When AI Becomes Smarter Than Lawyers, Who Protects the Child?

Family Justice, Recoded — 30 October 2026 | An evidence-led JSH Law commentary on Bill Gates, judicial AI, children’s welfare and the future of human decision-making.

Bill Gates forecasts a world in which machines exceed human intellectual capabilities. England and Wales’ judiciary has just declared that judicial decisions must remain human. Meanwhile, independent research asks whether checking AI outputs really works. For families whose lives depend on the justice system, the difference between intelligence and accountability is everything.

1. A walk with Bill Gates — and a question for the Family Court

On 25 October 2026, The Observer published James Harding’s interview with Bill Gates, conducted during a walk around the Royal Society in London. Gates describes advanced artificial intelligence as an evolutionary turning point: systems could outperform humans in tasks ranging from coding and scientific discovery to much of today’s professional work. That remains a forecast, not a finding that a machine is already better at every legal or welfare assessment.

What particularly interests me is Gates’ suggestion that certain roles should be designated “human reserved”. His proposal appears in the context of employment and technological disruption; it is not a proposal he specifically made for English family courts. But the idea has a powerful application here.

What should artificial intelligence be permitted to assist with in family proceedings — and what must remain the responsibility of a human being?

The answer matters to judges, advocates, Cafcass practitioners, children, parents and litigants in person. As Gates argues in his own August 2026 essay on the turbulent AI era, capability is advancing quickly, but institutions have difficult choices to make about who benefits, who bears the risks and which activities society protects.

2. A remarkably timely intervention from the judiciary

On 9 October 2026, the Courts and Tribunals Judiciary issued its Position Statement on the Judicial Use of Artificial Intelligence. It expressly adopts a cautious but positive approach: AI can support judicial work, but must not replace judicial decision-making. The statement says that decisions must be taken by a judicial office holder, and that AI must not displace legal reasoning, evaluation of evidence or judicial discretion.

Its practical requirements matter as much as its principles. Each AI output used in judicial work is to be reviewed, verified and approved by the responsible judicial office holder, who remains accountable. Only approved tools may be used for judicial purposes. The statement specifies UK data residency, enterprise data protection and disabled web access among their required features.

It also describes controlled pilots in summarisation, transcription, redaction, accessible writing, administrative work, chronology-building and document analysis. The statement does not establish that a judge in a particular family case is using all these systems, nor that every pilot has proved successful. Deployment is subject to governance, oversight and further evaluation.

This is a defensible boundary. But merely assigning final responsibility to a human does not establish that an AI-assisted process will be reliable, fair or intelligible.

3. The evidence gap: is “human oversight” enough?

In July 2026, Dr Holli Sargeant of the University of Cambridge published Artificial Intelligence and Justice: An Evidence Scoping Review for the Nuffield Foundation. The review covered administrative, civil and family justice in the UK and mapped 45 AI or data-driven tools in use or development. Public evaluations were available for only a small minority.

Her 20 October 2026 commentary highlights a particular weakness: the review found no UK evaluation specifically testing whether human checking of AI outputs works in justice settings. That is a statement about the evidence found by the review, not proof that human review never works.

Consider a 500-page family-court bundle condensed by a machine into a 12-page chronology. If the summary omits a crucial safeguarding disclosure or changes an allegation into a finding, how will a busy reviewer detect the mistake? Checking every factual proposition against the source may erode some time saving. Checking only a sample may leave serious omissions undetected. A confident writing style can conceal uncertainty.

The consequence is a risk of automation bias: people may place unwarranted trust in computer-generated material because it appears complete, consistent and professional. The proper test is therefore not simply whether a person clicked “approve”. It is whether that person’s checking demonstrably identifies material errors in realistic cases.

4. Government progress: AI is no longer a distant prospect

The Ministry of Justice’s AI Action Plan for Justice: One Year On, published 24 September 2026, reports large-scale use of AI in parts of the justice system. For example, its Justice Transcribe tool had processed more than 1.5 million probation meetings between October 2025 and September 2026, and the department reports broader adoption of enterprise productivity tools. It describes non-generative decision-support applications in probation, prisons and courts and continued exploration of AI-enabled court administration.

These examples should not be confused with autonomous adjudication in family proceedings. They do, however, show why discussions about governance cannot be postponed. The plan also adds a fourth priority addressing emerging AI risks, including synthetic and manipulated material.

For family justice, that makes the question immediate: how do we obtain the gains in speed and accessibility without weakening the reliability of evidence or the procedural fairness owed to the people affected?

5. The welfare principle is not a predictive algorithm

Under section 1 of the Children Act 1989, the child’s welfare is the court’s paramount consideration when it determines the relevant questions concerning upbringing. Where the statutory welfare checklist applies, the court must consider matters including ascertainable wishes and feelings in light of age and understanding, needs, the likely effect of changes, harm, parenting capability and the available options.

