Lammy backs faster AI rollout in UK courts with record funding

David Lammy’s decision to publicly back a faster roll‑out of artificial intelligence across England and Wales’ courts — coupled with a record funding settlement for 2026/27 — marks a pivotal moment in the justice system’s attempt to fix chronic backlog problems, but it also raises urgent questions about reliability, oversight and the risks of delegating sensitive tasks to generative models.

Background: the crisis that forced a rethink​

The criminal courts in England and Wales have been operating under severe strain for several years. Open caseloads in the Crown Court more than doubled after the pandemic, with projections showing the backlog could reach between 99,000 and 114,000 cases by 2029 if no structural reforms are made. Those numbers prompted the Government to commission Sir Brian Leveson’s Independent Review of the Criminal Courts and to pursue a package of measures collectively described as the Government’s “Plan for Change.”
At the same time, Ministers and senior judges have signalled that investment alone will not resolve the problem. The new funding settlement — a £2.785 billion package for courts and tribunals in 2026/27, with £287 million ring‑fenced for capital repairs — is the largest on record, and lifts previous limits on Crown Court sitting days so courts can run at maximum capacity. But Government leaders say money must be combined with modernisation and procedural reform to make a lasting dent in delays.

What Lammy said — and why AI is now part of the policy mix​

Speaking at the Microsoft AI Tour in London, Justice Secretary and Deputy Prime Minister David Lammy framed digital modernisation — including greater use of AI — as a pragmatic response to unacceptable delay for victims and witnesses. The speech aligned tech adoption with other reform proposals from the Leveson review: expanding judge‑only trials for certain lower‑harm offences, increasing magistrates’ sentencing powers, and shifting cases away from the most time‑consuming Crown Court jury trials. Lammy argued that AI tools could streamline routine processes such as scheduling, document summarisation and case‑management decisions, freeing human time for adjudication and complex legal judgment.
Those remarks were paired with binding budgetary commitments agreed between the Ministry of Justice and the senior judiciary: unlimited sitting days in the Crown Court for the next year, and three‑year certainty on funding through to 2028/29 for parts of the settlement. The political calculation is plain — marry extra resources with technology and process reform to hit backlogs quickly.

---

Leveson’s recommendations: where AI fits in the picture​

A short recap of Part I of the review​

Sir Brian Leveson’s first report set out an ambitious reconfiguration of how criminal cases are processed. Key measures include:

• Creating a Crown Court Bench Division (judge‑only courts) to hear “triable either‑way” offences up to a specified tariff.
• Increasing magistrates’ sentencing powers to relieve pressure on the Crown Court.
• Reclassifying certain offences and strengthening plea incentives to drive earlier guilty pleas.

The review frames these proposals as a package; Leveson warned that picking and choosing without complementary reforms would blunt the impact on delay.

Where AI was proposed to be used​

Leveson — and the Government’s implementation plans — specifically identified targeted uses of AI to reduce administrative friction and accelerate case processing. One of the more notable suggestions is that courts could use AI‑generated summaries of witness statements when making case‑management decisions. The rationale: well‑summarised witness material could let judges and administrators triage hearings more quickly, identify overlapping issues, and reduce the non‑essential time spent on reading long bundles. That proposal is explicitly designed to augment, not replace, judicial decision‑making — although the boundary between assistance and substitution is a central battleground.

The pitch: what proponents say AI can deliver​

Advocates for limited, governed adoption of AI in courts list several plausible benefits:

• Rapid extraction of factual timelines, names and core disputed issues from long written statements.
• Automated scheduling optimisation that reduces idle courtroom time and keeps trial timetables tight.
• Document triage for disclosure, allowing legal teams to prioritise genuinely relevant items.
• Administrative automation (forms, listing adjustments, boilerplate directions) that reduces clerical backlog and speeds case progression.

When combined with record funding and procedural reform, supporters argue these capabilities can materially reduce sitting‑day requirements and shorten the average wait for a hearing — particularly for cases that do not require the full resources of a jury trial. The Government’s public messaging stresses assistance rather than substitution: AI as a force multiplier for a chronically under‑resourced system.

---

The counter‑argument: why the legal profession and civil‑liberties groups are cautious​

Law Society and professional bodies​

The Law Society of England and Wales has publicly signalled reservations about using AI in court processes. Its vice‑president warned that while digital modernisation is welcome, AI is “not a silver bullet” and must demonstrably enhance access to justice, be reliable and ensure fairness. The Society emphasised that interpretation and the capturing of cultural nuance require human involvement, and cautioned that unverified automation could distort evidence and produce miscarriages of justice.
Other bodies — including the Bar Council, the Criminal Bar Association and civil‑liberties groups — have raised overlapping concerns. The core themes are consistent:

• The risk of hallucination (AI inventing facts or producing plausible‑sounding but false citations).
• The danger of embedding biases into processes that already have unequal outcomes.
• Problems of explainability and the difficulty of auditing complex model outputs in adversarial legal settings.
• The erosion of procedural safeguards if human oversight is thin or under‑resourced.

