AI Is Already in the Courts: Why Family Justice Needs Hard Red Lines

The short version

• Sir Geoffrey identifies three realities: legal professionals must use AI responsibly; autonomous judicial decisions should not be imposed without informed consent; and the law must clarify responsibility for AI-caused harm.
• He is right that AI can help litigants in person organise and communicate their cases. But a more polished document is not necessarily a more accurate, relevant or evidentially sound case.
• “Human checking” is a safeguard only when the human has the time, competence, source access, independence and authority to challenge the machine.
• In children cases, adult consent cannot be the whole answer. The child may not be a party, cannot simply be treated as having consented, and the court retains an independent welfare duty.
• Family justice needs clear red lines, disclosure of material AI involvement, secure audit trails, meaningful human review, tested systems and equal access to reliable tools.

1. The three realities in the lecture

Sir Geoffrey’s argument is more careful than the headline “judges must embrace AI” might suggest. He identifies three connected realities.

1. Lawyers and judges must adopt AI. He points to useful work including summaries, chronologies, transcripts, translations, drafting improvements and preliminary research. Used properly, these tools can reduce cost and delay.
2. Autonomous judicial decisions must not be imposed without fully informed consent. The rule-of-law concern is the right to an independent and impartial human tribunal, reflected in Article 6 of the European Convention on Human Rights as given domestic effect by the Human Rights Act 1998.
3. The law must give clearer answers about AI-caused harm. Developers, deployers, professional users and individual users may each bear responsibility in different circumstances. Sir Geoffrey also raises the more uncomfortable future question: could a professional eventually be criticised for failing to use AI where its use would have protected a client?

These propositions are not contradictory. We can support useful technology while resisting the transfer of judicial power to a system that cannot itself hold office, explain its conscience, accept legal responsibility or be disciplined for failing in its duty.

The difficulty lies in the space between them. A machine does not need to deliver the final words of an order to exercise decisive influence over the outcome.

2. Why the access-to-justice argument matters

One of the most important parts of the lecture is its frank recognition of what litigants are already doing. Sir Geoffrey says, in deliberately universal language, that “every litigant” now puts a legal problem into a mainstream AI system before approaching a lawyer. That may be rhetorical rather than measurable, but the underlying reality is undeniable.

For a parent who cannot obtain legal aid and cannot afford private representation, a chatbot is available immediately. It does not charge by the hour. It can explain unfamiliar words without impatience, turn scattered notes into headings and help somebody begin when stress has made the blank page feel impossible.

Sir Geoffrey also reports that judges are seeing AI-assisted pleadings and skeleton arguments from litigants in person that are of “higher quality” than many documents received before these tools became widely available.

That observation deserves to be taken seriously. Poor presentation has always distorted access to justice. A legally relevant point can disappear inside an emotional narrative. A parent may possess important evidence but not know how to identify it, date it or connect it to the issue before the court. A well-used AI assistant can help with structure, plain language and organisation.

This is why banning AI use by litigants in person would be unrealistic and unfair. It would remove a tool from the people with the least support while leaving well-resourced organisations and legal teams able to purchase sophisticated systems.

3. Access to filing is not access to justice

I part company with the lecture where it appears to treat an increase in the number of claims as, in itself, improved access to justice.

More people being able to start proceedings may reveal previously unmet legal need. It may also produce claims that are unfocused, legally misconceived or far larger than the dispute requires. In a system already struggling with delay, volume can displace rather than deliver justice.

A document can be immaculate and still be wrong. It can contain twenty authorities and still fail to identify the actual test. It can sound restrained while exaggerating the evidence. It can reproduce every possible complaint—the so-called “kitchen sink” approach—while making the decisive welfare issue harder to find.

Real access to justice includes the ability to:

• understand the process and the legal test;
• identify the issues the court must actually decide;
• place relevant and reliable evidence before the court;
• participate on an equal footing;
• receive a fair and reasoned decision within a reasonable time; and
• understand, challenge and comply with the resulting order.

Those are also reflected in the Family Procedure Rules’ overriding objective: cases should be dealt with justly, fairly, proportionately and expeditiously, with the parties on an equal footing and with regard to welfare issues.

