Can a lawyer be disbarred for using AI? What the 2026 cases actually punished

Yes, and it has now happened. On 3 September 2026 the Solicitors Disciplinary Tribunal struck Abhishek Kumar off the register of registered foreign lawyers in case 12884/2026, the first time a UK tribunal has ruled on a lawyer’s use of a large language model. No American lawyer has been disbarred over AI as far as the public record shows. Suspensions arrived anyway, from the Ninth Circuit in June and the Nebraska Supreme Court in April. Read the three together and the same thing escalated all of them. It was never the fake citation on its own. It was the document the lawyer filed after the fake citation was found.

The strike off the search results have not caught up with

Kumar reached the tribunal by a different road. He was convicted in January 2024 under section 21 of the Immigration, Asylum and Nationality Act 2006 for knowingly employing a man disqualified from work by his immigration status, and the SRA took that conviction to the SDT. The AI allegation was bolted on afterwards, because of what he did while defending himself.

His Answer to the Rule 12 allegation, dated 12 March 2026, contained fabricated and unsupported authorities. Counsel for the SRA spotted them. Kumar replied by email on 9 April 2026 admitting he had used generative AI, and that reply was itself drafted with AI and carried fresh false material into the same file. He told the tribunal he lacked the expertise to verify the output, and that Ayinde did not reach him because he was a registered foreign lawyer rather than a solicitor.

The tribunal was unmoved. Culpability “very high”, and a persistent lack of insight, according to Legal Futures’ report of the hearing. The detail that matters to everyone else is where the tribunal placed its emphasis: it gave very substantial weight to his use of AI to answer for the errors his first use of AI had caused. Two allegations, one sanction, and the SDT recorded that it would have reached the same outcome on either allegation standing alone. This is not a conviction case with an AI footnote attached.

The Ninth Circuit said the quiet part in a published order

Lnu v. Blanche, No. 24-4790, filed 3 June 2026 before Judges Paez, Bea and Forrest. Marked for publication, deliberately, as a warning to the court’s own bar.

Mike Singh Sethi signed an opening immigration brief citing two cases that have never existed, and hung quotations on two real ones that do not contain them. When the fabrications surfaced he filed a Motion to Correct describing them as typographical errors and substituting real cases with vaguely similar names. At oral argument William Rounds was asked three times whether generative AI had been involved and said no each time, then conceded minutes later that it was possible. The briefs had been written by unlicensed law graduates. No admitted lawyer had read the cited cases.

The panel was precise about what it was not doing. It did not sanction anyone for using AI. In its words, the rules “are not violated at the point of research and drafting, but at the point of signing and filing.” The signature was the attestation, and Sethi’s own admission that the firm did not normally vet its brief writers’ citations got called an extraordinary confession.

Then the sentence to read twice. Had they disclosed in the Motion to Correct that AI had been used against firm policy and apologised for not checking, the court said, “lesser sanctions may have been warranted.” They did not. The gravity of the discipline, including the suspension, was owed to that repeated failure of candor. Each lawyer paid $2,500 and lost six months of practice before the court. Every attorney at the firm now files a sworn statement about AI use with every filing for two years, and the clerk was ordered to send the order to the State Bar of California.

Nebraska, where the expensive word was no

Greg Lake argued a divorce appeal before the Nebraska Supreme Court in February 2026. Opposing counsel counted 57 problems across 63 references. Asked from the bench whether he had used artificial intelligence, Lake said he had not.

On 20 March the court answered in a unanimous published opinion, Prososki v. Regan, 321 Neb. 38. It listed twenty fictional quotes and misrepresented holdings, noted the errors were discoverable with ordinary research tools, and referred him to the Counsel for Discipline for what it called a failure of his duty of candor toward the court. Two days before his suspension he filed an affidavit admitting the AI use and calling his own lack of candor a grave error of judgment. On 15 April the chief justice signed a one page order suspending him until further notice, with a referee still to recommend final discipline.

He had two months between the hearing and the referral. He spent them holding the line.

What a real correction has to contain

Every firm has had the cite-check conversation by now. Almost none have written down what happens in the hour after somebody finds a bad citation in a filing that has already gone out, and on the 2026 record that hour is where the damage is done.

The Ninth Circuit spelled the duty out. Tell the court and opposing counsel immediately. Say what kind of error it was, a fabrication or a gross misrepresentation of a real case, and say how it came about, which in that matter meant misused generative AI.

There is a trap inside this that deserves its own warning. In a separate case, Contreras Pelayo v. Bondi, No. 24-5168, Sethi filed a Notice of Errata that listed the bad citations as errors and slotted real cases into their place, without ever saying they were hallucinations. The panel held that insufficient, because swapping a hallucination out for a real case does not correct the earlier claim that the fake case was real. A tidy errata that avoids the word is worse than no errata at all. It is a second uncorrected false statement to the court, and it is the one the panel treated as subterfuge.

One more line worth pinning above the desk. Candor, the court said, requires admissions of uncertainty just as much as it requires revealing facts known for certain. You are allowed to tell a judge you do not yet know whether AI was involved and that you are finding out. You are not allowed to say no.

What is actually on the table, in both countries

Be accurate about the downside, because the content farms are not. In the United States the realistic 2026 exposure is a court-specific suspension, a monetary sanction, a referral to your state bar, and standing filing conditions that attach to everyone in your firm rather than to you alone. Disbarment has not happened. It is the reciprocal discipline that can follow a referral, not the sanctions order itself, that would get anyone there. Our running account of the hallucination sanctions tracks how the money side has moved.

In England and Wales the top of the range is strike off, and it has now been used. The route was signposted in Ayinde, decided by the Divisional Court on 6 June 2025 under the Hamid jurisdiction, which asked the regulators to consider urgently what was needed beyond guidance. The SDT accepted the SRA’s position that this is an issue of profound public importance. If you want the baseline you are being measured against, start with what the SRA actually expects.

A last point, and it is the one I would act on tomorrow. The Ninth Circuit adopted the broader definition of hallucination from the Stanford study and warned that inaccuracies, meaning real cases cited for propositions they do not support, may prove more dangerous in the long run than invented ones, because they survive a facial check. A tool that has stopped making up case names has not stopped hallucinating. We worked through what that does to a cite-check in our piece on the Sixth Circuit and legal research tools.

Write the disclosure paragraph now, while nothing is on fire, and decide today who in the firm gets told first. The confidentiality half of this problem is in our free guide, and our tools page says which AI we would let near a filing at all.

FAQ