There is a quiet transformation happening inside British justice, and it is moving considerably faster than the public conversation about it. AI in UK courts is no longer a theoretical concern or a distant Silicon Valley export. It is here, operating now, embedded in risk assessment tools used by the Crown Prosecution Service, trialled in document review software at legal aid firms barely keeping their lights on, and quietly piloted in judicial decision-support systems that most defendants, and their solicitors, know nothing about.
I’ve spent several weeks examining what’s actually been disclosed by the Ministry of Justice, what the CPS has published about its own digital transformation programme, and what legal professionals are saying privately about what they’re encountering day to day. The picture that emerges is one of significant ambition, modest transparency, and almost no meaningful public debate.
What the CPS is actually using AI for
The Crown Prosecution Service has been trialling AI-assisted disclosure tools since at least 2023. Disclosure, the legal obligation to share evidence with the defence, has long been one of the most resource-intensive parts of criminal prosecution, and the CPS has openly acknowledged it as a pressure point. Their 2025 digital strategy references machine learning tools designed to accelerate the review of large volumes of electronic material, flagging potentially relevant documents rather than requiring prosecutors to read everything manually.
On the surface, this sounds sensible. The CPS handles enormous quantities of digital evidence in modern cases. But the question that doesn’t appear anywhere in their published documentation is this: when an AI tool flags something as “not relevant” and a prosecutor acts on that categorisation without re-reviewing, what is the accountability mechanism? The defence won’t know what was excluded. The judge won’t know. And the algorithm producing those categorisations is, in all but a handful of cases, a commercial product whose internal workings are protected as a trade secret.
Legal aid firms and the productivity trap
On the other side of the courtroom, the story is different in character but equally concerning in its implications. Legal aid firms, many of which have been running at the margin for years, the collapse of legal aid funding has been in slow motion since the Legal Aid, Sentencing and Punishment of Offenders Act 2012, are turning to AI tools out of sheer necessity rather than enthusiasm. Document summarisation, case chronology tools, and automated legal research platforms are being adopted because firms cannot afford the hours that proper preparation used to require.
I spoke to a criminal defence solicitor working in Manchester who told me, on condition of anonymity, that her firm had started using a commercial AI research tool earlier this year. “It’s not that we think it’s perfect,” she said. “It’s that without it, we physically cannot get through the caseload at the rates legal aid pays.” She was clear that the tool occasionally produced plausible-sounding but incorrect citations, and that supervision was essential. But she was equally clear that the supervision was often rushed. That is not a technology problem. It is a funding problem wearing a technology mask.
What the Ministry of Justice framework actually says
The Ministry of Justice published its principles for use of AI in the justice system in 2024. They are principles, not rules, and that distinction matters enormously. The document articulates commitments to fairness, transparency and accountability, but it does not establish enforceable standards, mandate disclosure of AI use to defendants, or require independent auditing of tools before they enter operational use. There is no requirement that a court be informed when a prosecution case has been shaped, even partially, by algorithmic review.
This is the transparency gap. Not that AI is being used, that is to some extent inevitable and not inherently problematic, but that there is no legal architecture requiring disclosure of that use to the people most affected by it. A defendant whose case was shaped by an AI disclosure tool, triaged by an algorithmic risk assessment, and processed through a predictive bail recommendation system has, under current frameworks, no right to know any of that happened.
Bias risks that nobody wants to name
Algorithmic bias in criminal justice contexts is documented extensively in other jurisdictions. The COMPAS recidivism tool used in American courts was the subject of a 2016 ProPublica analysis that found it significantly more likely to incorrectly flag Black defendants as future criminals. Britain’s legal establishment often responds to such examples by noting that the UK’s approach is different, more cautious. That may be true as a cultural disposition, but it is not yet true as a regulatory reality.
Risk assessment tools are already used in the UK context, particularly in probation and parole decisions. The Offender Assessment System (OASys), managed by His Majesty’s Prison and Probation Service, has incorporated algorithmic elements for years. What has changed in 2026 is the speed at which more sophisticated AI tools are entering adjacent stages of the process, often through commercial contracts where evaluation is superficial and post-deployment monitoring is negligible.
