lawyery
Two weeks ago Crystal and I relaunched the firm around a proposition we had arrived at by doing the work: the machine can draft, but it cannot take responsibility. We had no rulebook to point at when we wrote it. There was plenty of commentary, most of it either breathless or doom-laden, and very little from the regulator. So you build the systems you think are right, supervise the work, sign your name and hope.
On Monday the SRA published a warning notice on the misuse of AI. I have read it twice. There is no surprise in it. There is, for the first time, an authoritative statement of where the duty sits.
Welcoming an instruction from a regulator does not come naturally to me. I spend my working life advising people how to satisfy a system I do not much admire, and I have never been at ease with institutions telling professionals how to think. This one I welcome, and the reason is narrow. It settles an argument that was costing the profession more than the answer will.
The format matters. A warning notice is not general guidance. The SRA uses it for conduct it is prepared to act on, and this one ends by saying that failure to have proper regard to it puts you at risk of disciplinary action, and that the SRA can and will act where it finds contravention. It applies to every SRA regulated firm and individual.
Accountability does not move
"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."
There is nobody else to blame. You cannot delegate a duty to something that owes none. The model has taken no vow, sits on no roll and loses nothing when it is wrong. The notice then shuts the escape route: reliance on AI output is not a suitable defence when false citations reach a court.
That sentence needed writing. A good part of the market has been assuming the opposite, that an error produced by the tool belongs to the tool, and that the profession is a bystander to software it did not build.
Four tests, not one
The notice sets a standard for authorities put before a court. Anything cited must be genuine, must be relevant, must carry a verifiable citation, and must advance the argument being made.
Most AI policies I have seen stop at the first of those. The instruction to fee earners is some version of check the case exists. Genuineness is the easy test. It can be done at six on a Friday with a bundle due Monday, and it can be done by someone who knows nothing about the subject. Relevance, and whether the case actually helps you rather than merely sounding as though it might, survive a Google check untouched. A fabricated authority is the spectacular failure and it makes the papers. A real authority that does not support the sentence it is attached to is quieter, and I suspect far more common.
The notice grounds this in Ayinde v Haringey LBC [2025] EWHC 1383 (Admin), where false citations were placed before the Divisional Court and the judgment observed that a reference to the regulator is likely to be appropriate. It cites three 2026 authorities as well, including an Upper Tribunal decision on privilege and decisions in the Family Court and the Chancery Division. I have those from the notice rather than from my own reading, so check them before relying on them.
I used to do this myself
When I was a junior solicitor I cited cases that were borderline at best. Not invented. Real cases, correctly cited, which I had read. The failure was subtler than fabrication. Sometimes the proposition I was leaning on sat in obiter rather than the ratio. Sometimes the authority came from another jurisdiction and I had not asked whether a court here was bound by it, persuaded by it, or free to ignore it. Sometimes the case supported a narrower version of my point and I left the gap there, hoping nobody would look.
Nobody called that hallucination. It was called being a junior. The remedy was a supervisor who knew the area well enough to read the draft and tell me paragraph 47 was obiter and the argument could not hang on it.
I teach on Free Movement's advanced immigration course now, so I spend a certain amount of time telling people not to do what I did. It has left me with more sympathy for the trainee who overreaches than for the partner who did not read the draft. The trainee is learning. The partner is the control.
The four tests are not new law. They are what a competent supervisor has always applied to a junior's work, and paragraph 2.4 of the Code, which allows you to make only properly arguable assertions to a court, has been there all along. The SRA has taken a standard that lived in supervisory habit and written it down, at the moment the source of the draft changed.
The source changes one thing. A junior's overreach is usually visible. You feel the strain in the drafting, the case is cited a shade too triumphantly, an experienced eye snags on it. A model's overreach is fluent. It arrives formatted and confident, in the register of something already checked, and that fluency defeats the instinct that used to catch the error. There is a second difference nobody much wants to discuss. When I was corrected, I was also being taught. Nobody is teaching the model anything.
Where legitimate use ends
The four tests divide in two, and the division tells you who is capable of supervising.
Genuineness and a verifiable citation are mechanical. A paralegal can check them, and so, increasingly, can software. Relevance and whether the authority advances the argument are not mechanical. To know that the point relied on is obiter, that the Court of Appeal has since doubted it, or that the Upper Tribunal reads the same paragraph of the Rules differently, you have to know the area. No checklist gets you there and no tool does either.
