On 9 October 2026 the UK judiciary published a Position Statement on the Judicial use of Artificial Intelligence. Its approach is "cautious but positive". The central rule is that "AI tools must not determine the outcome of a case", and that "Judicial decisions will be taken by a judicial office holder." It is a sensible framework. Its real test is whether the human review it requires stays substantive when courts are under pressure.
What the statement actually does
The statement says AI may assist judges but cannot take over their reasoning, their evaluation of evidence or their exercise of discretion. Every AI output used in judicial work must be checked and approved by the responsible judicial office holder, who remains fully accountable. Only approved tools may be used. The tools are not listed publicly, which the statement attributes to the pace of change. The required features include UK data residency, enterprise data protection and disabled web access.
Governance is centralised. A steering group chaired by the Lead AI Judge, Sir Colin Birss, reviews new use cases, products and pilots for legal, ethical, operational, security and data protection issues before they proceed. The statement describes it as the sole oversight route for development, piloting and deployment, and as the escalation route if concerns arise. It also keeps the statement itself under review.
The pilots have been deliberately modest. The judiciary lists summarisation, transcription, redaction, writing in accessible language, administrative and leadership tasks, chronology-building and document analysis. They tested how reliable the tools are, how much human checking they need, and how sensitive material is handled. None of these tasks is adjudication.
The strongest case for tighter limits
The case for caution deserves a fair hearing. Courts exercise coercive state power, and errors in them are costly and hard to detect. Generative models can produce confident falsehoods. The judiciary's own October 2025 guidance, published on 31 October 2025, expanded its warnings on hallucinations and bias in training data. It reminded judicial office holders not to enter private information into public AI tools. It also stressed, in Lord Justice Birss's words, the "personal responsibility judicial office holders have for all material produced in their name." A sceptic would argue that a summary is never neutral, because deciding what goes in a chronology shapes how a case is understood. On that view, even "administrative" uses can quietly steer outcomes.
That argument is right that the line between assistance and influence is not sharp. It does not follow that the tools should be banned. It means oversight has to be designed for the way human reviewers actually behave.
Why the pro-innovation approach is the right one
The statement gets several things right. First, it starts with low-risk, verifiable tasks. Transcription, redaction and document analysis are labour-intensive, and a reviewer can check the output against the source document. Delays in the courts are themselves an access-to-justice harm, so refusing useful tools has a cost. Second, it routes decisions through one accountable body rather than leaving each judge to improvise. Third, it avoids prescriptive technical rules that would be out of date within a year. It sets principles and puts a standing review mechanism behind them. Fourth, the human-in-the-loop requirement attaches accountability to a named person, which is the right place for it.
The hard problem: oversight under pressure
The statement followed the Master of the Rolls, Sir Geoffrey Vos, who gave the John Lehane Memorial Lecture at the Supreme Court of New South Wales on 24 August 2026. As reported by Legal Cheek, Vos predicted that the public will come to accept machine-made decisions in some disputes. He suggested litigants may eventually prefer decisions "in minutes by a machine at little or no cost" if the machines prove at least as reliable as people. He drew a firm line between consent and compulsion: "Consensual machine-made decisions are one thing, and non-consensual ones are quite another." He said courts may be legally unable to impose machine judgments on unwilling parties, and he treated the right to a human judge as deeply entrenched.
He also identified the weak point in any human-sign-off model. Heavy caseloads, he warned, mean "economic and time pressures will inevitably lead to judicial approval of the machine's recommendations." That is the rubber-stamping risk. The position statement says outputs must be reviewed, verified and approved. It does not say how the judiciary will check that approval remains meaningful. A rule that a human must approve an output is only as strong as the human's practical ability to disagree with it.
The two documents are compatible. The position statement governs what judges do now. Vos was describing a possible future in which parties consent to machine resolution of small, algorithm-friendly disputes. Any such step would, by his own account, need informed consent and likely legislation. The statement does not pre-empt that debate, but it also does not answer it.
What would make this credible
Three additions would strengthen it without adding heavy regulation:
- Publish evaluation results. The statement promises pilots, evaluation and transparency. Reporting error rates and the share of AI outputs that reviewers changed would show whether review is real. A reviewer who never edits an output is a warning sign.
- Keep an audit trail. Recording when approved tools were used on a matter lets parties and appellate courts see where AI assisted, which supports open justice.
- Hold the line on consent. If machine-determined small claims ever arrive, they should be opt-in with an appeal to a human judge, as Vos himself suggested in 2023.
Conclusion
The UK judiciary has chosen a model that is neither a ban nor a free-for-all. The ban on AI deciding outcomes, mandatory human accountability, an approved-tools list and a single steering body is proportionate and evidence-led. What remains is empirical. Over the next year the question is whether the review step stays substantive, and that will show up in published pilot data rather than in the text of any statement.