Lammy's AI Push in the Justice System: Reform, Risks and Governance

David Lammy’s recent public push to widen the use of artificial intelligence across England and Wales’ justice system marks a decisive turn in how ministers propose to tackle chronic court backlogs — but it also exposes the reform to an array of technical, legal and ethical risks that have not yet been fully confronted.

Background​

David Lammy, the Secretary of State for Justice and deputy prime minister, used a high-profile Microsoft AI event in London to set out a sharper policy direction: expand pilots of AI in the Ministry of Justice (MoJ), scale up an “in-house justice AI” capability, and press ahead with reforms that will reduce the number of jury trials by moving a larger share of lower and intermediate cases to magistrates or judge-only sittings.
The stated rationale is straightforward: cut delay, free judicial time, and relieve pressure on overcrowded Crown Courts. The MoJ points to measurable administrative savings from trials already under way — primarily transcription and note-taking pilots in the Probation Service — and argues that automation can return officer time to front-line rehabilitative work while accelerating case progression in tribunals and lower courts.
These announcements arrtiny over how public bodies are using commercially supplied AI tools. A high-profile policing scandal last year — in which an AI “hallucination” produced false details that contributed to a decision to bar visiting football fans — has made MPs and civil society more wary of blindly trusting generative models. That episode, and subsequent Sky News reporting that at least 21 police forces continued to use Microsoft’s Copilot tool, sits uncomfortably alongside the MoJ’s enthusiasm for rapid adoption.

---

What Lammy said — the headline policy points​

• Lammy announced that the MoJ will increase investment in its internal “justice AI” unit to accelerate pilots and embed AI tools across probation, tribunals, magistrates’ courts and administrative workflows.
• He highlighted existing pilots that have used automated transcription in the Probation Service, which the department reports have transcribed more than 150,000 meetings and saved around 25,000 hours of administrative time. The MoJ said similar transcription tests are being extended to courts and tribunals, and that judges in the Immigration and Asylum Chamber have begun using AI to formulate notes and draft remarks.
• Lammy reiterated plans to reduce the number of jury trials, arguing that only a small fraction of criminal matters currently go to jury trial — a figure ministers consistently cite as about 3% of criminal cases — and that the bulk of cases are resolved fairly in magistrates’ courts. He framed the reforms as a continuity of past changes to case allocation, not a departure from the principle of a fair trial.

Each of these claims carries operational promise, but also invites scrutiny when the details are unpacked: what exactly will AI do, how will outputs be verified, and who remains accountable when the technology errs?

---

The pilots: transcription and summarisation — what they promise​

Where AI is already used​

The MoJ’s most mature pilots focus on automated transcription and document summarisation. These use-cases are natural first steps: converting audio recordings of meetings and hearings to searchable text, generating draft notes for legal advisers, and producing concise summaries to speed early case management decisions. The benefits are clear on paper: reduced admin overhead, faster case progression, and better information flows between agencies.

• Practical wins the MoJ cites:
• 150,000 meetings transcribed in probation pilots.
• ~25,000 hours of staff time saved from transcription alone.
• Early trials of transcription in courts, and summarisation tools used by some judges in immigration chambers.

Why transcription looks attractive operationally​

Transcription and summarisation are efficiency-focused tasks where current AI models have demonstrable competence. In many administrative environments the technology reduces repetitive typing, makes records searchable, and accelerates routine correspondence. For the justice system, accurate transcripts could shorten the time judges and legal advisers spend on minute-taking and drafting, permitting more sitting days to focus on adjudication rather than paperwork.
But operational competence is not the same as legal readiness. The acceptance of an AI transcript or summary in a justice context must rest on rigorous standards of accuracy, provenance, and verifiable chain-of-evidence — all of which are hard to guarantee at scale with current generative models.

---

The elephant in the room: hallucinations, fabrication, and police use-cases​

The Maccabi Tel Aviv / Aston Villa controversy was the public jolt that brought generative AI failures into the realm of national politics. An AI-generated or AI-assisted piece of content — later attributed to Microsoft Copilot in multiple reports — contained fabricated details about a match and contributed to policing decisions with real-world consequences. That scandal culminated in senior resignations and a parliamentary inquiry into the role of AI in operational policing.
Sky News reported that at least 21 police forces continued to use Microsoft’s Copilot despite the incident, which highlights a patchwork of governance across public bodies and inconsistent acceptance of AI’s limitations. The National Police Chiefs’ Council’s guidance has emphasized caution, but local force policies remain heterogeneous.
This is not a minor or theoretical risk. When AI outputs feed into risk assessments, tactical plans, or evidential packages, fabrication and incorrect association can distort decisions that affect liberty, safety, and reputations. In other words, the justice system’s stakes magnify the consequences of model failures.

