TL;DR

A scoping review by Dr Holli Sargeant of Cambridge, commissioned by the Nuffield Foundation, finds the justice system is tracking what AI does to processing times but not what it does to the people using it. Of 45 tools identified, 38 have no evaluation data in the public domain. Where participants were asked, perceived fairness fell sharply once AI entered the hearing.

The measurement gap

Sargeant examined the evidence accumulated since July 2025, when Whitehall began rolling out its AI Action Plan for Justice across administrative, civil and family settings. Efficiency and task performance are well covered. Almost nothing addresses comprehension, procedural legitimacy or lived experience — the conditions she argues any meaningful right to justice rests on.

Deployment has run ahead of scrutiny. Courts and tribunals lead with a dozen tools handling scheduling, transcription, listing, form processing and case management support. Civil law follows with eight, spread across small claims, negligence and legal aid work. Family law accounts for six. Suppliers advertise faster processing and higher satisfaction scores, but the review found those numbers self-reported and published without any stated methodology.

What happens to the litigant in person

The sharpest evidence concerns people representing themselves. Sargeant analysed over 3,800 questions put to a GPT-4-powered legal assistant by more than a thousand users. Seven in ten arrived with no factual context whatsoever. Under 4% combined grounding, specificity and structure the way a lawyer framing a question would. The review warns this reproduces existing inequality through a new route: better-educated, more digitally confident users simply get better answers.

The tribunal findings are starker still. Agreement that proceedings had been fair stood at 58% without AI and 22% once AI summarised the bundles. Confidence that the judge had what they needed to decide fell from 33% to 7.5%.

Looking forward

Sir Geoffrey Vos told an audience last week he expects AI to decide small claims. That direction of travel makes missing outcome data harder to defend, not easier. For UK firms putting AI into any adjudicative or assessment role, the transferable finding is that efficiency metrics can improve while the people subject to the decision trust it markedly less — and only one of those is currently being counted.