The Tribunal Did Not Ask What the Tool Got Wrong. It Asked Who Signed.

On 25 August 2026 the Solicitors Disciplinary Tribunal struck a registered foreign lawyer off its register. One of the two allegations was that his written defence, and the email meant to correct it, contained AI-generated citations that did not hold up. His answer was that he lacked the expertise to check them. That defence has now been tried and rejected.

The person who signs a document is responsible for every authority in it, whatever produced the draft. That is the whole of the finding, and it is not new law.

In July we argued that you cannot rent case law discipline. This is what its absence costs.

What did the tribunal decide, and what did it not decide?

The tribunal struck the respondent off the register of foreign lawyers on two allegations and said it would have reached the same sanction on either one alone.

The first allegation had nothing to do with AI. He had been convicted in 2024 of knowingly employing a man with no right to work in the UK, and the SRA referred him on that conviction in January 2026. The AI misuse was added to the charges later, after he filed his defence. Calling this a strike-off for AI, full stop, leaves out half the case.

The second allegation is the one this piece is about. His written answer, filed in March 2026, cited four authorities. Two had citations that did not resolve to the case named. Two were real cases relied on for propositions they did not support. Counsel for the SRA spotted the problems on 8 April and pointed him to Ayinde.

His reply the next day was meant to correct the errors. It had also been drafted with AI, and it contained a corrected citation that was still wrong, quoted passages that do not appear in the judgment, and a word in quotation marks that the judgment does not contain. The SRA added the AI conduct as a second allegation in June. Legal Futures, reporting the judgment on 4 September, describes it as the first time the tribunal has had to deal with AI hallucinations.

What it did not decide matters too. Dishonesty was not alleged. The respondent was not a solicitor and was not acting for a client. Typographical errors in citations were not serious enough to be misconduct. The SRA accepted he was impecunious and sought no costs.

What was the defence, and why did it fail?

Every limb of the defence placed the fault somewhere other than the person who filed the document.

He had no intention to mislead. He "simply did not have the expertise to verify the AI output". He was unaware of the regulatory warnings. It was "technological negligence", not a lack of integrity. He was an unqualified lawyer used to being supervised. Ayinde was about solicitors and he was a registered foreign lawyer. He was not practising and considered himself a member of the public.

The tribunal accepted none of it. He was a regulated individual obliged to comply with the standards of his profession. It found he had started using AI without proper thought to the consequences and had failed to ensure his output was accurate or to satisfy himself the tool was reliable. On culpability it put the principle in one line: "A lawyer is always responsible for the accuracy of his or her output."

The repetition is what carried the sanction. Having had the errors explained, and having been directed to the leading judgment on exactly this, he went back to the tool and produced a second set. The tribunal gave that very substantial weight. Acting for himself was not mitigation either, because he was the intended beneficiary of his own submissions.

Why is paragraph 2.4 the rule that caught it, rather than a new AI rule?

Because the duty breached is the duty not to put something before a court or tribunal that is not properly arguable, and that duty predates the tool.

Paragraph 2.4 of the Code says you only make assertions or submissions "which are properly arguable". The tribunal found it breached in respect of the two real cases relied on for propositions they did not support, alongside paragraph 1.4 and Principles 1, 2 and 5. None of those provisions mention AI, and none needed to.

Two passages settle where responsibility sits. The tribunal said there was no distinction in principle between material a lawyer sends to a regulator, a disciplinary tribunal or a court. Then it said AI was the medium here, but the same issue would arise with a search engine, or with another person's work for which the lawyer had ultimate responsibility.

The regulatory event is not the tool's mistake. It is the filing.

What did Ayinde already say, and who does it now clearly cover?

Ayinde v Haringey, decided by the Divisional Court in June 2025, held that lawyers who use AI for legal research have a professional duty to check it against authoritative sources before relying on it.

The court joined two Hamid referrals, one involving a barrister and one a solicitor, where unverified citations had reached the court. It listed the sources that count: legislation.gov.uk, the National Archives case law database, the official Law Reports and reputable publisher databases. It referred the lawyers to their regulators and asked the regulators to consider urgent steps beyond guidance. The High Court instalment that followed was the Pinsent Masons ruling, where the same duty reached a large firm's supervision chain.

The open question after Ayinde was scope. The SDT has answered part of it: a registered foreign lawyer, not in practice, representing himself before a tribunal rather than a court, is inside it. The SRA's warning notice agrees. Reliance on an AI output is not a suitable defence.

Invented cases or wrong citations for real ones: the distinction the tribunal left open

The tribunal said that, case by case, there may be a difference in harm and culpability between citing cases or propositions that do not exist and providing incorrect citations for cases that do.

It did not draw the line, and said it was not fettering other divisions on different facts. We are not going to draw it either. Two points will matter later. The SRA's own counsel conceded that a bare citation error would not, on its own, amount to an unarguable case, and typographical errors fell below the misconduct threshold. But a real case relied on for a proposition it does not support was an unarguable submission. The judgment describes this respondent's conduct in both terms in different places, so the line is still open.

