AI Explained
How AI is changing legal work and the way disputes are decided
Part 4 · Can a Machine Judge?
By Dr Peter Fields
THE 2-MINUTE SUMMARY
The judiciary of England and Wales allows AI to assist with judicial work within stated limits. The judicial office holder remains responsible for the decision, including the assessment of evidence and the exercise of discretion.
The American Arbitration Association offers an opt-in procedure for eligible documents-only commercial disputes. The parties can check the AI summary of their submissions. The software drafts a proposed award and a human arbitrator decides the case and issues the award.
UK data protection law now allows significant solely automated decisions in more circumstances, with safeguards and restrictions. The separate question of who may exercise judicial power remains.
The practical test is what the human actually did. A person who can explain why a piece of evidence mattered, and how they tested a draft against it, is exercising a different role from someone who merely signs the result.
Two ways to use a draft
A judge has a bundle to read and a decision to make. A tool might summarise the parties’ submissions or put the documents in date order. Those tasks still shape what the reader sees first. The harder question comes when the tool says what the evidence means and proposes who should win. Who has actually assessed the evidence?
The October 2026 Position Statement for the judiciary of England and Wales sets out the approach to adoption and oversight. AI can assist, while the judicial office holder retains the duty to reason, evaluate evidence and exercise discretion. The office holder must review, verify and approve any AI output used in judicial work. The statement describes pilots including summarisation, transcription, redaction and chronologies. The specified protections for approved tools include UK data residency, enterprise data protection and disabled web access. Read it alongside the separate guidance for judicial office holders. [1]
Why the source of the decision matters
Imagine a case in which an email looks decisive until you read the chain and discover that it answers a different question. Summarising the email can help the judge find it. Deciding what weight it deserves calls for an assessment of the competing arguments and the rest of the evidence. The parties need a fair opportunity to address the points on which the decision will turn. They also need reasons they can understand and, where there is a route of challenge, take issue with.
The Law Society’s foresight study raises a question about how much independent judgement a human can exercise at the end if software has already selected the material, framed the issues and laid out the reasoning. The study draws on interviews and discussions and identifies a risk for the future. Sir Geoffrey Vos, speaking in June 2026, considered the possibility of eventual machine decisions in some settings. He also identified constitutional and fair-hearing obstacles to replacing judges in England and Wales. [2, 3]
Research by Parasuraman and Manzey on decision aids explains why a person can miss errors or omissions when they rely too heavily on automated assistance. Applying that concern to a draft award is an inference. A practical response would be to assess the main issues and read the important documents before looking at the proposed outcome. [4]
Geek Box – When earlier cases point in different directions
Donoghue v Stevenson reached the House of Lords as a Scottish appeal. The majority held that the pleadings disclosed a basis on which the manufacturer could owe the ultimate consumer a duty of care. Lord Buckmaster read the earlier case of Winterbottom v Wright as support for denying a duty to a third party. Lord Atkin read it more narrowly, as a case about a duty said to arise from a contract. A tool might help find the authorities and the speeches. Deciding which reading should govern the case, and giving reasons for it, is the judicial work. [5]
What the parties can agree in arbitration
In November 2026 the American Arbitration Association extended its AI Arbitrator procedure to eligible two-party, documents-only business disputes. Both parties must agree to use it. Before the case goes to the human arbitrator, the parties can review the platform’s summary of their submissions and confirm it or give feedback. The parties see the final award but do not see the proposed award drafted by the software. The human arbitrator decides the case and issues the award. If a point emerges later in the process, how will the parties have the opportunity to address it? [6]
The Chartered Institute of Arbitrators’ guideline advises arbitrators to avoid delegating tasks to AI if that could influence their procedural or substantive decisions. It calls for independent checks of AI output on which they rely and addresses when to consult the parties about AI use. For an arbitration governed by English arbitration legislation, a challenge on the ground of serious irregularity under section 68 of the Arbitration Act 1996 requires substantial injustice. One statutory ground is failure to comply with the tribunal’s general duty under section 33. The question would be what happened in the particular proceedings. [7, 8]
Where data protection fits
From 5 February 2026, section 80 of the Data (Use and Access) Act 2025 widened the circumstances in which the UK GDPR permits significant solely automated decisions about individuals based at least partly on personal data. A decision is solely automated if there is no meaningful human involvement. Safeguards include giving information and allowing the person to make representations, obtain human intervention and contest the decision. Special-category data is subject to tighter restrictions. A controller cannot rely on the recognised legitimate interests basis for a significant solely automated decision. The judiciary’s Position Statement separately governs how AI may assist a judicial office holder, who must exercise the reasoning and make the decision. [1, 9]
How to use it better
Before using a tool in adjudication, identify exactly what you are asking it to do. If it is preparing a chronology, check the dates against the documents and identify which events are disputed. If it proposes reasons or an outcome, ask who has tested the reasoning against both sides’ submissions and how the parties could challenge a material mistake. The label on the software tells you much less than the procedure being followed.
Sources and notes
1. Judiciary of England and Wales (9 October 2026), Position Statement on the Judicial Use of Artificial Intelligence, ‘Core principles’, points 2 and 4, and ‘Tools’, points 9–10. See also the October 2025 guidance for judicial office holders. Position Statement • Judicial guidance
2. The Law Society (published 5 October 2026), The Future of Agentic AI in the Justice System, full foresight report, pp. 52–66, especially Table 9. Study
3. Sir Geoffrey Vos (4 June 2026), AI in Arbitration – Promises and Pitfalls, §§1–7. Speech
4. Raja Parasuraman and Dietrich H. Manzey (2010), Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors 52(3), 381–410. Review
5. Donoghue v Stevenson [1932] AC 562 (HL). See the speeches of Lord Atkin and Lord Buckmaster. Judgment
6. American Arbitration Association (9 November 2026), expansion of its AI Arbitrator to commercial cases, and the service’s eligibility and procedure. Announcement • Procedure and FAQ
7. Chartered Institute of Arbitrators, Guideline on the Use of AI in Arbitration, updated September 2025, arts. 8.2–8.4 and 9.1. Guideline
8. Arbitration Act 1996, ss. 33 and 68(2)(a), as amended. Section 33 • Section 68
9. Data (Use and Access) Act 2025, s. 80, inserting UK GDPR arts. 22A–22D. Commencement under SI 2026/82, reg. 2(j), with the saving for specified earlier decisions in reg. 5. Section 80 • Commencement regulations • ICO summary