Paste, Wait, Waive: Legal Professional Privilege in the Age of AI

I wrote previously about the danger of pasting privileged material into AI tools, in Pasting Away Privilege: Lawyers, Clients and AI in England and the BVI. This article repeats that same message, but this time in the words of the Chancellor of the High Court himself.

At the City of London Law Society on 22 April 2026, Sir Colin Birss, Chancellor of the High Court, framed a question every solicitor should be able to answer before their next AI prompt. When a client, or their lawyer, uses a large language model, does anything about that exchange remain privileged?

The doctrinal starting point

Birss grounds legal professional privilege in the rule of law rationale endorsed in Three Rivers [2004] UKHL 48, drawn from Professor Zuckerman. Zuckerman’s argument is that it is only fair to say ignorance of the law is no excuse if everyone has the facility to discover what the law actually requires. The lawyer’s role, on this view, is to level the playing field between those who are legally sophisticated and those who are not.

Privilege, however, stops at the legal profession. In R (Prudential) v Special Commissioners of Income Tax [2013] UKSC 1, the Supreme Court refused to extend common law privilege beyond the legal profession to other advisers, including tax accountants. Any argument that a client’s exchange with an AI system should attract privilege has to confront Prudential first.

The confidentiality problem, and the case that crystallises it

Munir v Secretary of State for the Home Department [2026] UKUT 81 (IAC), usually cited simply as Munir, saw a panel of three Upper Tribunal judges (Lindsley, Keith and Blundell JJ) address two cases in which legal representatives had cited AI-hallucinated authorities. Buried within the tribunal’s general conclusions, at paragraph 60, sits the passage every practitioner needs to remember. The tribunal held that uploading confidential documents into an open-source AI tool, such as ChatGPT, places that material “on the internet in the public domain”, thereby breaching client confidentiality and waiving privilege. Such conduct, the tribunal added, might itself warrant referral to the regulator, and should in any event be referred to the Information Commissioner’s Office.

That is not a stray remark. It sits alongside a wider point on supervision. A solicitor who delegates work remains responsible for supervising it, must ensure fee-earners understand the dangers of relying on non-specialist AI for legal research and drafting, and is likely to be judged more harshly than a lawyer who fails to check their own work.

Why the doctrine gives the same answer

Confidentiality has always been a prerequisite of privilege. Even if the common law concept were extended to cover advice generated by AI, it could not attach to interactions with public systems that are not confidential, since one cannot assume confidentiality has been preserved.

The narrow safe harbour

Where a lawyer uses AI to help formulate advice, and provided the system is secure, Birss suggests it is hard to see how that could affect the client’s privilege. He draws an analogy with consulting a textbook or an online source. That is broadly the model the judiciary applies to itself. Judges are not prohibited from using AI, but they take full personal responsibility for what goes out under their name, and are directed to use only systems they are confident are secure, currently a protected Copilot deployment and an in-house HMCTS and MoJ build.

Practical takeaways

• Public large language models are not confidential recipients. On the current authority, uploading privileged material to one of them amounts to a waiver, and probably an incident reportable to the Information Commissioner’s Office.
• An exchange between a client and an AI system is not privileged. Prudential remains the answer to any argument otherwise, unless legislation changes the position.
• The safe zone is a properly assessed, secure system, supported by contractual and technical guarantees against training and retention on client data (the same standard the judiciary applies before permitting judicial use).
• Supervision is a training obligation with regulatory teeth. Munir makes this explicit: a failure to warn staff about the pitfalls of non-specialist AI is likely to result in a referral to the Solicitors Regulation Authority or another regulatory body.
• Update engagement letters and AI policies now, well before a paralegal, a junior colleague, or a client decides to save time on a Friday afternoon.

Birss’s tone throughout was one of cautious optimism, but the message on privilege is not ambiguous. One careless paste, and privilege is gone.