For a firm the practical reading is narrower. The check has to cover whether the paragraph relied on says what the document claims, not just whether the case exists. That is a pinpoint question, and it is the one general tools are worst at.

What does the UK Jurisdiction Taskforce say about who is liable?

The same thing, from the other direction. The tribunal decided one case. The UKJT set out the general rule a month earlier, and it is the rule the tribunal applied.

On 7 July 2026 the UK Jurisdiction Taskforce published its legal statement on liability for AI harms under the private law of England and Wales. Its conclusion is that no new law is needed. AI has no legal personality, so it cannot be liable for anything. Liability sits with the person who used it, under contract where there is one and under negligence where there is not. The statement draws a line between perceived uncertainty and actual uncertainty: the sense that AI liability is unsettled is a perception, not the state of the law.

For a professional, the statement puts the standard where it always was: reasonable skill and care, judged against a competent member of the profession, with regulatory guidance shaping what that means. It then lists the ways a professional can fall short. Using AI inappropriately. Using an unsuitable model. Failing to do proper due diligence on it. Failing to test it. Failing to validate its output. And, cutting the other way, failing to use AI at all where a competent peer would have done. That last one is the statement's, not ours, and it is worth quoting because it stops this being read as an argument against the tools. The duty runs both ways.

The statement's own worked example is a barrister who files AI-generated authorities without removing the invented ones. It says the barrister has breached the duty of care and may be in contempt. That was published before the SDT judgment was reported. The tribunal did not need it, but it is useful to a firm reading both, because the answer to "I could not have known" is now in the doctrine as well as in the outcome.

The Master of the Rolls has been saying the practical version for two years. His three rules for lawyers using large language models, repeated at Legal Geek in October 2025, are that you understand what the model is doing before you use it, that you keep private data out of a public model, and that you check what comes out before you use it for any purpose at all. Rule three is the whole of this piece in a sentence.

What does "the check sits before the signature" look like as a workflow?

The check happens before signature, by the person whose name goes on the document, against the source rather than inside the tool that produced the draft.

1. Every authority resolves to a source the signer can open. Neutral citation and one of the Ayinde sources. A case name alone is a lead, not a citation.
2. Every quotation is checked against the paragraph, not the headnote. If the words are in quotation marks they must appear in the judgment.
3. Every proposition is checked at the pinpoint. The draft carries the paragraph number relied on and the signer reads that paragraph. This step would have caught two of the four authorities in the SDT case.
4. The check is recorded. Who did it, on which version, when. On witness statements PD 32 already requires the statement to disclose how it was prepared, so the record serves twice.
5. Corrections come from the source, never from the tool. The 9 April email is the case study. Asking the tool to fix its own output produced a second set of errors and turned one allegation into a repeated one.

None of that is new work. What has changed is how finished the draft looks before anyone has read it.

What does the SRA warning notice ask of a firm, in practice?

The notice asks a firm to be able to show who checked each authority, against what, and that supervision covers AI-assisted work.

Its central sentence, published 17 August 2026, is this: "AI has no separate legal personality; solicitors and regulated individuals who use AI in the course of delivering legal services remain accountable for their work and outputs, regardless of how that work has been prepared." Before relying on a named authority, it says, you should be assured it is genuine, relevant, has a verifiable citation and advances the argument being put forward.

The notice then extends the duty upward. Supervisors remain accountable for work carried out through them (Code paragraphs 3.5 and 3.6), firms must have effective systems, controls and supervision (firm Code 2.1, 4.3 and 4.4), and the COLP must take all reasonable steps to ensure compliance. The accompanying press release put a number on it: 42 reports of potential AI misuse between July 2025 and July 2026, with investigations ongoing on citations, supervision and confidentiality.

So the question a COLP should expect is not "do you have an AI policy" but "show me the check on this document".

What should a claimant firm change this month?

Four things, none of which need a procurement.

Write the check down as a drafting standard, not an AI policy. Authorities are checked at source by the signer before signature. One paragraph, where fee earners will see it.

Require pinpoints on every authority in every draft, whatever produced it. A citation without a paragraph number goes back.

Fix errors from the source. If an authority fails the check, the correction comes from the judgment, not the tool.

Give supervisors sight of what was AI-assisted. Pinsent Masons turned on a supervisor who did not know. The SDT case turned on a respondent who checked nothing. The record closes both gaps.

On our side, every authority we have checked at source in LegalDocs Assist and CreditHire Assist carries a pinpoint marker to the paragraph relied on, checked against the judgment by a credit hire lawyer, and a weekly integrity check confirms those markers and flags are still intact. Not every case in the library is there yet; the ones without a judgment we can open are marked as such rather than dressed up. That does not replace the signer's check. It means the signer is reading a paragraph rather than hunting for one, which is why the curated content layer is where we think the durable value sits. The product page shows how LegalDocs Assist handles authorities.

The tribunal's question was not what the tool got wrong. It asked who signed. Every draft in your firm will be measured the same way.

© LegalDocs Assist — www.legaldocs-assist.co.uk