AI can help meet that objective. Uncontrolled AI-generated volume can frustrate every part of it.

The relevant measure is not how many pages were generated or how many applications were issued. It is whether the technology enabled a fairer, clearer and more reliable route to a lawful outcome.

4. Why responsibility alone is not a safety system

Sir Geoffrey describes the classic hallucination objection as flawed. His essential point is that a lawyer remains responsible for inaccurate material whether the error came from the lawyer, a pupil, a judicial assistant or an AI system. On that point, he is plainly right.

A machine is not an excuse. The person who signs, files or relies on the work remains accountable.

But responsibility after failure is not the same thing as prevention before failure. “Human checking” is not a magic phrase. It is an effective control only where the reviewer:

• has enough subject knowledge to recognise a plausible error;
• can access current and authoritative legal sources;
• has enough time to check the source rather than merely reread the summary;
• understands how the particular system can fail;
• is independent enough to reject a convenient answer; and
• remains personally engaged with the evidence and reasoning.

Those conditions cannot be assumed. An unrepresented parent may not know how to locate a judgment, identify whether it has been overturned or distinguish an English authority from an American one. A busy professional may be tempted to check a sample rather than every proposition. A judge facing a large list may unconsciously treat a machine-generated recommendation as the starting point and look for reasons to approve it.

The judiciary’s own AI Guidance for Judicial Office Holders is candid about the risk. It says public chatbots are poor tools for discovering legal information that the user cannot independently verify. It warns that wrong answers are not infrequent, sources may be invented, the law may be outdated or drawn from another jurisdiction, and apparently persuasive text may contain obvious substantive errors on closer inspection.

The rule that humans remain responsible is essential. It is the beginning of governance, not the end of it.

5. What the family cases already tell us

This is not a theoretical concern imported from commercial litigation.

In D (A Child) (Recusal) [2025] EWCA Civ 1570, a mother acting in person used AI when preparing a lengthy skeleton argument. Some authorities were relevant, others did not support the propositions advanced, and some did not exist. The Court of Appeal absolved her of an intention to mislead. Baker LJ recognised that it was “entirely understandable” for litigants in person to seek AI help, while making clear that represented and unrepresented parties alike must ensure that cited cases are real and support the point for which they are used.

That is the correct balance: realism and education, without pretending that an invented authority becomes harmless because the person who relied upon it was unrepresented.

By contrast, the Solicitors Regulation Authority’s warning notice on misuse of AI, published on 17 August 2026—one day before Sir Geoffrey’s lecture—refers to Re A, B, C and D (Extension of Assessment; Use of AI: Hallucinations) [2026] EWFC 71 (B). In those care proceedings, an unregistered barrister who held herself out as a lawyer placed hallucinated authorities before the court. The court did not find a deliberate intention to mislead, but was concerned that the seriousness of misleading the court had been minimised.

The distinction matters. A vulnerable parent using a free tool because legal support is unavailable is not in the same position as a person selling or presenting legal expertise. The court must protect its process in both situations, but regulation, education and sanction must be proportionate to role, knowledge, conduct and harm.

The SRA warning also follows Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), where false citations placed before the court led to serious professional consequences and referrals to regulators. Its message is simple: reliance on AI does not answer a failure to verify.

These cases show why the debate cannot be reduced to “AI good” or “AI bad”. The relevant questions are who used it, for what task, with what safeguards, in what setting, with what ability to check, and with what effect on the proceeding.

6. The “human in the loop” problem

The most important part of Sir Geoffrey’s lecture may be his warning about machine-assisted judicial decisions.

A system can retain a “human in the loop” on paper while removing meaningful human judgment in practice. If AI analyses a file, selects the relevant facts, ranks credibility indicators, recommends an outcome and drafts the reasons, what remains for the judge who approves the result?

Sir Geoffrey himself imagines a judge presented with 1,000 machine recommendations. Economic and time pressure, he says, may lead to their routine approval. That is not a remote technical concern. It is a description of automation bias at institutional scale.

The problem is not solved by placing a human name at the bottom of the order. The human must have:

• direct access to the underlying evidence;
• enough time to engage with the disputed material;
• authority and confidence to reject the system’s recommendation;
• an explanation of the system’s limits and the route by which it reached the output;
• a record sufficient to identify error after the event; and
• personal responsibility for independent reasons.

