AI and expert evidence in care proceedings

The judgment in Cornwall Council v M, F, A, B and C [2026] EWFC 306 is worth reading for anyone involved in public law proceedings. It does not say that experts must not use artificial intelligence. It addresses the more difficult question of where legitimate assistance ends and the expert’s own professional judgment must begin. The issue mattered: HHJ Flexman ultimately concluded that no significant weight could be placed on the independent social worker’s evidence. [40]

What had AI actually been used for?

The position emerged in stages. There had been no disclosure of AI use in the first report. When questions were subsequently raised, the independent social worker explained that Microsoft Copilot had been used to organise information and assist with drafting, including spelling and grammar. She maintained that the contents were her professional opinion and said that she used AI in that way because “it produces better English than I use myself”. [33]

The reports, however, contained a warning that the tool had gone further. Text which was plainly not intended to appear in a finished assessment had been left in place. It included Copilot commentary explaining that material was now “about half the length” and was ‘more likely to withstand cross-examination’, together with explanatory drafting commentary about why passages had been written in a particular way. [31]. Those remnants were clearly not evidence about the family or intended to have been left in the report. At paragraph [35] the Judge records the analysis of the evidence, the sub-headings, layout, construction and development of the report as AI generated. The Judge went on to say, having heard the expert give oral evidence, that the tone of the report and use of language was very different from her own and that the use of AI in this way “has placed a significant gap between the voice and opinion of the expert and the report which she has filed.” [35].

The assessment also contained factual errors. It stated that the father had accepted the court’s findings when he had not, and described a child as autistic although there had been no diagnosis. The judge could not determine whether those errors had originated in the AI output or in the underlying notes. Either way, their presence illustrated the risk of submitting generated text without rigorous checking against the source material. The report even purported to adopt a Resolutions model of risk assessment without the expert having been instructed to use it or her CV indicating that she was qualified to use it. [36]

The line drawn by the court

There is a material difference between saying that AI helped to organise or improve the wording of an assessment and accepting that it converted the assessor’s notes into the report which constituted her expert evidence. The Court summarised at [37]:

37. I accept that there is a sliding scale of use from what is acceptable to what is not and it is not always straightforward to differentiate between what might arguably be assistance with presentation and what in fact amounts to AI authorship. There are areas between the categories of risk identified with the currently available guidance. I am concerned that the use of AI to generate an entire report from notes strays into the latter. I accept that the primary material was [the ISW)’s. AI was not asked to write a report on its own, it used her source material. But this is a higher risk approach which in my view impacts on the weight that can be attached to her written document.

The useful principle is straightforward: AI may assist an expert, but it cannot become the expert. An expert remains personally responsible for every statement in the report. The professional evaluation and opinion for which the witness takes responsibility must genuinely be theirs, supported by their own reading of the evidence and their own application of expertise.

Good practice and disclosure

The judgment drew attention to the Academy of Experts’ Guidance for Expert Witnesses on the use of Artificial Intelligence, published in January 2026. That guidance adopts a risk-based approach, stresses that experts remain ultimately responsible for their work product and recommends early discussion with instructing lawyers where a proposed use is high risk. [25]. Where AI has been used in preparing an expert report, that should be stated in the report. The statement of truth may also appropriately explain whether, and how, AI has been used. [28]

The requirement of paragraph 9.1(c) of FPR Practice Direction 25B to identify “who carried out any test, examination or interview which the expert has used for the report and whether or not the test examination and interview has been carried out under the expert’s supervision” was not drafted with AI in mind but the use of AI could, the Judge’s view, be akin to third party involvement, depending upon the way that it is used [21].

Solicitors already have relevant professional guidance of their own. The Law Society’s practice note, Using artificial intelligence in legal practice, published on 21 October 2026, emphasises human oversight, informed professional judgment and a proportionate, risk-based approach. Its central message is that professional responsibility cannot be transferred to an AI system and that AI output must be checked rather than validated by asking another AI tool.

Lessons for local authority lawyers and social workers

This is not only an issue for independently instructed experts. A social worker giving evidence in care proceedings provides professional evidence which incorporates expertise, analysis and judgment. Even where Part 25 is not formally engaged, local authority lawyers advising on statements should consider the same safeguards.

In practical terms:

• AI must not replace the author’s reading of the source material. A social worker must be familiar with the records and evidence relied upon rather than relying on an AI-generated summary of them.
• Every factual assertion must be checked against the primary material. Names, dates, diagnoses, findings, admissions and quotations are particularly vulnerable to apparently plausible error.
• The analysis and opinions must be those of the author. AI may assist with expression, but any professional opinion should reflect the author’s own conclusion, or a conclusion which the author has first formulated and instructed the programme to help express, rather than an opinion generated by the programme.
• The author must read the whole finished statement and be able to explain and defend every material proposition in it.
• The use of AI should be transparent where it has contributed materially to the preparation of the evidence, with consideration given to recording the tool used and the nature and extent of its use.
• Only an approved and suitably secure system should be used. Confidentiality, privilege, data protection, information security and the organisation’s AI policy remain relevant even where the resulting text is accurate.

When the local authority commissions an ISW

A further practical risk arises where an independent social worker is engaged by the local authority because of capacity pressures, but without a formal letter of instruction. Operationally, that may feel different from the joint instruction of a Part 25 expert. The evidence may nevertheless be placed before the Family Court and relied upon as independent professional assessment.

It may be a safer course in such circumstances to use a written letter of instruction or commissioning document which addresses AI expressly. It should identify:

• whether AI may be used and, if so, the permitted purposes and approved systems;
• that AI must not substitute for reading the source documents, interviewing those assessed or applying professional expertise;
• that the assessor must verify every factual proposition against the underlying evidence;
• that all analysis, evaluative judgments and opinions must be the assessor’s own, even if AI has assisted with wording;
• the requirement for meaningful human oversight and a final line-by-line review;
• confidentiality, data protection, privilege, retention and information-security requirements;
• the need to keep an appropriate record of the use made of AI; and
• the form of disclosure to be made in the assessment, including any explanation accompanying the statement of truth.

The letter should not simply prohibit ‘AI-generated reports’ without explaining what that means. It should set a workable boundary: technology may assist the assessor, but it must not undertake the substantive assessment or supply professional conclusions which are then presented as those of the assessor.

The real lesson

Cornwall Council v M is not an argument for pretending that AI does not exist. It is an argument for transparency, professional ownership and human oversight. An expert can use a tool to help communicate their judgment. They cannot safely allow the tool to supply the judgment which the court believes it is receiving from them.

And, in the interests of practising what I have just preached, I should disclose one final matter: AI was used in the preparation of this article. I have, however, read the source material, checked the finished text and, so far as I know, not asked Copilot to make it withstand cross-examination. And hopefully those who know me will recognise the voice.

Graeme Bentley is a Principal Solicitor in Invicta Law’s Child Protection team, with over 40 years of experience in family law.

Sources and guidance

Cornwall Council v M, F, A, B and C [2026] EWFC 306, particularly [28]-[41].

FPR Practice Direction 25B, paragraph 9.1(c).

The Academy of Experts, Guidance for Expert Witnesses on the use of Artificial Intelligence (January 2026).

The Law Society, Using artificial intelligence in legal practice (21 October 2026).