In partnership with
The SRA has moved from talking about AI to writing a warning notice about it. Published on 17 August, it confirms that the regulator has received 42 reports of potential AI misuse in a year, that some came from senior members of the judiciary, and investigations are already running. Elsewhere, Harvey has launched a version of its platform that remembers how you draft, everything Claude writes is now watermarked, and the numbers on in-house AI adoption turn out to be thinner than the press releases suggest.
AI in Practice
The SRA publishes a warning notice on AI, and reveals it is already investigating
The Solicitors Regulation Authority published a warning notice titled "Misuse of AI" on 17 August. Warning notices sit in a particular place in the SRA's architecture. They are not new rules, but the SRA says it will "have regard to" them when exercising its regulatory functions. This one states "If you fail to have proper regard to this warning notice, you are at risk of disciplinary action. We can and will act where we find evidence that solicitors, firms and/or their employees contravene our rules."
It applies to every firm and individual the SRA regulates.
The accompanying press release carries a number that should concentrate attention. The SRA received 42 reports relating to potential misuse of AI between July 2025 and July 2026, and there are "a number of ongoing investigations", covering inaccurate legal citations, supervision and confidentiality. The notice also adds a detail the press release misses: some of those reports came from senior members of the judiciary, and others were solicitors self-reporting after relying on a tool that produced inaccurate content.
Aileen Armstrong, the SRA's Executive Director of Strategy and Policy, put the position in one line: "Individuals remain responsible for the work they produce and the advice they provide, regardless of whether AI is used, or not."
Two concerns run through the document. The first is hallucinated material reaching the court, and the SRA sets its reasoning on cases, citing R (Ayinde) v Haringey LBC [2025] EWHC 1383 (Admin), BCP v A Mother [2026] EWFC 71 (B) and Cork and another v Smith [2026] EWHC 1199 (Ch).
The key point that this author cannot stress enough is: "Reliance on an output of AI would not be a suitable defence."
The notice also picks up the point at paragraph 29 of Ayinde, that a reference to the regulator is likely to be appropriate where a lawyer puts false citations before the court, and it reminds readers that misleading the court is (as said in Brett v SRA [2014] EWHC 2974 (Admin)) "one of the most serious offences that an advocate or litigator can commit".
The second concern is confidentiality, and here the SRA adopts the Upper Tribunal's observation in UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC) (covered in this newsletter here), that putting client letters and Home Office decision letters into an open source tool "is to place this information on the internet in the public domain". The consequence the SRA draws is blunt: privilege "may be permanently waived and unable to be recovered".
The part this author found most useful is the supervision analysis, because it is where the notice goes further than a general reminder to be careful.
Supervisors may themselves be found in breach where false citations reach the court without adequate review, under paragraphs 3.5 and 3.6 of the solicitors' Code, and firms carry parallel obligations under paragraphs 2.1, 4.3 and 4.4 of the firms' Code, with the COLP's duty under paragraph 9.1 sitting on top.
The SRA also points to Rule 9.4 of the Authorisation of Firms Rules, which requires regulated work to be supervised by someone who has practised for at least three years.
Taken together, this is the regulator saying that the risk does not stop with the fee earner who accepted the citation. It travels up. Firms that have handed AI tools to junior staff without changing anything about how their work is checked now have a documented standard against which that decision will be measured, and the SRA quietly added AI material to its effective supervision guidance back in June.
Takeaways
Act: Read the "Standards and Regulations" section at the foot of the notice, which lists the exact Code paragraphs the SRA will measure you against, and check your firm's AI policy names them. If your policy is a page of general encouragement to be careful, it will not survive contact with this document.
Watch: What the ongoing investigations produce. Warning notices tend to be followed by published outcomes, and the first Solicitors Disciplinary Tribunal decision built on this notice will tell the profession far more about the SRA's threshold than the notice itself does.
Risk: Supervision, more than the citation. A partner who never touched the tool can still be in breach of paragraphs 3.5 and 3.6 if the work went out unchecked, and "I did not know they were using it" describes the problem rather than answering it.