---

Legal, ethical and constitutional implications​

The right to a fair trial and human decision-making​

Ministers have been at pains to stress that the right to a fair trial remains sacrosanct and that the reforms do not remove the right to a fair hearing. Yet fairness has many dimensions: procedural fairness, transparency, and the ability to challenge evidence. Introducing AI — particularly opaque, proprietary models — complicates each of these.
Legal practitioners and representative bodies have cautioned that automation cannot substitute for investment in courts or staff, and that decisions bearing on liberty and reputation should remain firmly in human hands. As Richard Atkinson, former head of the Law Society, observed, modernization must “enhance access to justice, be reliable and ensure fairness” — and courts must be wary of allowing AI to displace human judgment where rights are at stake.

Admissibility, disclosure and challenge​

The introduction of AI raises immediate questions about disclosure obligations and the ability to scrutinize how an AI output was produced. If an AI-generated transcript or summary forms part of a prosecution or case file, defendants must be able to know:

• which system produced it,
• what prompts or settings were used,
• whether any human editing occurred,
• and the error rates and provenance of the recorded data.

Without these disclosures, defence teams cannot meaningfully test the reliability of AI-derived material. That threatens the adversarial process where evidence must survive challenge, cross-examination or expert scrutiny.

Bias and differential impact​

AI models are trained on large datasets that reflect historical patterns, including social and systemic biases. When applied to policing, sentencing, or case triage, these patterns can reproduce or amplify unequal treatment. Any widescale rollout must therefore include comprehensive fairness audits, differential impact assessments, and accessible remedies for those adversely affected.

---

Technical governance: what the MoJ must build before scale​

If the MoJ’s in-house “justice AI” unit is to be more than a procurement arm, it must mature into a full governance engine with six core capabilities:

• Model validation and independent auditing — formalised, frequent tests of accuracy, calibration, and failure modes, carried out by external auditors.
• Data provenance and chain-of-custody — immutable logging of inputs, prompts, human edits, and timestamps for every AI-produced artefact used in casework.
• Explainability and disclosure standards — a minimum dataset for disclosure to opposing parties and the court about any AI-generated material.
• Human-in-the-loop controls — mandatory human review thresholds, with clear rules on what may be delegated to automation and what cannot.
• Red-team testing and adversarial evaluation — simulation of worst-case hallucinations, spoofed inputs, and deliberate attempts to subvert outputs.
• Governance across suppliers — standard contractual safeguards, liability allocations, and incident response obligations with commercial vendors.

These are not novel asks; they are the baseline controls demanded by regulators and independent reviewers in other sensitive domains. The speed of model improvement should not substitute for the depth of governance.

---

Operational challenges on the ground​

Skills, staffing and digital maturity​

Rolling out AI across magistrates’ benches and tribunals will require training judges, legal advisers, clerks and court staff to understand both the capabilities and the limits of the systems they will use. Digital maturity varies dramatically between courts: some have modern case management while others still rely on manual paper processes. AI tools layered on top of fragile digital infrastructure risk producing brittle outcomes.

Infrastructure and security​

Large-scale transcription and summarisation at the scale the MoJ anticipates demands robust computing infrastructure, secure data pipelines, and strict access controls. Sensitive personal data — including health and offence details — are frequently present in court records. Any cloud or third-party processing must be accompanied by rigorous security assessments, encryption-at-rest and in-transit, and clear data retention policies.

Cost, procurement and vendor lock-in​

The temptation to adopt off-the-shelf commercial models is understandable, but the MoJ must avoid vendor lock-in and opaque licensing that could hinder transparency. Procuring bespoke services with enforceable SLAs, audit rights, and the ability to extract data and models for independent review is essential. The department should also budget for the ongoing human oversight costs that automation creates, rather than assuming it will be a net saving.

---

International comparisons and precedents​

Across Europe and North America, governments have taken varied approaches to judicial use of AI: pilots focused on administrative tasks are common, but jurisdictions diverge on the extent of permitted automation in decision-making. The recurring lesson is that narrow, well-scoped uses (e.g., redaction, transcription) are the least risky near-term options; anything that impacts judicial reasoning or sentencing demands far stronger safeguards and often explicit statutory authorization.
Comparative scrutiny shows that countries which prematurely expanded AI into core decision-making have faced political blowback, court challenges, and costly rollbacks. The MoJ can learn from these cases by sequencing adoption: stabilise transcription accuracy and governance first, then cautiously broaden capabilities once independent evaluation validates safety and fairness.

