Artificial intelligence (AI) is moving rapidly into the UK’s justice system. Amid growing pressures, ministers see AI as a way to “deliver swifter, fairer, and more accessible justice for all”.
This ambition is reflected in rising government investment in AI tools for the courts and tribunals through initiatives including the Incubator for AI (i.AI) and Justice AI Unit. Meanwhile, the Ministry of Justice is one of the fastest-growing users of Microsoft Copilot across government.
Current policy presents AI as a route to a faster and more efficient justice system beset by significant backlogs and delays that are limiting access. But will AI make the system faster? And if it does, will the benefits reach those seeking justice?
In my report for the Nuffield Foundation’s Public right to justice programme, Artificial intelligence and justice, I reviewed the available evidence. However, AI is rolling out across the justice system faster than its effects can be independently evaluated, making it difficult for research to keep pace with the scale and speed of deployment.
Will AI make justice faster?
Some limited evidence on efficiency exists. For example, the i.AI reports that Caddy, a chatbot tool used by Citizens Advice advisers, wrote responses in half the time of an adviser.
Beyond Caddy, justice-specific evaluations are scarce. The following studies are drawn from other public sector contexts that are adjacent or comparable to the justice system.
Most evidence however comes from self-reporting. In a Home Office pilot, asylum caseworkers reported saving significant time when using AI to summarise interview transcripts and search for country information. A Department for Business and Trade study of Microsoft Copilot found small time savings on some tasks, while others took longer. Importantly, it found no evidence that self-reported efficiencies translated into departmental productivity gains.
Checking and correcting AI outputs can offset gains. Police officers saw no time saved when trialling AI-assisted witness statements after they had verified the outputs. In Ada Lovelace Institute research some social workers found that managers’ caseload expectations and longer documentation absorbed time saved. These examples illustrate that efficiency gains do not necessarily improve people’s outcomes or experiences.
Who benefits from AI in justice?
It is important to ask not only what AI can do, but who can benefit from it. Around two-thirds of the AI tools mapped in my report are used by professionals. Judges use them to write notes and summarise judgments; lawyers to research and draft; civil servants to process casework.
There are good reasons for this. Tools for legal experts are easier to design, manage and test against existing ways of working. Their risks are contained, at least in theory, by the expertise of the user. Building tools for the public is harder, and the caution shown by public bodies is in some ways justified.
Errors, oversight and bias
Expertise alone cannot mitigate all the risks of AI tools. In justice, small errors can have large consequences. AI-generated documents or AI-assisted summaries or transcripts can form part of a record that may follow a person across their life. There are three specific risks associated with AI tools that are particularly challenging in legal settings:
- Hallucination and error. Hallucination, where AI produces fluent but invented content, is built into how these models work. In one US study, specialist legal research tools still hallucinated on 17–33% of test questions. The Home Office’s asylum AI tools produced inaccurate outputs in about 9% of cases, and Microsoft Copilot hallucinations were only measured through staff reporting. A judge recently suggested the Home Office may have refused an asylum claim using “AI hallucinated” information.
- Human oversight. Policy treats humans checking AI outputs as a primary safeguard, yet the review found no UK evaluation that tests whether that checking works in practice. Judges and lawyers have raised concerns about over-reliance, but how professionals verify AI outputs remains untested. These risks are also high for the public: in one US study, people trusted an AI legal tool more after using it, overlooking its mistakes.
- Fairness and bias. Many AI tools learn from past data and can therefore mirror existing inequalities and perform less accurately for some groups. However, no UK study has tested this in justice settings. An Australian legal aid triage tool was evaluated as markedly less accurate at routing requests from older people and from Aboriginal and Torres Strait Islander users.
Finally, deployment of AI into the justice system may impact perceived legitimacy. UK research suggests the public is wary of AI in the justice system. When researchers asked people to imagine AI summarising the papers in a tribunal case, the share who thought that was fair fell to 22% compared to 58% when carried out by a tribunal judge.
AI is already shaping how people encounter the justice system. Whether it advances the public’s right to justice remains an open question. Answering it will require researchers, providers, government, and the legal profession to facilitate independent evaluation of AI’s effects on the system and the people it serves.
The challenge for policymakers will not simply be to roll out ever greater use of AI tools, but to demonstrate that they improve outcomes for people seeking justice, not just those administering it.