The judicial guidance rightly states that “Judges must always read the underlying documents.” It also says that judges are not generally required to describe the research or preparatory work used to produce a judgment.

That second position will need refinement as AI becomes more influential. A spellchecker does not need to be disclosed. Nor should every administrative use create procedural clutter. But where AI materially summarises contested evidence, classifies allegations, evaluates risk, proposes findings, recommends an outcome or drafts substantive reasoning, the parties should be told enough to understand and, where appropriate, challenge that involvement.

The current ICO guidance on significant automated decisions affecting children concerns data protection rather than the judicial function, but its principle is instructive: human involvement must be active and meaningful, not tokenistic. Family justice should demand at least as much where the decision affects a child’s life.

7. Why consent is different where children are concerned

Sir Geoffrey’s proposed red line is that there should be no machine-made judicial decisions without the parties’ freely given and fully informed consent. That may offer a sensible starting point for competent commercial parties choosing an arbitral process.

It is not enough for family justice.

First, the child whose life is affected may not be a party to private law proceedings. Agreement between the adults does not amount to the child’s informed consent.

Secondly, under section 1 of the Children Act 1989, the child’s welfare is the court’s paramount consideration when determining questions about upbringing. Adult agreement cannot displace that independent obligation.

Thirdly, family proceedings may involve domestic abuse, coercive control, fear, financial inequality, disability, trauma, language barriers or a desperate wish to end delay. A person may “choose” an instant machine decision because the alternative is another year without contact, another unaffordable hearing or another period of uncertainty. That is not the same type of choice as two businesses agreeing an arbitration clause.

Fourthly, some family decisions require moral and constitutional judgment, not merely prediction. The court may need to assess proportionality, risk, credibility, the child’s wishes and feelings, the possibility of change and the least interventionist lawful outcome. Historical data can show what courts did before. It cannot determine what justice requires in this child’s circumstances.

There is an additional data problem. Family proceedings are generally private, and published judgments represent only part of the work of the family courts. A system trained or evaluated mainly on published cases may learn from a dataset skewed towards appellate, unusual or publication-worthy disputes. Past decisions may also contain the very human assumptions and institutional inequalities that the system is expected to overcome.

My proposed red line is firmer: no autonomous machine decision should determine a child’s living arrangements, contact, removal from parental care, placement for adoption, findings of abuse or other core welfare outcome. AI may support administration and carefully bounded analysis. It should not hold the judicial power.

8. The privacy trap for litigants in person

The access-to-justice benefit comes with a serious privacy risk.

Family litigants hold intensely sensitive material: children’s names and addresses, medical records, school reports, allegations of abuse, police disclosure, psychological assessments, photographs, messages and court orders. A distressed parent may upload an entire bundle because the chatbot invites them to “attach the document” and promises an immediate summary.

The judicial guidance says that information entered into a public AI chatbot should be treated as capable of becoming public. The SRA’s August 2026 warning likewise identifies confidential information being placed into public tools and warns of risks to confidentiality, legal professional privilege and data protection.

This cannot be dealt with by a small disclaimer that nobody reads. Systems and support services must distinguish clearly between:

• a public consumer chatbot;
• a paid tool whose terms may still permit retention or secondary use;
• a professionally procured system with appropriate contractual and technical controls; and
• a genuinely closed, approved environment designed for sensitive justice data.

Payment alone does not make a tool confidential. Removing a name does not always make a family story anonymous. A combination of school, location, dates, medical detail and procedural history may identify a child even where initials are used.

9. The risk of a two-tier AI justice system

Sir Geoffrey notes that lawyers use general or specialist systems according to their means and their confidentiality needs. That observation points to an inequality the justice system must confront directly.

A large firm may have a secured legal AI system connected to authoritative databases, controlled client workspaces, trained staff, audit logs and professional indemnity insurance. A litigant in person may have a free public chatbot, an old phone, no subscription database and no reliable way to check whether a case exists.

If the state responds by holding both users to the same practical standard while providing neither verified tools nor human support to the unrepresented person, AI may widen the gap it was expected to close.