---

Political and public trust dimensions​

The Maccabi-related policing scandal has eroded public confidence in some uses of AI within public bodies. For justice reform to succeed politically, the government must rebuild trust through transparent pilots, open evaluation reports, and clear lines of accountability when technology contributes to error.
Lammy’s framing — that AI can “smash through delays” and free up human time — is rhetorically persuasive. But political buy-in also requires demonstrating that AI will not quietly displace rights or obscure the reasoning that leads to judicial outcomes. Citizens will judge success not by lower waiting lists alone, but by demonstrable fairness and the capacity to contest AI-assisted findings.

---

Legal profession response and the civic balance​

Legal representative bodies are not opposed to modernization per se; instead they emphasise that justice reforms should not substitute technology for staffing, building maintenance, or funding. The Law Society and former senior figures in the legal profession have called for AI adoption only where it demonstrably enhances access to justice and where human decision-making remains central for consequential judgments.
Their practical concerns reflect a cautionary principle: reforms must be phased, transparent, and supplemented by adequate resources for defence and prosecution alike to contest AI-influenced materials.

---

Recommendations: a practical roadmap for safe adoption​

To reconcile urgency with prudence, the MoJ should adopt a staged, auditable program with the following steps:

• Consolidate and publish pilot results — release redacted datasets and evaluation metrics on the probation transcription pilot so independent researchers can validate the claimed 150,000 meeting transcriptions and 25,000 hours saved. Openness will build credibility.
• Establish an independent oversight board — include judges, defence counsel, civil liberties advocates, technologists and statisticians to review systems before wider deployment.
• Mandate disclosure in court — require courts to log the use of AI for any material in evidence or case management, including model version, prompt text, and human edits.
• Limit scope where rights are engaged — avoid permitting AI to make or replace judgments involving liberty, mental capacity, or credibility determinations. Use AI only for administrative augmentation unless proven beyond a high evidentiary bar.
• Commission regular, public audits — external auditors should publish findings on accuracy, bias, and incident reports (with appropriate redaction), at least annually.
• Invest in digital maturity and staff — pair AI adoption with concrete funding for court estates, staff recruitment, and training — otherwise automation will mask underlying resource gaps rather than solve them.

---

Accountability and liability — who answers when AI is wrong?​

One of the trickiest unresolved practicalities is liability. When an AI transcription contains an error that materially affects a sentencing timetable, or when an AI-generated summary leads to an evidential omission, who bears responsibility? The MoJ must clarify whether liability rests with the human supervisor, the vendor, or the department — and set contractual and statutory frameworks accordingly.
Contracts with vendors should include:

• indemnities for demonstrable model failures,
• robust incident-response timelines,
• audit and data access rights,
• and explicit obligations to remedy or retract affected outputs.

Accountability must be visible to maintain confidence that redress is possible and timely.

---

Conclusion​

The MoJ is right to seek efficiency and to experiment with technology that can reduce administrative burdens. Transcription and summarisation are sensible first steps that can save staff hours and potentially speed case progression. Lammy’s explicit commitment to invest in an in-house justice AI unit recognises that public-sector deployment requires sustained capability, not a scattergun adoption of private tools.
Yet the path from pilot to scale is treacherous unless accompanied by rigorous governance, full transparency, and statutory clarity about the role of AI in decisions that affect rights. The policing scandal that revealed a Copilot “hallucination” shows precisely why a cautious, evidence-driven approach is essential: technology that amplifies error can do outsized damage when applied inside institutions that make liberty decisions.
If the MoJ uses this moment wisely — publishing pilot data, opening systems to external audit, and pairing automation with investments in human capacity — it can achieve meaningful productivity gains without eroding the safeguards at the heart of the justice system. If it moves too quickly, or obscures the mechanics of how AI is used, the result will be public mistrust, avoidable legal challenges, and the very delays the reforms aim to cure.
The debate is not about whether AI will be part of the future of law — it will. The question is whether the Ministry of Justice will shape that future with transparency, accountability and rights‑respecting controls, or whether it will allow operational expedience to outpace the legal and ethical frameworks that must govern technologies with power over people’s lives.

---

Source: coastfm.co.uk Magistrates and judges to use more AI, says Lammy - as jury trials reduced