Read: SRA, Misuse of AI warning notice, the SRA press release, and the analysis at 4 New Square
On your radar
Harvey has launched a version that remembers how you draft: Harvey unveiled Harvey II on 18 August, built around a Memory function which retains an individual lawyer's drafting style, preferred structure, citation habits and level of detail, and carries those preferences into Word and Outlook as well as across the platform. It sits alongside matter-based "spaces", which bring documents, tasks, permissions and matter history into one place so that agents work with context already loaded. Harvey says users can see what it remembers, change it, or switch it off, and that Memory is never used to train models. Personal memory comes first, with shared and eventually organisation-wide memory promised. Why it matters for UK lawyers: a tool that accumulates a picture of how you work is a different data proposition from one that answers a question and forgets, and the SRA's confidentiality section this week turns on exactly what a system retains and who can reach it. If your firm is on Harvey, ask what Memory stores, where it is held, and whether a departing fee earner's memory travels with them or stays. (Harvey, via Artificial Lawyer, Legal IT Insider)
In-house teams have adopted AI, and then largely stopped: LegalOn's 2026 In-house Pulse survey, written up by Artificial Lawyer on 18 August, found that 54.9% of in-house respondents are still experimenting with prompts and tools without shared workflows or standards. Standardising prompts and workflows puts a team ahead of roughly two thirds of the market at 26.5%, grounding AI in the organisation's own playbooks and policies puts it in the top 11% at 8.8%, and around 2% are using in-house legal AI platforms to anything like their full extent. The piece concludes that the obstacle is organisational culture rather than the technology. Why it matters for UK lawyers: the general counsel who is said to be about to bring your work in-house is, on this data, probably still at the experimentation stage, which is useful context for anyone reading last month's warnings with a sinking feeling. (Artificial Lawyer, The Global Legal Post)
Everything Claude writes is now watermarked, and it survives a light edit: Anthropic has begun embedding machine-readable marking in the output of all Claude products released from 2 August, to meet the transparency obligation in Article 50(2) of the EU AI Act. The watermark works by selecting between equivalent word choices against a key, so it is invisible on the page, does not change the meaning, and stays in place when text is copied, pasted or lightly edited. Artificial Lawyer reported on 17 August that some vendors and lawyers are weighing whether to stop using Claude for drafting as a result. Why it matters for UK lawyers: Claude sits underneath a good deal of the legal AI now being rolled out in large firms, so this is not confined to people who use the chatbot directly, and a document that can be tested after the event turns AI disclosure into a question of what is discoverable rather than what you chose to volunteer. It proves nothing about accuracy, only about production method. Settle what your firm tells clients about AI-assisted drafting now, rather than after a client runs a detector over a draft you sent. (Artificial Lawyer, Legal IT Insider)
In case you missed it: the SRA rewrote its supervision guidance in June: The SRA added sections on AI to its effective supervision guidance in June, which passed with little comment at the time. This week's warning notice leans on it heavily, and the guidance is where the practical detail is such as, what appropriate human review of AI-assisted work looks like, how delegation should be recorded, and why an authorised individual retains responsibility for work done by unauthorised staff. Why it matters for UK lawyers: the warning notice tells you the standard, and the supervision guidance is the closest thing to an explanation of how to meet it. Read the two together before writing anything into a firm policy, because a policy drafted from the notice alone will miss the supervision detail the SRA has already published. (SRA, effective supervision guidance)
Ad Break
In order to help cover the running costs of this newsletter, please check out the advert below. In line with my promises from the start, adverts will always be declared.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
For Review
Misuse of AI: warning notice (Solicitors Regulation Authority)
The document itself, and short enough to read properly in twenty minutes. If you are pressed, go straight to "AI hallucinations – relevant considerations" and "Client confidentiality – relevant considerations", then open the collapsible "Standards and Regulations" section at the foot, which lists every Principle and Code paragraph the SRA considers engaged. That list is the closest thing to a compliance checklist the regulator has published on AI.
Read or listen: Solicitors Regulation Authority
SRA Warning Notice on "Misuse of AI" dated 17 August 2026: a timely reminder to legal practitioners (4 New Square Chambers)
A practitioners' analysis from a set that does a great deal of professional negligence and regulatory work, which is the right lens for this document. Useful if you need to explain to a management board why a warning notice matters more than its tone suggests.
Read or listen: 4 New Square Chambers
Speed Without Structure: What Legal AI Is Really Exposing (Dan Hoadley, Artificial Lawyer)
Hoadley's argument is that AI has made individual tasks faster while leaving the disconnected systems around them untouched, so the gains disappear into the gaps between tools. It is the counterweight to a week dominated by regulatory risk: the answer to both the accuracy problem and the productivity problem may be less about which model you buy and more about whether your matter information is joined up in the first place.
Read or listen: Artificial Lawyer
Practice Prompt
Try the below prompt to turn this week's warning notice into a firm AI policy, which is the step the SRA implies but does not spell out. It is written for a partner or COLP who has to put something in front of the board, and it works in two stages: it extracts the obligations from the source documents first, then drafts the policy against them, so what you get back is anchored in what the regulator actually wrote rather than in what the model remembers.
Both source documents are public, so pasting them in raises no confidentiality issue, and doing so is what keeps the output honest.
Ensure you fill in context and constraints and other aspects marked with {}.
Remember to adhere to the Golden Rules and do not upload confidential or privileged information to public tools.
You are assisting a partner or COLP of a law firm in England and Wales. Your task has two stages: first extract the obligations from the SRA's own documents, then draft a firm AI policy against them. The output is a draft for the partnership to consider and approve, and it is a planning aid only.