The answer is not to lower the requirement of accuracy. The answer is to make accuracy achievable. That means publicly accessible, source-linked legal information; plain-language guidance; safe document tools; court-based support; digital-literacy training; and routes to limited human assistance before unreliable material reaches a hearing.

It also means that courts should not mistake professional-looking drafting for professional knowledge. Fluency can disguise vulnerability. A parent may submit a document that reads like counsel’s skeleton argument while understanding none of the authorities or procedural consequences within it.

Equality of arms is not achieved by giving everybody a chatbot. It requires a fair opportunity to understand, verify and challenge what the technology produces.

10. Liability: using AI—and failing to use it

The lecture’s third reality concerns legal responsibility for AI-caused harm. Sir Geoffrey refers to the UK Jurisdiction Taskforce’s July 2026 Legal Statement on liability for non-deliberate AI harms. The statement considers liability across the AI supply chain, the responsibility of professional and individual users, and the circumstances in which a professional might eventually face liability for failing to use AI.

That final question will become increasingly important. If a reliable system can identify a missed deadline, locate a decisive authority, detect a mathematical error or reveal an overlooked document, professional competence may one day require its use in defined circumstances.

But “AI is available” cannot become a shortcut to negligence. Before non-use attracts criticism, the profession needs evidence about the tool’s reliability for the task, approved procurement, training, secure data handling, clear responsibility and a realistic standard for human review.

There must also be accountability upstream. It is inadequate to say that the last human to touch an output carries every consequence while a developer, vendor or institution controls the model, interface, testing, updates, warnings and commercial pressure under which it is deployed.

Responsibility should follow control, knowledge, role and contribution to the harm. It will rarely sit in only one place.

11. A proposed Family Justice AI Accountability Framework

The legal community does need to move quickly, as Sir Geoffrey urges. But adoption should follow a published framework rather than a series of quiet experiments.

I propose ten minimum requirements.

1. Protected judicial functions. AI must not autonomously determine findings of fact, credibility, domestic-abuse risk, a child’s wishes and feelings, welfare outcomes, removal, contact, placement or adoption.
2. A materiality-based disclosure rule. Routine spelling, formatting and administration need not be disclosed. Material AI involvement in evidence synthesis, risk assessment, legal analysis, outcome recommendation or substantive reasons should be recorded and disclosed sufficiently for procedural fairness.
3. Direct judicial engagement. The judge must read and assess the evidence material to the decision. An AI summary must remain a map back to the record, never a substitute for it.
4. Secure and approved environments. Sensitive family data should be processed only within systems meeting defined contractual, technical, access, retention and data-protection requirements.
5. Traceable outputs. Every material factual proposition in an AI-assisted chronology or summary should link to the source document, page, message or recording from which it was drawn.
6. Auditability. The system, version, task, relevant inputs, output, reviewer and checks should be securely recorded where AI materially contributes to a judicial or professional decision.
7. Independent testing. Tools should be evaluated on family-specific tasks, including missing data, conflicting accounts, non-linear disclosure, disability, language, culture, domestic abuse and the differential impact on children and vulnerable adults.
8. Meaningful contestability. A party affected by material AI use should have a practical route to identify the contribution, correct input errors, challenge the output and obtain genuine human reconsideration.
9. No penalty for refusing machine adjudication. A person should not face greater delay, cost or adverse inference because they insist upon a human decision in a matter affecting fundamental or family rights.
10. Incident reporting and public learning. Significant errors, bias, confidentiality failures and near misses should be recorded, investigated and reported in a way that protects children while allowing the system to learn.

Commercial confidentiality cannot be permitted to conceal how a tool used in public justice was tested or what its known limitations are. A vendor may own software. It does not own the legitimacy of a judicial outcome.

12. What litigants in person can safely do now

Litigants in person should not be frightened away from using a tool that may genuinely help. The safer approach is to use AI for organisation and explanation, while keeping evidence, law and responsibility outside the machine.