Anyone building a practice around AI and looking at senior time as the saving should sit with that. The savings are in the first two tests. The last two are why you keep expensive people, and the notice has attached your name and your disciplinary exposure to them.
We are a boutique firm, which means the person who signs the work also writes the policy, answers the compliance question and takes the client's call at half past six. There is no committee to hide behind. The distance between what our policy says and what actually happens on a file is about one desk wide, and I would rather have that than a manual nobody reads.
It also means I can tell you that we have caught things in review which should never have reached review. That is what the review is for. Any firm claiming a spotless record here is either not using the tools or not looking properly.
So where does supervision end and legitimate use begin? Not at any particular quantity of AI, and firms will waste months trying to draw the line there. The workable test is whether the person signing could have written and defended the same paragraph without the machine. If they could, the machine saved them labour. We have done that with precedent banks and counsel's opinions for a generation and nobody called it a scandal. If they could not, the machine supplied judgement, and judgement cannot be handed to something that owes no duty.
Avoiding the irate judge is mostly unglamorous. Read every authority in full and in the original, and note the paragraph you are relying on. If the point sits in obiter, say so on the face of the document instead of hoping the other side does not notice. If the case comes from another jurisdiction, say that too, and say why the court should be persuaded by it. Check what has happened to the case since, because genuine and still good law are different questions. Decide in advance who may sign what, so that risk rather than availability determines who reviews. And if you find you have got it wrong, tell the court straight away. The Hamid line has always been harsher on the failure to correct than on the original mistake.
The obligations are filed under supervision
Look at where the notice puts things. Paragraphs 3.5 and 3.6 of the Code for Solicitors, on accountability for supervised work and the competence of those you manage. Paragraphs 2.3, 4.3, 4.4 and 8.1 of the Code for Firms, on accountability for work done through others, staff competence, effective supervision of client matters, and managers' responsibility for compliance. Paragraph 7.2, on being able to justify your decisions and actions.
There is no new AI rulebook. The SRA has said the existing supervision rules already cover this, which is harder to live with than a new rulebook, because it removes any argument that you were waiting for one.
Supervision must be proportionate to the complexity and risk of the work, and carried out by someone with the experience to do it. The Authorisation of Firms rules require supervision by a person who has practised as a lawyer for at least three years. A file showing a junior signing off a complex matter records a supervision failure with a signature on it.
Much of this is aimed at the firm rather than the individual. Governance structures, systems and controls are firm level obligations. One conscientious partner reading everything carefully is not a system. It is a person, and people go on holiday.
What it does not say
It does not prescribe a tool. It does not ban AI or discourage it; the SRA has authorised firms delivering legal services through AI and that has not changed. It does not certify anybody's compliance product.
I should be plain about my position. We built Countersigned, which runs AI-assisted work through a structured supervision process ending in a sealed record and a named solicitor's signature, and NativeLaw, which turns the same discipline into infrastructure other firms can adopt. Both came out of needing them inside a regulated firm, before there was any guidance to build against. So I have an interest in you thinking about this.
Over the next month a number of vendors, possibly including people who sound like me, will tell you the notice means you must buy something this quarter. It does not. A record can evidence supervision. It cannot perform it. Where the solicitor has not engaged with the draft, a well formatted audit trail documents an unread document with a signature attached, in a form that is disclosable. Read the notice, work out what your firm actually does today, and buy nothing until you can describe the gap in a sentence.
For two years the live question was whether supervision of AI was a professional duty or a good habit. It is settled now, in writing, by the body that can strike you off. What is left is what good supervision looks like on a Tuesday afternoon with three deadlines running and a client on the phone. No warning notice answers that. We will answer it by doing it, and by being honest about the times we get it wrong.
Profession is an old word and it means a vow said out loud. I have never minded that the vow comes with somebody checking. What I minded was not knowing what they would check.
Part two takes the other half of the notice, on confidentiality and privilege, which is getting far less attention and which in immigration practice matters more.
The SRA's warning notice on the misuse of AI was published on 17 August 2026. Quotations and paragraph references are taken from the notice as published on that date; guidance changes, so check the current version before relying on it. Nothing here is legal advice.