An AI tool might organise the evidence relevant to those statutory factors. It may make a chronology searchable or help the court locate passages that have not previously been connected. It does not follow that its numerical predictions, classifications or summaries should decide whether a child lives with a parent, spends time with a relative or needs protection.

A child’s wish may be clear but its meaning contested. A parent may present consistent messages while other evidence tells a different story. What looks like a recurring pattern to a language model may be explained by context, or may be a genuine warning sign. Those distinctions require fair investigation and an accountable judicial assessment. The child’s welfare is not the average outcome for a superficially similar cohort.

The strongest claim for AI is therefore augmented human attention: helping the decision-maker find what must be examined, not deciding what must be believed.

6. Domestic abuse, coercive control and the difference between patterns and proof

In proceedings involving alleged domestic abuse, evidence can be scattered across text messages, financial records, safeguarding documents, medical records, police material and years of correspondence. Carefully constrained AI may be useful in indexing dates, identifying references and suggesting lines of inquiry.

But Practice Direction 12J requires a case-specific approach to domestic abuse and harm in relevant child-arrangements proceedings. The proper distinction between an allegation, an admission, a factual finding and a professional opinion is indispensable. An AI output must not silently promote one category into another.

Coercive control may be cumulative and contextual: an apparently ordinary message may carry significance in a known history, while an alarming excerpt can be misread when separated from its context. Technology might help researchers and lawyers see patterns, but it may also magnify selection bias, translate ambiguity into certainty or miss relevant contrary evidence.

Practical rule: every material AI-assisted allegation summary should be linked to the original source, with its date, author, status, context and any competing account clearly recorded. The responsible person should decide what can properly be put before the court.

7. A real Family Court warning: Re A, B, C, D [2026] EWFC 71 (B)

In Re A, B, C, D (Extension of Assessment; Use of AI: Hallucinations) [2026] EWFC 71 (B), Recorder Howard considered inaccurate legal references in submissions made by a friend of the mother and prospective kinship carer who appeared in person. It would be inaccurate to describe her simply as one of the parents. The court examined the failure to verify legal authorities and the risk of misleading the court.

The judgment quotes the Court of Appeal’s observations in Re D (A Child) (Recusal) [2025] EWCA Civ 1570 at paragraph 83. AI, used properly and responsibly, can assist litigants and lawyers; it is neither an authoritative nor infallible source of law. Parties remain responsible for checking whether cited cases exist and actually support the legal proposition advanced.

Recorder Howard accepted that the person had not intended to mislead the court, but remained concerned about the seriousness of the errors. The judgment’s publication conditions also underline the need to preserve the anonymity of the children and family members. The lesson is not “never use AI”; it is that plausible text must not replace verified authority.

The broader professional warning is also apparent in Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), where the Divisional Court addressed fabricated or unreliable legal citations and lawyers’ duties to the court.

8. Litigants in person: an opportunity that must not be squandered

For someone without legal representation, affordable AI could be genuinely useful. It can help explain directions, turn correspondence into a dated list, identify questions to ask, create a first draft of a chronology and make technical language more accessible. It can help a person recognise when a source is missing or a procedural step needs checking.

But a litigant who needs an AI assistant may lack the tools or legal expertise to identify its mistakes. Courts should neither assume that every AI-assisted document is defective nor treat a polished AI output as evidence of legal knowledge. Accessibility, clear guidance and proportionate opportunities to correct error are critical.

At JSH Law, the appropriate principle is straightforward: AI may help organise and draft, but source documents and cited authorities must be independently checked; disputed allegations must stay labelled as disputed; and final court documents require human review.

9. Confidentiality: a particular danger in private children proceedings

Family-court papers routinely contain intimate information about children, parents, health, safeguarding and allegations. The October 2025 judicial AI guidance warns judicial office holders against placing private information into public AI tools. Wider obligations also arise from data protection and confidentiality law, and from the rules governing communication of information from private children proceedings.

The Family Procedure Rules, Part 12, Chapter 7 and Practice Direction 12G regulate permissible communication of information in proceedings involving children. A permission to share information with a specified person for support or advice must not be casually assumed to authorise uploading the same material to an external AI service. Whether any processing or disclosure is lawful depends on the circumstances, the service, the recipient, the contractual arrangements and any relevant court direction.

Do not upload a confidential family-court bundle, identifiable child’s records or sensitive safeguarding material into a general-purpose AI platform without first establishing an appropriate and lawful basis, security controls and any necessary court permission. Where legitimate use is contemplated, documented due diligence, data minimisation and secure approved systems matter.

10. AI-generated evidence and synthetic reality

Even where AI is never used by a judge, it may still influence what the court sees. Generated screenshots, manipulated voices, fabricated recordings, edited images or synthetic correspondence can be presented as apparent evidence. At the other extreme, genuine evidence may be wrongly dismissed as machine-generated. Both risks can damage procedural fairness.

The practical response is proportionate provenance checking: retain original files and metadata where available, preserve disclosure chains, identify transformations, avoid repeatedly re-exporting important material and obtain specialist analysis where warranted. Unsupported claims that a disputed item “looks AI-generated” should not become findings.

11. What a genuine human-reserved approach could mean

Gates’ employment proposal prompts a different and more fundamental question about public authority. The relevant distinction is not merely between a fast machine and a slower professional. It is between technical performance and legitimate responsibility.

Judicial decisions require reasons that can be understood and challenged, procedures that permit participation, and a person accountable for the exercise of public power. Even if future systems outperform individual humans at many analytical tasks, that would not automatically settle the democratic, legal or ethical question of whether a machine should determine a child’s future.

Human-reserved decision-making should not mean refusing technological support. It should mean preserving the substantive decisions, hearings and accountability mechanisms that make family justice justice, while permitting evidence-led assistance where its benefits are demonstrated.

12. JSH Law’s proposed Family Justice AI Accountability Framework

The following is a proposed policy framework, not a claim that seven identical legal obligations already exist in statute.

1. Reserved judgment: legal findings, welfare decisions and judicial discretion remain genuinely human and accountable.
2. Source traceability: every material AI-assisted statement can be checked against original evidence and authorities.
3. Safeguarding testing: tools are evaluated against realistic scenarios involving domestic abuse, vulnerability, children’s voices and disputed facts.
4. Lawful confidentiality: sensitive case material is processed only through appropriately authorised and secure arrangements.
5. Meaningful transparency: parties and courts can identify material AI influence where fairness requires disclosure.
6. Effective challenge: individuals have a practical route to correct material mistakes and contest conclusions.
7. Independent evaluation: policymakers test not only time savings but accuracy, fairness, public comprehension, unequal impacts and lived experience.

There should also be independent scrutiny of the cost of checking AI outputs. A tool which saves 30 minutes in drafting but requires 45 minutes to verify may still offer accessibility benefits, but it should not be marketed as a proven productivity gain. Equally, a tool which produces a faster decision but conceals a decisive omission would represent a failure of justice, whatever its efficiency score.

13. Seven practical checks before relying on AI in a family case

1. Check every quoted authority against an authoritative judgment or legislative source.
2. Keep originals; do not replace evidence with AI-created paraphrases.
3. Clearly separate what someone alleges from what has been admitted or found.
4. Confirm dates, identities, relationships and chronology manually.
5. Challenge summaries by searching for omitted or contrary material.
6. Protect children’s information and confirm disclosure is lawful before using any external service.
7. Ensure a person understands, approves and can explain everything submitted to court in their name.

For related guidance see JSH Law’s articles on who checks AI in a family-court case, AI transparency in public decisions, and child-focused courts and procedural fairness.

Conclusion: the real test is not whether AI is smarter

Bill Gates’ forecast may prove optimistic, pessimistic or incomplete on the pace of technological change. But his question about which human activities society chooses to preserve is already urgent. The judiciary has made an important commitment to human judgment. The Ministry of Justice has demonstrated that AI deployment is expanding. Independent researchers have shown how much remains untested.

Family justice should welcome technology that makes evidence comprehensible, reduces avoidable delay and gives unrepresented parents a fairer opportunity to participate. It should be uncompromising about confidentiality, verifiable evidence, welfare analysis and accountability.

The defining question is not whether AI becomes smarter than lawyers. It is whether we are wise enough to decide what it must never be permitted to replace.

---

Primary sources and further reading

• James Harding, The Observer, 25 October 2026
• Bill Gates, Gates Notes, August 2026
• Judicial AI Position Statement, 9 October 2026
• Judicial AI Guidance, October 2025
• Holli Sargeant, Nuffield Foundation review, 14 July 2026
• Holli Sargeant, Nuffield commentary, 20 October 2026
• Ministry of Justice, AI Action Plan for Justice: One Year On, September 2026
• Re A, B, C, D [2026] EWFC 71 (B), approved judgment
• Ayinde and Al-Haroun [2025] EWHC 1383 (Admin)
• Children Act 1989, section 1
• Family Procedure Rules, PD12J
• Family Procedure Rules, Part 12
• Family Procedure Rules, PD12G

About this series: Family Justice, Recoded examines law, technology and the future of access to family justice in England and Wales. JSH Law Ltd provides non-reserved family-court support and is not an SRA-regulated law firm. This is general commentary, not advice on an individual case. The proposals identified as JSH Law’s are policy recommendations rather than statements of existing law.