More suitable uses

• asking for a plain-English explanation of a general legal or procedural term, then checking it against an official source;
• turning non-confidential, placeholder-based notes into a proposed heading structure;
• reordering dates already supplied into a draft chronology;
• shortening repetitive text without changing the facts;
• checking spelling, tone and readability;
• creating a list of questions to take to a lawyer, McKenzie Friend or support service; and
• building a checklist from a court order, provided every deadline and obligation is checked against the order itself.

Uses requiring particular caution

• asking AI to find case law or tell you what the law is;
• asking it who is telling the truth or what a judge will decide;
• asking it to diagnose a parent, child or professional;
• allowing it to add facts, allegations, quotations or references;
• using it to reconstruct a message, image, audio recording or missing document;
• pasting private court documents or identifying information into a public tool; and
• filing any output that you have not read, understood and checked against the original sources.

A seven-point check before anything goes to court

1. Facts: Can I identify the evidence supporting every factual statement?
2. Accuracy: Did the tool add a date, quotation, event or description that I did not provide?
3. Authorities: Does every cited case exist on an authoritative source, and does the judgment actually support the proposition?
4. Current law: Is the legislation or rule still in force and applicable in England and Wales?
5. Relevance: Does this point help answer an issue the court must decide, or does it merely increase the volume?
6. Privacy: Have I kept children’s information and private proceedings out of public AI systems?
7. Ownership: Can I explain and stand behind every sentence under my name?

If the answer to any of those questions is no, the document is not ready to be filed.

Final thoughts

Sir Geoffrey Vos is right that the legal system cannot turn back the technological tide. He is right that AI can remove drudgery, improve presentation and make legal processes more usable. He is right that a justice system which becomes irrelevant to the people it serves will ultimately damage the rule of law.

But speed is not neutrality. Fluency is not accuracy. A recommendation is not harmless merely because a human later clicks “approve”. And consent is not a sufficient constitutional safeguard where the person most affected is a child who never chose the process.

The real task is not to decide whether we are for or against AI. It is to decide which human duties must never be outsourced, what evidence is required before a system is trusted, when its involvement must be disclosed and how an ordinary person can challenge it.

Family justice should use technology to make human judgment better informed, more consistent, more timely and more accessible. It must not use a human judge to make machine judgment look legitimate.

Sources and further reading

• Sir Geoffrey Vos, John Lehane Memorial Lecture: “The implications of the development of artificial intelligence on the judiciary and the legal profession”, 18 August 2026
• Judiciary of England and Wales, Artificial Intelligence Guidance for Judicial Office Holders, 31 October 2025
• Solicitors Regulation Authority, Misuse of AI warning notice, 17 August 2026
• Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)
• D (A Child) (Recusal) [2025] EWCA Civ 1570
• Re A, B, C and D (Extension of Assessment; Use of AI: Hallucinations) [2026] EWFC 71 (B)
• UK Jurisdiction Taskforce, Liability for AI Harms under the private law of England and Wales, July 2026
• Human Rights Act 1998, Schedule 1, Article 6
• Children Act 1989, section 1
• Family Procedure Rules 2010, Part 1: Overriding Objective
• Information Commissioner’s Office, automated decision-making and children

Sources checked on 24 August 2026. The analysis and reform proposals in this article are the author’s commentary.

Important information

This article provides general legal information and commentary about England and Wales. It is not legal advice and should not be relied upon as a substitute for advice on the facts of an individual case. Laws, court rules, guidance and technology change. Always check current primary sources before relying on legal information generated or summarised by an AI tool.

JSH Law Ltd is not regulated by the Solicitors Regulation Authority. It does not conduct litigation, exercise rights of audience or undertake reserved legal activities. JSH Law provides practical, structured and procedural support to litigants in person.

How JSH Law can help

JSH Law supports litigants in person to turn large amounts of material into a clear, proportionate and source-linked case. This can include chronologies, statements, schedules, document organisation, bundle preparation, report review and hearing preparation.

Where AI has been used, JSH Law can help you separate source evidence from generated wording, identify material requiring verification, remove repetition and make sure the document remains yours—factually accurate, understandable and focused on the issues the court must decide.

Book a free 15-minute introductory call:

Connect with JSH Law

Follow JSH Law’s work on family justice, access to justice and responsible legal technology.

JSH Law links and social channels