otherwise. Here is a verification protocol that would have stopped both of these problems before filing.
First, pull and read every case before you cite it. Not the AI's summary of the case, the actual opinion. Open it in Westlaw, Lexis, PACER, or a court's own database and confirm the case exists, the citation is correct, the holding is what you are claiming, and the court and year match. If a citation returns no results in a real database, it is not a formatting error, it is a fabrication. Delete it.
Second, verify quotations and pin cites against the source text. Hallucinating tools invent not just cases but language and page numbers. If you are quoting a proposition to a court, the quote has to appear where you say it appears.
Third, do not let AI check its own work. Asking the same tool to confirm its citations reproduces the error, as the Sabatini reply brief shows. The correction has to run through an independent, authoritative source, meaning a human reading a real opinion.
Fourth, keep a verification record. Note who checked each authority and against what database, and keep it. If a court ever raises a question, contemporaneous proof that you verified is the difference between a correctable mistake and a candor problem.
Fifth, protect confidential information before it goes near a tool. Understand a given AI product's data-retention and training practices before you paste in a privileged document or a client's personal data. That is a separate duty, but it lives in the same workflow.
None of this is exotic. It is the work you were already supposed to do. The change is that courts are now checking, and they are willing to name you when the work was not done.
Questions professionals are asking
Does the Eleventh Circuit ruling ban lawyers from using AI?
No. The court did not prohibit AI tools. It held that a lawyer who signs a brief remains personally responsible for verifying every citation, and that failing to do so violates duties owed to the client and the court. The problem was not that Sabatini used AI, it was that he filed fabricated citations without checking them, twice.
Has the attorney been disbarred or suspended?
No. The opinion refers the matter to the Eleventh Circuit's Committee on Lawyer Qualifications and Conduct through a separate order. That is the start of a disciplinary process, not its conclusion. The committee has not ruled, and no sanction has been imposed by the opinion itself.
Does this apply to self-represented parties, or only to licensed attorneys?
Both. The Eleventh Circuit matter involved a licensed attorney, but the District of Maryland order in Johnson v. Nationstar warned a self-represented plaintiff that Rule 11 still applies to him and that future citation of hallucinated authority could draw a show cause order and sanctions. Courts extend some indulgence to pro se filers, but not permission to cite cases that do not exist.
What is the practical duty these rulings impose?
Verify AI-assisted output against primary sources before filing. Read each cited case in an authoritative database, confirm the citation, holding, and quotations, do not rely on the AI to check itself, and keep a record of your verification. Certification of a filing means you did that work, not the model.
Is one hallucinated citation enough to get sanctioned?
It can be enough to draw a warning, as Johnson shows, and courts have issued show cause orders on their own initiative in other cases. Whether sanctions follow depends on the court, the conduct, and whether it repeats. The safest assumption is that a single fabricated citation is a serious problem, and that a pattern is a discipline problem.
RELATED BRIEFINGS
Browse the full AI Regulation News tracker
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel in the relevant jurisdiction.