On 18 May 2026 the Bar Standards Board published its first dedicated guidance on the use of artificial intelligence at the Bar. It bans nothing, blesses nothing, and creates no new duties — and that is precisely what makes it demanding. Here is what it actually asks of practitioners, and where the operational burden now sits.
The guidance arrives with unfortunate timeliness. In Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin), two matters were referred to a Divisional Court and listed together under the court’s Hamid jurisdiction [2] — in the first, a wasted costs application made on the grounds that those acting for the claimant “had cited five fake cases” [43]; in the conjoined Al-Haroun matter, witness statements relying on numerous authorities, many of which could not be found to exist. The court’s framing was not subtle: artificial intelligence is “a tool that carries with it risks as well as opportunities”, and the administration of justice depends upon the court being able to “rely without question on the integrity of those who appear before it” [5].
The President of the King’s Bench Division put the technical point in terms no practitioner can now claim to have missed: freely available generative AI tools “are not capable of conducting reliable legal research” [6]. The court set out the armoury available where that warning goes unheeded — up to and including proceedings for contempt under CPR Part 81 [27] — and stated plainly that lawyers who do not comply with their professional obligations in this respect “risk severe sanction” [69].
Then in March, a recorder at Bournemouth Family Court considered whether her judgment should be published and the lay advocate before her named [27], describing the case as “another example where AI hallucinations have led to the court being misled” [86]: Re A, B, C, D (Extension of assessment; Use of AI: hallucinations) [2026] EWFC 71 (B) — our analysis of that decision is here. Public trackers of court-recorded AI incidents in the United Kingdom now run to dozens of entries. The Law Society published its own warning to solicitors this week. The regulators have stopped treating hallucinated authority as a curiosity.
What the guidance requires
Four obligations do the real work.
Competence in the tools themselves. Under Core Duty 7, the guidance expects barristers to maintain “a sufficient level of competence in technology and AI” — not expertise, but enough to understand what the tools in their workflow actually do, and enough to understand the AI used by instructing solicitors, opponents and the courts. Ignorance of how a language model behaves is no longer a professional posture available to the Bar.
Transparency with clients where it matters. Under rC19, barristers must be transparent about their use of AI where that use “materially impacts the nature or scope of your service”. The guidance leaves materiality to judgment, which is where the difficulty will live: a drafting assistant probably falls below the line; research synthesis that shapes the advice probably does not.
Caution amounting to prohibition on free public tools for client work. The guidance notes that free generative tools’ terms typically allow providers to retain and use input data, including for model training. For material attracting legal professional privilege, that sits somewhere between an outsourcing-rules problem and a confidentiality breach. The practical effect is that the free tiers of general-purpose chatbots are unsuitable for client work — full stop.
Responsibility that does not delegate. The line that matters most is the simplest: inaccuracies introduced through AI remain the barrister’s responsibility. Systems that act with autonomy are singled out — agentic AI features “are likely to be of high risk for most applications at the Bar”. There is no safe harbour for having trusted the machine.
What it deliberately does not do
The guidance creates no checklist, certifies no tools, and offers no compliance formula. This is consistent with how the BSB regulates everything else: existing duties — competence, integrity, not misleading the court, confidentiality — applied to a new instrument. The Bar Council’s parallel guidance, updated alongside, takes the same shape. Some will read this as regulatory timidity. The better reading is that it locates the obligation exactly where it has always been: the practitioner vouches for what is put before a court, whatever produced the first draft.
The verification gap
Which leaves the operational question the guidance cannot answer: how. The duty to verify is clear. The cost of verification is the variable.
It is worth being precise about why the problem exists at all. A language model asked for a citation will always produce one, real or not — fluency is the objective; the truth of the output never enters the calculation. That is not a defect to be patched out of next year’s models; it is what generation is, and it is why the Divisional Court’s description of such tools as “not capable of conducting reliable legal research” [6] is an observation about architecture, not about this year’s products. The only structural fix is verification between the model and the reader: every citation checked against the official record before a human relies on it, every proposition traceable to the paragraph of the judgment said to support it.
Done by hand, that is the familiar discipline — pull the authority, confirm it says what is claimed, confirm it is still good law. Nobody at the Bar needs to be taught it. The question the last twelve months have posed is narrower: whether the checking happens every time, including the night before a morning hearing, and including by the lay advocates and litigants in person whom the courts increasingly receive. Every incident in the public record so far is a case where it didn’t.
Practical points
A barrister wanting to stay on the right side of this guidance needs five things in place: knowledge of which tools in their workflow generate content as opposed to retrieving it; a confidentiality posture for anything client-related; a verification habit that treats every machine-produced authority as unverified until checked against the record; a transparency judgment for when AI materially shapes the service; and an honest answer ready for the question a judge is now entitled to ask — how was this research produced and checked?
The standard the profession should converge on — and should demand of its tools — is that unverifiable material never reaches the practitioner at all. That is the design principle behind Search the Law: authorities are checked against The National Archives’ record before they render, linked to the paragraph relied upon — as the citations in this article are — and anything that fails verification is excluded rather than presented with confidence. The machine’s job is the mechanical part of checking. The lawyering — reading the case, judging its weight, deciding what it proves — was never the machine’s to do.
A postscript on method: in writing this article, three widely-reported quotations from the BSB guidance failed verification against the source document and were corrected to the regulator’s actual words. The discipline applies to commentary too.