AI Agents for Lawyers

Regulation & Ethics

The SRA's 2026 Warning Notice on AI Misuse: What UK Solicitors Must Know

The SRA AI warning notice of 17 August 2026 followed 42 misuse reports in a year. Solicitors must understand a tool well enough to review its output critically.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. Forty-two reports in a year, and why that number is a floor
  2. The three cases the SRA chose to cite
  3. "Well enough to review its output critically": what the standard means
  4. The enforcement ladder, from Ayinde to Kumar
  5. The Law Society and the Bar Council layer
  6. A compliance checklist for firms
  7. A policy clause to adopt
  8. What to do this month

The SRA's Warning Notice on the misuse of AI, published on 17 August 2026, followed 42 reports of AI misuse between July 2025 and July 2026. Its core standard is short: a solicitor must understand any AI tool well enough to review its output critically. It cites three 2026 cases, and eight days after it appeared the Solicitors Disciplinary Tribunal struck off its first lawyer for defending himself with AI-fabricated authorities.

Forty-two reports in a year, and why that number is a floor

Forty-two reports to a regulator over twelve months sounds modest. Read against the other counts, it is the visible part of a larger problem. Matthew Lee's UK tracker at naturalandartificiallaw.com listed 76 UK cases of confirmed or suspected AI-hallucinated citations in September 2026; Damien Charlotin's global database counted 69 UK decisions on 14 September 2026 out of 2,041 worldwide. Those are only the incidents that reached a judgment. A Clio survey reported by the Law Gazette on 8 April 2026 found that 89% of UK and Irish legal professionals use AI, 70% adopted it in the past year, 17% of firms have no policy, and only 7% of clients recall being told. A LexisNexis survey of around 900 UK lawyers in February 2026 found 69% rated verification as weak.

The SRA's own count, then, is the number of times something went wrong badly enough for someone to tell the regulator. The notice exists because the SRA expects the real figure to be higher and wants firms to close the gap before it does so through enforcement.

The three cases the SRA chose to cite

The notice does not argue from abstractions. It names three decisions, each illustrating a different failure.

CaseCourtWhat went wrongWhat the SRA takes from it
BCP v A Mother [2026] EWFC 71 (B)Family CourtA 2026 family decision the SRA cites as an instance of AI misuse; read it before relying on AI-assisted material in any family matterAI-assisted material in family proceedings attracts the same scrutiny as in any other court
Cork v Smith [2026] EWHC 1199 (Ch), 4 June 2026High Court, Chancery DivisionA junior associate at Pinsent Masons asked an AI tool to research insolvency law; it invented "Insolvency Rule 12.37(5)". The tool had flagged that its answer might be inaccurate; the associate did not check. The judge criticised a "cavalier attitude"; the firm was referred to the SRA and agreed to bear the extra costsSupervision is a firm duty; a warning label on the output is not a substitute for opening the source
UK v SSHD [2026] UKUT 81 (IAC)Upper Tribunal, Immigration and Asylum ChamberClient correspondence pasted into public AI toolsDoing so puts the material "in the public domain"; confidentiality under the Code is engaged at the moment of upload

Cork v Smith deserves a second look because it involves a large firm with policies, training and supervision, and the failure still happened. The UK Human Rights Blog has the detail. The tool told the associate its answer might be wrong. The associate filed it anyway. That is the behaviour the SRA's standard is aimed at.

"Well enough to review its output critically": what the standard means

The phrase is deliberately not a technical benchmark. It does not require a solicitor to explain transformer architecture, and it does not require a certificate. It requires enough understanding to know where a given tool fails and to check those places. The plain-language explainer on this site covers the minimum: a language model predicts plausible text, has no database of authorities unless one is connected, will invent citations that look right, and will agree with a false premise if you assert it confidently.

Applied to daily work, the standard breaks into four expectations that map onto the existing Code of Conduct for Solicitors, RELs and RFLs.