The constitutional argument: jury trials and public trust​

Beyond technical risks, critics focus on constitutional and cultural dimensions. Sir Brian’s recommendations that would scale back defendants’ right to elect jury trials — combined with faster case processing aided by AI — raise the spectre of a two‑tier system where speed efficiently displaces important participatory safeguards. Opponents warn that trimming jury access risks public confidence in the justice system if it is perceived to trade fairness for throughput. Parliamentary debate and judicial commentary on Leveson’s proposals repeatedly emphasise the need to preserve the jury for the gravest offences and to treat any changes as experimental with clear sunset clauses or robust review mechanisms.

---

The Copilot controversy: a live case study of AI failure in policing​

The debate about courts is not an abstract one: recent events show how generative models can actively mislead operational decision‑makers. An independent inquiry into the West Midlands Police decision to recommend banning Maccabi Tel Aviv fans from an Aston Villa match found that a Microsoft Copilot output — an AI hallucination — referenced a non‑existent match between Maccabi Tel Aviv and West Ham. That fabricated detail formed part of material used to justify the ban and only later emerged as an AI error after public and parliamentary scrutiny. Senior police figures initially denied using AI, then corrected the record and apologised to MPs. The episode has become a political flashpoint about the appropriateness of deploying unverified generative tools in high‑stakes civic decisions.
Why this matters to courts: if policing decisions that shape civil liberties can be influenced by hallucinations, the risk is amplified when the outputs feed into adjudication — where a wrong fact or an invented “precedent” could materially affect liberty, reputation and life outcomes. The Copilot case is shorthand for what can go wrong when AI outputs are consumed uncritically.

---

What the evidence says about reliability: hallucinations and error rates​

There is now a growing empirical literature showing that even purpose‑built legal AI tools hallucinate with non‑trivial frequency. A widely cited study from Stanford’s human‑AI research group found that specialist legal generative systems still produced misleading or incorrect legal propositions roughly 17% of the time under benchmarked testing conditions; general large language models fared worse on legal tasks. Industry trackers and compendia of sanction cases show a steady stream of court decisions where parties relied on AI outputs that turned out to be false, resulting in sanctions or reputational damage.
The practical consequence is clear: every automated summary, citation or factual extraction must be verifiable against authoritative source material. Verification is costly in time and expertise; if every AI output requires human re‑checking, the efficiency gains can quickly evaporate — and the system risks a transfer of liability to under‑trained or under‑resourced individuals who are asked to perform that verification under pressure.

---

Technical, legal and governance safeguards that must accompany adoption​

If the Government and judiciary intend to proceed with wider AI use, the following minimum safeguards should be considered and written into policy and procurement:

• Human‑in‑the‑loop requirement: every AI summary or recommendation used in case management must be certified by a qualified legal professional who accepts responsibility for its accuracy.
• Provenance and auditable logs: systems must retain immutable records of prompt inputs, model versions, response timestamps and the document sources used to generate any output.
• Retrieval‑augmented grounding (RAG): models must be configured to retrieve and cite specific, verifiable documents from court‑controlled repositories rather than relying on web‑scale hallucination‑prone completions.
• Versioning and change control: model and dataset updates require approval and staged rollouts with regression testing against held‑out legal benchmarks.
• Red‑team testing and external audit: independent adversarial evaluations should be mandatory before deployment into operational workflows.
• Privacy and data‑protection controls: witness statements and suppressed material must not be used to train models unless there is explicit legal authority and robust technical guarantees that the data will not leak.
• Liability framework: contracts and statutory instruments must clarify who is legally accountable for errors — vendors, the MoJ, judiciary or the individual lawyer/judge who relied on or certified the output.

These are not optional technicalities: they are the mechanics that determine whether AI helps or harms the lived reality of litigants. International precedents — including conservative practice notes issued by some Australian courts — show that the judiciary can and will restrict AI’s role if safeguards are absent.

