Ministers, the courts service and senior judges are putting a growing share of digital justice in private hands, by default rather than by design. That is the argument of Delegated by Default?, a report by Dr Natalie Byrom for the Nuffield Family Justice Observatory (NFJO), setting out 26 recommendations grouped under four themes. Its summary calls the current model “a policy choice”. On 29 October, days before it appeared, the Civil Justice Council (CJC) published 15 recommendations on how parties should use AI in civil court documents. Read together, the two reports expose a gap: the CJC’s remit covered what lawyers and witnesses file in civil courts, but not the AI judges use, which is where the NFJO finds little public information. Our news report on the Byrom paper covers its headline warnings. This analysis asks whether “delegation by default” is something UK justice should reverse, and what reversing it would actually mean.
Strategic Insight: The report does not ask government to stop buying from private vendors. In its own words it is “not an argument against digital or AI-mediated justice, or against private-sector involvement”. What it wants reversed is the default, not the market.
How did the courts end up delegating by default?
The NFJO summary describes a widening handover to private suppliers, covering online legal information, dispute resolution and the software justice agencies run internally. It presents this as a departure from an earlier preference for technology the state built and owned itself. Three forces lie behind it, on the summary’s account: money pressures, the troubled court modernisation effort and how fast generative AI arrived. Regulation and governance, it says, have not kept pace.
The Law Society Gazette reports that Byrom sees the turn to the market as an “understandable consequence” of the HMCTS court reform programme failing to achieve its aims. Her verdict on the rules around it is harsher: in the Gazette’s paraphrase, the regulatory and governance framework for AI falls far short of what such reliance needs.
The real story is about accountability, not procurement
It is easy to read this as a familiar argument about outsourcing. The more pointed claim is about who can check the tools once they are in use. Judges, the NFJO notes, now have a locked-down edition of Microsoft Copilot. Yet there is “limited public information” on four points: how it was judged safe, what it is meant to be used for, who checks how it is used, and what a litigant can do to object.
The government’s own language shows how far Copilot has spread. Speaking at a Microsoft AI Tour event in London in February, David Lammy said the Ministry of Justice “has been one of the fastest growing users across government” of Copilot, as The Register reported. He also said AI was helping some judges hearing immigration and asylum cases draft notes and remarks.
| Measure | Figure | Strategic implication |
|---|---|---|
| NFJO recommendations | 26, across four themes | The report’s own count; it covers regulation, consumer tools, assurance and court governance |
| CJC recommendations | 15 | Focused on court documents from lawyers and expert witnesses, plus declarations by witnesses of fact, in civil proceedings |
| Family law providers on LawtechUK’s tracker | 8, of which 3 plausibly offer dispute resolution | The report adds that none is covered by legal services regulation, so few regulated alternatives to general chatbots exist for families |
| Extra funding for the Justice AI Unit | More than £12 million next financial year | Announced by Lammy in February, per The Register; investment is going into AI capability inside the justice system |
| References to Copilot in the judicial AI guidance | 9 lines in April 2025; none in October 2025 | Our count of the two published versions; the current guidance does not name the tool judges were given |
Strategic Reality: The April 2025 judicial guidance introduced Copilot Chat on judges’ devices through eJudiciary, saying it “provides enterprise data protection”. The October 2025 guidance that replaced it does not mention Copilot at all. Whatever the reason, the tool is no longer described in the public guidance addressed to its users.
What does the Copilot question actually involve?
The NFJO’s concern is specific. When general-purpose language models summarise or edit text, it says, they can get facts wrong, invent material and introduce “subtle shifts in meaning”. In family work one such shift is the flattening of what the report calls “justified asymmetry”, which could deepen a known pattern in which courts treat allegations of domestic abuse as two-sided conflict.
The usual safeguard is the judge as a “human in the loop”. The NFJO argues that this is unreliable in practice, because “the superficial fluency of AI outputs makes users less likely to scrutinise them even under incentives to check”. It also wants the ministry, the courts service and the judiciary to set out publicly the evidence behind “disregarding Microsoft’s own advice against using Copilot where it could affect legal status or rights”. The advice in question comes from the Application Card Microsoft publishes for Copilot, which the report says warns against uses that could affect a person’s legal status or rights.
Critical Context: The current judicial guidance states that judges “are not generally obliged to describe the research or preparatory work which may have been done in order to produce a judgment”. That is a long-standing principle, but it sits awkwardly with the NFJO’s call for safeguards that let litigants contest AI use. A litigant cannot easily challenge a use of AI they are not told about.
Where the CJC report fits
The CJC’s final report, from a working group chaired by Sir Colin Birss, deals with a different side of the same courtroom. It recommends a Practice Direction for civil proceedings built on a functional, risk-based approach, and a specific declaration from witnesses of fact that AI did not generate or influence the substance of their statements. Lawyers drafting pleadings and skeleton arguments would not have to make a general statement about their AI use.
The report is explicit about what it left out. Neither judges’ use of AI nor that of unrepresented parties fell within the remit it was given. On unrepresented parties it went further anyway, after strong consultation support for guidance, asking the CJC and the procedure rule committee to look at the question “as a matter of urgency”. Judicial use stays out of scope. Our earlier report on the CJC findings sets out the full recommendations.
If adopted, the CJC’s recommendations would require witnesses of fact to declare that AI did not shape their statements. Neither the CJC report nor the current judicial guidance mentions Copilot, the tool the bench itself was given.
Who carries the risk?
The groups the reports name are affected in different ways.
| Group | What changes for them | What is still unknown |
|---|---|---|
| Litigants in person in family cases | The NFJO says they are turning to generative AI for advice and documents, with limited redress when it gets the law wrong | Whether any civil-court guidance on AI for litigants in person will reach family proceedings |
| Survivors of domestic abuse | The report warns sycophantic tools “may affirm and enflame vexatious litigants and domestic abusers” and help them produce convincing documents at scale | How family courts will build the capacity the report says they need to tell weaponised litigation from genuine high-conflict cases |
| Judges and legal advisers | Answerable in person for anything issued under their name, under the current guidance | Which tasks Copilot is approved for, and who monitors it |
| MoJ and HMCTS | Asked to assess vendor lock-in and sovereignty risks before buying, and to build in-house capability | What that capability would cost |
| Legal technology providers | Byrom suggests a registration regime for UK-built AI tools, rather than turning legal advice into a reserved activity | Whether any future review of how legal services are regulated takes that route |
The report cites research in which users of sycophantic chatbots grew “more convinced of their own rightness and less willing to repair relationships”. The Gazette calls the implications for unrepresented family parties “particularly alarming”, and it is easy to see why: in a family dispute, a tool that hardens one side’s position works against the settlement the process is trying to reach.
⚠️ Warning: The Gazette reports that the paper cites studies finding people who search with generative AI reach official sources less often than those using ordinary search engines. The Gazette adds that efforts to build trustworthy support are undercut by the easy appeal of free chatbots, including Google’s AI Overview.
What would reversing the default take?
The four themes give a fair picture of what “deliberate” would mean in practice. According to Local Government Lawyer’s account of the report, they are: filling the regulatory holes left by chatbots and UK-built tools that offer legal help, with a new national commission to oversee AI in justice; fixing a failing consumer market by funding reliable public legal information and setting quality standards; making court AI visible, with evidence, lists of approved tools and pilot results published and independent scrutiny on top; and running court AI under shared standards, with monitoring, disclosure and a route for litigants to object.
None of that removes private vendors. It changes who sets the terms. That is our reading of the angle this article set out to test: the case the report makes is mainly for reversing the absence of public standards. It also wants less dependence on a few foreign developers, but it does not argue against private supply.
Challenge 1: The drivers have not gone away
The summary names financial pressure and an underperforming court reform programme as reasons the system turned to the market. The report’s answer includes an oversight committee for judicial AI with a budget of at least £1 million a year, scaling up in-house capability and reducing dependence on what the Gazette quotes as “a small number of foreign AI developers”. Building capability costs money, and fiscal constraint was one of the original drivers. The one recent public figure, the Justice AI Unit’s extra £12 million, was announced in February, months before the report, and comes from the same Copilot-using ministry the report is asking to change course.
Challenge 2: Two frameworks, one courtroom
The NFJO wants the ministry’s AI Action Plan brought into line with the principles set by HMCTS and the Judicial AI Working Group, and with the judiciary’s own position statement, giving “a coherent rather than competing framework”. The CJC report shows why that is harder than it sounds: it says HMCTS runs court administration only, and that guiding how participants behave is a job split between judges, the courts’ inherent jurisdiction, the Lord Chancellor and the procedure rule committees. A single framework has to work across all of them.
Challenge 3: Contestability without disclosure
The report wants litigants to have a way to object to AI use in their case. The judicial guidance does not generally require judges to describe their preparatory work, and the report notes that judges face no requirement to disclose Copilot use. Its answer is guidance from the Lady Chief Justice requiring judges to document and disclose AI use. Until something like that exists, a right to challenge AI use will be hard to exercise.
Reality Check: The CJC has committed to checking, in its next yearly review, how much of its report has been acted on. The NFJO’s recommendations are calls on ministers, the courts service and judges, so they depend on those bodies choosing to act.
What to watch
Byrom’s report was published in the week Sir Colin Birss became Master of the Rolls, according to the Gazette, which describes him as an AI champion. He chaired the CJC working group and is named on the current judicial guidance as Lead Judge for Artificial Intelligence, so he has been close to both strands this article compares.
Four things would show whether “delegation by default” is being replaced by design:
- AI guidance for unrepresented parties. The CJC wants it considered urgently for civil proceedings. Guidance that also covers family proceedings, or a parallel effort for the family courts, would show the NFJO’s family-specific warnings have landed.
- A published Copilot evidence base. Disclosure by the MoJ, HMCTS or the judiciary of how Copilot’s safe use was assessed, which tasks it is approved for and how it is monitored would answer one of the report’s main transparency requests. Silence would confirm its diagnosis.
- The next revision of the judicial AI guidance. A version that names the tools judges are given, and says what litigants can expect to be told, would move towards the contestability the report asks for.
- Procurement language. Lock-in and sovereignty assessments appearing in HMCTS or Justice AI Unit procurement would show the first theme turning into practice.
The question for UK justice is no longer whether to use AI or whether to buy it from vendors. Both are happening. It is whether the public, and the people whose cases are decided, get to see the standards those tools are held to. In our view, the past fortnight shows where attention has gone first: to what parties file. The tools on the bench have yet to get the same treatment.
Sources: Delegated by Default? Privately owned technology in the family justice system by Dr Natalie Byrom (Nuffield Family Justice Observatory, November 2026), full report PDF, also reported by the Law Society Gazette and Local Government Lawyer (Lottie Winson); the Civil Justice Council’s final report on AI for preparing court documents (29 October 2026); the judiciary’s AI guidance for judicial office holders (April and October 2025 versions); and The Register (26 February 2026).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the courts publish the evidence behind their own AI tools as readily as they now ask parties to declare theirs. Read more analysis at Insights, or get in touch.