Code provisionWhat the notice expects
Paragraph 1.4 (do not mislead the court) and 2.4 (only properly arguable assertions)Every authority, quotation and factual statement produced by AI is verified against the source before it reaches a court or counterparty
Paragraph 3.2 (competent service) and 3.3 (maintain competence)The solicitor knows which tool was used, what it is grounded on, and what its known failure modes are
Paragraph 3.5 (supervision)Partners and supervisors know when juniors are using AI and check the work with that in mind; Cork v Smith is the reference case
Paragraph 6.3 (confidentiality)Client data goes only into tools the firm has assessed; public consumer tools are out, per UK v SSHD
Paragraph 7.7 (report serious breaches)An AI incident that reaches a court or a client is assessed for reporting like any other breach
Principles 1, 2, 5 and 7Administration of justice, public trust, integrity and the client's best interests all point the same way

The Divisional Court's own formulation in Ayinde v Haringey [2025] EWHC 1383 (Admin) is blunter: "freely available generative AI tools ... are not capable of conducting reliable legal research". The SRA's version is more careful because enterprise tools grounded on real authorities exist and are used across the profession. The obligation is the same either way: the tool proposes, the solicitor checks. The verification protocol sets out a procedure that satisfies it.

The enforcement ladder, from Ayinde to Kumar

The notice sits on top of fourteen months of judgments that have already established what misuse costs.

StepExampleOutcome
Wasted costsAyinde v Haringey and Al-Haroun v Qatar National Bank, 6 June 2025£2,000 plus VAT wasted costs in each case; five fake authorities in Ayinde, 18 of 45 fictitious in Al-Haroun
Regulatory referralAyinde (barrister and solicitor referred to the BSB and SRA); Cork v Smith (firm referred)The referral, not the costs order, is the lasting consequence
Adverse credibility findingAyinde: barrister Sarah Forey's denial of AI use found "unpersuasive"The court reached the contempt threshold but brought no proceedings because she was junior and already reported, per the Law Gazette
Sanctions the court listedAyindePublic admonition, costs, wasted costs, strike-out, referral, contempt, referral to the police
Fake authorities against the regulator itselfBandla v SRA [2025] EWHC 1167 (Admin), May 2025Multiple fictitious authorities cited to the High Court; see Mountford Chambers
Striking offSRA v Kumar, SDT, decided 25 August 2026, reported 4 September 2026A registered foreign lawyer used AI-generated quotations and citations in his own SDT defence, then filed a "corrected" response with more fabricated material; struck off the Register of Foreign Lawyers, per Legal Futures

Two features of this ladder matter for firm policy. First, the Divisional Court in Ayinde addressed "heads of chambers and managing partners" directly, which is why the SRA's notice speaks to firms and their compliance officers, not only to individuals. Second, in every escalated case the aggravating factor was the response: denial in Ayinde, a second fabricated filing in Kumar, silence about a warning label in Cork v Smith. A solicitor who discovers a hallucination and tells the court and the client the same day is in a different position from one who quietly refiles. The casebook shows the same pattern in the United States.

The Law Society and the Bar Council layer

The Law Society issued guidance and a press release on the SRA's notice; its practical value is in translating regulatory expectations into firm procedures, and it should be read alongside the notice rather than instead of it. The Bar Council's "Considerations when using ChatGPT and generative AI", last reviewed on 25 November 2025, is not formal BSB guidance but is the document the BSB would expect a barrister to have read. The judiciary's guidance for judicial office holders, updated on 31 October 2025, tells judges not to enter private information into public AI tools and makes them responsible for all material in their name. The UK guidance overview compares all three; the point here is that a solicitor instructing counsel or appearing before a judge is now dealing with people bound by the same expectation.

A compliance checklist for firms

The notice does not prescribe a programme. This one is built to answer the questions an SRA investigator would ask after an incident, which is the only test that matters.

Governance

  1. The COLP owns AI compliance and reports on it to the management board at least twice a year.
  2. A written AI policy exists, is dated, and names the approved tools; the policy page has an outline.