Source documents (paste the full text of each below, both are public):
- SRA warning notice, "Misuse of AI", published 17 August 2026: {paste}
- SRA guidance, "Effective supervision", including the AI sections added in June 2026: {paste}
- {Optional: anything else the firm wants applied, for example the Law Society's generative AI guidance, or your professional indemnity insurer's stated requirements}
Context to apply:
- The firm: {size, offices, number of fee earners, and whether there is a separate risk and compliance function or whether the COLP does it alongside a caseload}
- Practice areas: {list, and flag those with court-facing work or heightened confidentiality sensitivity}
- AI tools in use today: {name each, and say whether it is firm-licensed, individually subscribed, or embedded in another product such as a case management or research platform}
- AI tools you believe are being used without approval: {be honest, or say "unknown", which is itself a finding}
- Current AI policy: {paste it, or say "none", or "a paragraph in the IT acceptable use policy"}
- How work is reviewed before it leaves the firm, by role: {describe the actual practice, not the intention}
- What the partnership will not agree to: {e.g., "we will not ban AI outright", "we will not fund an enterprise licence this financial year"}
STAGE 1: EXTRACT THE OBLIGATIONS
1. Obligations table
Working only from the pasted documents, list every obligation the SRA states, with the Principle, Code paragraph or Rule it attaches to and a short verbatim quotation for each. Quote the regulator rather than paraphrasing it. Do not add obligations that are not in the text.
2. What the documents do not say
Identify where the SRA has deliberately left the method open, and where practitioners commonly assume a requirement that is not actually stated. This section exists to stop the policy committing the firm to controls the regulator has not asked for.
3. Gap analysis
Measure the firm's current position against the obligations table. For each gap, say whether it is a documentation gap, a process gap or a technology gap, since the three carry very different costs.
STAGE 2: DRAFT THE POLICY
4. Draft AI policy
Produce a policy the firm could adapt, using the sections below, and keep it short enough that a fee earner will actually read it:
a. Scope: who it binds, and that it covers AI embedded in other products as well as chatbots
b. Approved tools, and the route to get a new one approved
c. What may and may not be entered into each category of tool, expressed in terms of client confidentiality rather than in technical terms
d. Verification: the checks required before AI-assisted work leaves the firm, separating whether an authority exists, whether the citation is correct, and whether it supports the proposition it is cited for
e. Supervision and accountability: who signs off what, by role, and how that is recorded
f. Record keeping: what goes on the file, and what the firm would produce if the SRA asked how a particular document was made
g. Client communication: what clients are told about AI use, and when
h. Reporting: what a fee earner must do on discovering an error, and how the firm decides whether to self-report
i. Training and competence
j. Breach and consequences
Against each section, cite the obligation from the Stage 1 table that it answers. Mark any section you cannot tie to an obligation as a firm choice rather than a regulatory requirement.
5. Decisions only the partnership can make
Extract every point that needs a decision rather than a drafting choice, and present each as a discrete question with the realistic options and their trade-offs. Mark these as [DECISION REQUIRED] in the draft rather than choosing for the firm. Include at least: whether unapproved tools are prohibited or discouraged, whether AI use is disclosed to clients, and who holds final sign-off on court documents.
6. Implementation and evidence
Set out what the firm has to do beyond publishing the document, covering training, the file-level record, and the evidence the firm would rely on if its supervision arrangements were questioned. Give a realistic sequence rather than a wish list.
7. Review trigger
Say what should cause the policy to be reviewed before its next scheduled date, for example a further warning notice, a disciplinary outcome under this one, a new tool, or a change in what an existing tool retains.
Constraints:
- {Add firm-specific constraints, for example an insurer requirement, a client mandate on AI disclosure, a licence with time left to run, or a tool the partnership has already rejected.}
- Quote and cite only from the documents pasted above. If you think an obligation exists but it is not in the pasted text, say so and mark it as unverified rather than stating it.
- Do not invent regulatory requirements, SRA expectations, enforcement outcomes, case law, paragraph numbers or product capabilities.
- Do not state that the draft policy achieves compliance. Its purpose is to map obligations to firm practice and surface the decisions the partnership has to take.
- Write the policy in plain English and in the second person, and avoid defined terms where an ordinary word will do.
- This is a planning aid, not legal or compliance advice. The partners, the COLP and the firm's managers remain responsible for the policy and for everything done under it.How did we do?
Hit reply and tell me what you would like covered in future issues or any feedback. We read every email!
Thanks for reading,
Serhan, UK Legal AI Brief
Disclaimer
Guidance and news only. Not legal advice. Always use AI tools safely.
Recommended Newsletters
Below are a few newsletters that I recommend, for various reasons. Check them out!