---

Implementation pathways: five pragmatic steps the MoJ should take first​

• Pilot, don’t deploy: launch small, time‑bounded pilots limited to administrative tasks (scheduling, form generation, redaction assistance) with clear success metrics and independent evaluation.
• Build ground‑truth corpora: create a secure, curated corpus of court judgments, codes and rules to underpin retrieval systems, and prohibit external public web crawling for legal training data.
• Mandate independent verification: require that any AI‑assisted summary presented to a judge includes an attestation signed by a named lawyer or official confirming checks against source material.
• Fund oversight: allocate a portion of the 2026/27 capital and resource uplift to create a statutory AI oversight unit within the MoJ — independent, transparent and staffed with legal, technical and ethicists.
• Publish transparency reports: make anonymised metrics on hallucination rates, corrected outputs and near‑misses publicly available so oversight can be evidence‑based rather than rhetorical.

These steps put the emphasis on safe, measurable productivity gains rather than rapid wholesale substitution. Each step is designed to preserve legal safeguards while making realistic room for technology to reduce avoidable administrative burden.

---

Risks that the Government’s current framings underplay​

• Regulatory capture and vendor lock‑in: a close partnership with a single cloud vendor risks locking the justice system into opaque proprietary stacks and creates single points of failure. Procurement must insist on interoperability and on‑premise or government‑controlled options where needed.
• Mission creep: starting with summaries and scheduling, the system may drift toward automated risk assessment or predictive pre‑trial decisions without adequate trials or legal scrutiny.
• Inequality of access: if AI‑assisted processes speed up cases for represented parties while unrepresented litigants lack the means to verify AI outputs, justice outcomes could skew worse rather than better.
• Erosion of human judgment: repeated reliance on algorithmic summaries can deskill court staff and junior lawyers, weakening the human capacity to spot subtle context or cultural nuance.
• Legal standards mismatch: evidentiary rules and standards of proof were not designed for machine‑generated statements; courts will need explicit procedural rules about admissibility, provenance and challenges to AI outputs.

These are structural risks, not hypothetical edge cases. The Copilot affair in policing shows how quickly errors can propagate from tool output to policy decisions, amplified by organisational denial or poor transparency. Courts — which sit at the intersection of evidence and liberty — cannot afford to be the downstream receptors of the same mistakes.

---

International lessons and precedents​

Other jurisdictions have taken a measured approach. Some Australian supreme courts have restricted the use of generative AI for preparing affidavits and witness material, requiring explicit confirmation about machine use. Several US law firms now use RAG architectures and maintain human certifiers, and a body of case law documents sanctions where lawyers filed material containing AI‑generated fake citations. These precedents underline two truths: first, careful governance can enable useful legal AI; second, courts will hold professionals to traditional standards even when they use new tools.

---

Measuring success: what good looks like​

If AI is to be judged a success in this context, the metrics must be both operational and rights‑preserving. Successful deployment should demonstrate:

• Reduced median pre‑trial delay for targeted case types, without a measurable increase in appeal rates or overturned verdicts.
• A low, audited hallucination rate for any AI output presented to a judge (benchmarks set in advance and publicly reported).
• Evidence that AI adoption reduced administrative costs while not shifting verification burden to already stretched public defenders or litigants in person.
• Transparent accountability: clear incident reporting, timely correction procedures, and a record of remedial action when AI errors contribute to an adverse outcome.

These success criteria protect both the system’s efficiency goals and the central values of due process.

---

The politics of reform: trade‑offs and timing​

The Government has framed reform as a triage — blunt instruments to prevent the collapse of the criminal courts system. That framing exerts political pressure to move quickly and to prioritise throughput. But reformers should be wary of substituting speed for justice as an allocative choice. The next 12–24 months are politically sensitive: accelerated pilots and the new unlimited sitting days power create an appearance of momentum, but unless pilots are well governed, early mistakes will harden public scepticism and could invite judicial rebuke or legislative restrictions later.

---

Conclusion: prudence, not prohibition​

The combined package of record funding and a technology‑forward reform programme reflects an understandable urge to stop victims waiting in limbo. AI — when correctly architected, firmly audited and tightly bounded — has the potential to reduce clerical drag and help judges focus on core adjudication. But the technology’s known failure modes, documented hallucinations and real‑world policing fiascos make clear that governance is the operational risk. Without legally binding oversight, provenance, human attestation and independent audit, the very tools intended to accelerate justice could, in isolated but consequential cases, subvert it.
The way forward is neither an uncritical embrace nor a reflexive ban. It is a staged, transparent, and accountable programme: pilot the low‑risk administrative tasks; hardwire human verification for any output that touches evidence; fund independent oversight; and measure outcomes against both efficiency and fairness. That disciplined path offers the only realistic chance of using AI to speed justice while protecting the liberties the system is designed to uphold.

---

Source: Nation.Cymru Justice Secretary David Lammy to push for more use of AI in courts