Tools 3. Every approved tool has a recorded assessment covering training use, retention period, data location and access controls; consumer accounts are blocked for client work. 4. The privacy settings of each approved tool are set centrally, not left to individual users.

People 5. All fee earners and support staff complete AI training on joining and annually; attendance and content are logged. 6. Supervisors ask, as a matter of routine, whether AI was used on a piece of work and check accordingly; the junior lawyers page explains why the pressure falls on the most junior.

Matters 7. No authority, quotation, rule reference or factual assertion produced by AI leaves the firm until a named individual has opened the source and recorded that they did so. 8. Documents that may go before a court or tribunal carry an internal verification sign-off before release.

Clients 9. Client care letters state that AI tools may be used, what safeguards apply and that the client may object; the disclosure page has wording.

Incidents 10. Any AI-related error that reaches a court, a client or a counterparty is logged, corrected with candour, and assessed by the COLP for reporting under paragraph 7.7.

A policy clause to adopt

Use of AI tools

1. Fee earners and staff may use only the AI tools listed in Schedule A for
   work involving client information. Consumer accounts, including personal
   ChatGPT, Claude, Gemini or Copilot accounts, must not be used for client work.

2. Output from an AI tool is a draft. No authority, quotation, statutory or
   rule reference, or statement of fact produced by an AI tool may be included
   in any document sent to a court, tribunal, client or third party unless a
   named fee earner has personally checked it against the primary source and
   recorded the check in the matter file.

3. Where an AI tool flags that its output may be inaccurate or incomplete, that
   flag must be resolved by verification before the output is used.

4. Supervisors must ask whether AI was used in preparing work submitted to them
   and review the work on that basis.

5. Any error attributable to AI output that has reached a court, tribunal,
   client or third party must be reported to the COLP on the day it is
   discovered and corrected with full candour.

What to do this month

Send the notice to every fee earner with the three cited cases attached; Cork v Smith is a fifteen-minute read that changes behaviour. Check which AI accounts are actually in use across the firm, including personal ones on firm devices, and compare the list against Schedule A. Add the verification clause above to the policy and the sign-off step to the pre-filing checklist for litigation teams. Then run one training session that shows, on screen, a tool inventing an authority and how to catch it; Module 6 of the AI Lab for Lawyers course does exactly this with the UK, EU and US rules side by side. The question of whether a solicitor may use ChatGPT at all has been answered: yes, on the firm's terms, with the solicitor's name on the result.

Frequently asked questions

What does the SRA warning notice on AI actually require?

It does not create new rules. It applies the existing Principles and Code of Conduct to AI and states the expectation that solicitors understand any AI tool well enough to review its output critically. In practice that means verifying every authority and quotation, keeping client data out of public tools, supervising juniors who use AI and reporting serious breaches.

Can I put client information into ChatGPT as a solicitor?

Not into a consumer account that may retain or train on it. In UK v SSHD, the Upper Tribunal said that pasting client letters into public AI tools puts them in the public domain, and the SRA cites the case in its notice. Enterprise tiers with contractual no-training and retention terms are a different question, but they still require the firm's approval and a confidentiality assessment.

What happens if a solicitor files AI-hallucinated cases in England and Wales?

Since Ayinde v Haringey in June 2025 the Divisional Court has listed a ladder from public admonition through wasted costs to contempt and police referral, with referrals to the SRA. In 2026 the firm in Cork v Smith was referred to the SRA and bore extra costs, and the SDT struck off a lawyer in SRA v Kumar. Candour lowers the outcome; denial raises it.

Does the SRA warning notice apply to firms or only to individual solicitors?

Both. The Code of Conduct for Firms requires effective systems and supervision, and the compliance officer for legal practice is responsible for them. The Divisional Court in Ayinde addressed heads of chambers and managing partners directly, and the SRA's notice follows that lead. Individual solicitors remain personally responsible for what they sign.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion