Overview
Executive summary
Artificial intelligence (AI) is being deployed across the justice system in England and Wales faster than its full effects on the public can be independently or transparently evaluated. Some deployments include user-centred design or pilot processes; other use, including the informal use of AI by judges, practitioners, and litigants in person, sits outside any such process. In either case, deployment is materially outpacing the production of independent evidence about its effects on the people the system serves.
In July 2025 the Ministry of Justice (MoJ) published its AI Action Plan for Justice, setting out a three-year strategy for embedding AI across the department and the wider justice system, and established a dedicated Justice AI Unit to oversee implementation. Judges across His Majesty’s Courts and Tribunals Service (HMCTS) now have access to Microsoft Copilot on the eJudiciary platform, supported by recent judicial guidance. Proprietary large language models (LLMs) are in use across government agencies and legal advice organisations. Speaking at Microsoft’s AI Tour in February 2026, the Lord Chancellor described the MoJ as one of the fastest-growing users of Copilot across government. Senior judiciary have publicly endorsed and modelled the use of generative AI in judicial work. Adoption is being framed, in policy and in practice, as a core priority to achieve “swifter, fairer, and more accessible justice”.
There is well-motivated ambition and optimism for AI to improve access to justice. A strong case can be made especially where the system is currently under the most pressure – court and tribunal backlogs are creating significant delays, the legal aid and advice sector are capacity-constrained, and the level of unmet legal need is high. Impactful opportunities are visible across various stakeholders of the justice system, and a gain realised for one group frequently depends on or produces a gain for another. For people facing everyday legal problems, AI could drastically improve accessibility. Public-facing legal information, triage, navigation, and intake tools – available at low cost, at any time, and in plain and multiple languages – may reach those whom conventional provision does not, and help them recognise a problem as legal and actionable. For the judiciary, practitioners, and court staff, significant opportunities arise for efficiency and capacity. Automating transcription, document processing, legal research, and case management may reduce administrative load and redirect scarce professional time towards complex, people-centred work – provided AI augments rather than displaces professional judgment. Better access to empirical insight and evaluation may assist justice administrators and policymakers in understanding how the justice system functions. Data-driven analysis of otherwise poorly utilised justice data could show where delay accumulates and legal needs go unmet, and could support more accountable and responsive policy.
At the same time, the public have a right to a justice system that is accessible, fair, and effectively meets their legal needs. Whether the system is delivering on that right under conditions of AI deployment is not only a normative question – it is also an empirical one.
This evidence scoping review, commissioned by the Nuffield Foundation as part of its Public right to justice programme, examined the available evidence on AI and data-driven technologies in administrative, civil, and family justice settings in the UK, drawing on international evidence where transferability could be argued. The review mapped 45 tools in use or in development, concentrated in internal-facing and generative AI applications, with publicly available evaluation for a small minority. It characterised the academic, policy, and grey literature by temporal distribution, publication type, domain, and methodology, and synthesised the evidence thematically across individual and group experiences and outcomes, societal factors including public trust and legitimacy, and system efficiency and effectiveness.
The central finding of the review is that AI evaluations have not yet measured the outcomes that matter most from the standpoint of the people the justice system serves. Efficiency and task performance are well represented and show important opportunity: a faster, more accessible system is itself a contribution to access to justice. However, fairness, comprehension, lived experience, differential impact, and procedural legitimacy are under-represented. Tools may be assessed as effective against the metrics by which they are evaluated while the conditions on which the public right to justice depends remain under-measured. Four structural gaps run through the findings:
A measurement gap. No study identified in this review examined the effects on an individual whose matter was triaged, advised on, or decided through an AI-assisted process in a UK justice setting. Evaluations examine the experience of system operators rather than the people whose legal problems are being addressed. They rely on self-reported time savings and satisfaction with outputs rather than independent assessment of whether AI tools improved the efficiency of the resolution of legal issues or the broader justice system. They test tools on narrow tasks in controlled conditions rather than in the complex settings where real-world legal problems arise. Related evidence reveals that certain demographic groups experience worse outcomes in the justice system, and court digitisation projects revealed uneven uptake of digital processes across user groups. Evaluations should measure specifically whether similar patterns of disadvantage or exclusion will arise in the deployment of AI tools. No UK justice-specific study has measured how AI-assisted processes affect individuals, or how effects differ across groups.
This not only is a gap in what has been studied but also reflects a structural limit in the data itself. Many people’s legal needs never reach formal advice or adjudication and so are not recorded in legal data – a lack which only increases in publicly accessible legal data. Legal data also rarely contains ground truth. It reflects only those issues that are recorded in formal stages, for example in court judgments. It can be difficult to distinguish systematic biases across groups from valid legal reasoning and differing institutional contexts. The people most likely to be poorly served, and to have limited access to justice, are likely the least visible to empirical analysis of the system.
A deployment gap. The tools studied in the research literature are increasingly not the tools used in practice. UK deployments are dominated by general-purpose proprietary models, which cannot be subjected to the structural inspection of weights, training data, and architecture in the way that open-source or domain-specific systems permit. Behavioural evaluation through benchmarking, red-teaming,1 and human-interaction studies remains possible, and is being developed, but further work is required. Evaluative research indicates that legal-domain LLMs are more reliable than general-purpose LLMs for UK legal tasks, and that smaller non-generative models can outperform general-purpose LLMs for some legal applications. The current direction of deployment appears to be shaped by commercial availability rather than by evidence about which designs best serve users with unmet legal needs. These issues also extend beyond formal deployments: publicly accessible general-purpose LLMs such as ChatGPT and Claude are increasingly used informally, by litigants in person and practitioners, introducing challenges that currently sit beyond the purview of formal justice processes but already shape how people encounter them.
A legitimacy gap. While evaluations focus on performance metrics, users primarily experience AI in terms of procedural fairness, dignity, transparency, human engagement, and perceived independent legal judgment. These dimensions are under-represented in the current evidence. Even limited assistive uses of AI in tribunal processes may reduce perceived fairness and dignity among members of the public. Attitudinal and experimental research of this kind offers valuable early insight into how people may respond to AI in justice settings, though it cannot substitute for evidence of how individuals experience AI-mediated processes in practice. Human oversight is perceived as vital to the procedural fairness of AI use, and is treated in policy and judicial guidance as the principal safeguard, yet no UK deployment evaluation identified in this review examines whether that oversight is effective, calibrated, or sustainable in practice. Judicial and professional concerns about deskilling and over-reliance are articulated but untested.
A transparency gap. Where evaluation exists, it is often internal, partial, or methodologically thin, and the dominance of proprietary models makes external evaluation more difficult. Internal and vendor-led evaluations cannot easily escape the actual or perceived risks of optimism bias, and the apparent effectiveness of some tools may reflect selective reporting rather than demonstrated performance. Independent evaluation capacity needs to grow to enable verification of claims and assumptions about AI’s opportunities and risks. The justice system has, historically, a less developed tradition of independent evaluation than other parts of the public sector, and building that capacity is itself a precondition for meaningful scrutiny.
The conditions the review identifies – rapid adoption of proprietary general-purpose models, reliance on self-reported metrics, untested human oversight, and internal evaluation – are present across other high-stakes public service contexts in which AI is being deployed, including criminal justice, social care, and benefits administration. The findings of this review are therefore likely to have application beyond justice settings.
AI is already shaping the conditions under which people encounter the justice system. There is reasonable ground to think that AI, designed and used well, can extend access and reduce friction for users who would otherwise be poorly served. Whether it will advance or undermine the public’s right to justice is, at present, an open question, and one the public is entitled to have answered.