Training data is the core issue. If a predictive tool is trained on historical prosecution and conviction data, and that historical data reflects decades of disproportionate policing of certain communities, the model learns to reproduce those patterns. It does not know it is doing so. It simply finds the statistical regularities in what it was given. An AI system trained on London Metropolitan Police data from the stop-and-search era is not a neutral instrument of justice. It is a compression of past prejudice into a probability score.
The comparison to other fields is instructive. When engineers use Print Shape for rapid prototyping, the physical object produced is exactly what the machine was instructed to make. Judicial AI tools produce something murkier: outputs shaped by biases baked invisibly into training sets, and no physical prototype to inspect.
Where judicial use sits right now
The judiciary itself is the most cautious player in this space, and deliberately so. The Senior President of Tribunals and the Lord Chief Justice’s office have both issued guidance warning against AI use in drafting judgements, and emphasising that any decision must remain the product of human judicial reasoning. The Bar Council and Law Society have published similarly cautious positions.
But guidance is not prohibition, and the pressure on courts with enormous backlogs creates its own gravitational pull. Crown Court waiting times in England and Wales stood at over 67,000 cases outstanding as of late 2025, according to Ministry of Justice statistics. When a system is under that kind of strain, any tool that promises efficiency starts to look attractive regardless of its readiness. I’d argue that is precisely the environment in which corners get cut and oversight gets deferred.
This links to a broader pattern I’ve covered before. The use of AI hiring algorithms by British employers has demonstrated clearly that commercial AI tools enter consequential decisions with minimal regulatory scrutiny, and that the regulatory response follows years after the technology has already embedded itself. Justice cannot afford that lag.
What a genuine regulatory framework would need
If the Ministry of Justice is serious about responsible AI adoption, several things need to happen that are not yet on any published roadmap. First, mandatory disclosure: defendants must be told when AI-assisted tools have been used in the processing of their case, at every stage from initial charging decision to sentence recommendation. Second, independent pre-deployment auditing of any AI tool used in a criminal justice context, with published results that are not redacted for commercial sensitivity. Third, a right to contest: defendants should have a mechanism to challenge AI-generated outputs and demand human review.
None of this is radical. The ICO’s guidance on AI fairness, the Equality and Human Rights Commission’s work on algorithmic accountability, and academic legal scholars at institutions including UCL and the University of Edinburgh have all made versions of this argument. The gap is political will and legislative priority. Justice reform rarely wins elections. Algorithmic accountability in courtrooms is even less visible than the magistrates’ court backlog that has been quietly dismantling access to justice for years.
The technology is inside the building. The question now is whether the law will catch up before its absence causes harm we cannot easily undo.
Frequently Asked Questions
Is AI currently being used in UK criminal courts?
Yes. AI tools are being used at various stages of the criminal justice process in England and Wales, including disclosure review at the Crown Prosecution Service, risk assessment in probation decisions, and document analysis at some legal aid firms. Use within the judiciary itself in drafting judgements is officially discouraged but not formally prohibited.
Does a defendant have the right to know if AI was used in their case?
Currently, no. There is no legal requirement in England and Wales for prosecutors or courts to disclose to a defendant that AI tools were used in processing their case. The Ministry of Justice’s 2024 AI principles document sets out aspirations for transparency but does not create enforceable disclosure rights.
What is the risk of bias in AI criminal justice tools?
AI tools trained on historical criminal justice data risk reproducing historical biases, particularly those relating to race, class and geography. If training data reflects decades of disproportionate policing or prosecution patterns, the model will learn to replicate those patterns in its outputs, producing probability scores that appear objective but encode structural prejudice.
Which UK bodies regulate AI use in the justice system?
Responsibility is fragmented. The Ministry of Justice sets policy direction, the ICO covers data protection aspects, and the Equality and Human Rights Commission has jurisdiction over discrimination risks. There is no single regulator with a specific mandate to audit and approve AI tools before they enter criminal justice processes.
How does the UK's approach to AI in courts compare with other countries?
The UK has been more cautious than the United States, where tools like COMPAS have been used in sentencing recommendations for over a decade, often without meaningful scrutiny. However, the EU’s AI Act classifies AI use in justice as high-risk and mandates pre-deployment conformity assessments, placing the UK’s current framework behind what will soon be the European standard.