For medico-legal work, I would not recommend an unrestricted public chatbot or a system that claims to write the entire report automatically. I would recommend a secure, organisation-controlled drafting assistant designed to organise evidence and prepare text under the direct supervision of the named expert.
The distinction matters. One system assists professional judgement. The other encourages the expert to outsource it.
The recommended system
The safest model is a private AI workspace in which each case is isolated, and the system is restricted to the documents supplied for that instruction.
It should be able to:
extract dated events from medical records,
arrange those events into a reviewable chronology,
summarise the claimant’s reported symptoms and treatment,
identify apparent inconsistencies or gaps for the expert to examine,
populate approved report headings; and draft descriptive passages using only verified information.
It should not diagnose the claimant, decide causation, select a prognosis, assess credibility or produce a final opinion without the expert’s active input.
The system must distinguish clearly between the records, the claimant’s account and the expert’s own findings. A hospital entry should not be converted into an established fact merely because it appears in the notes. Nor should an allegation in the letter of instruction quietly become part of the clinical history.
The most useful AI system is therefore not the one that writes the most. It is the one that allows the expert to see where every statement came from.
Source-linked drafting is essential.
Every generated paragraph should be linked to its source document and, preferably, the relevant page or dated entry.
If the system states that the claimant attended their GP three times with neck pain, the expert should be able to open those three entries immediately. If it describes a previous psychiatric history, the underlying records should be visible rather than buried within a large bundle.
This allows errors to be found before they enter the signed report.
Generative AI can produce convincing but incorrect material. Current judicial guidance identifies fabricated information, inaccurate summaries and confident assertions without a proper evidential basis among the risks of these systems. It also stresses that the individual using AI remains personally responsible for material produced in their name. Although that guidance is directed principally at judicial office holders, the warning applies with force to expert evidence.
A system that cannot show its sources should not be used to prepare an expert report.
Confidentiality must be designed into the system
Medico-legal bundles contain health information, identification documents, employment records and sometimes material concerning criminal allegations or family circumstances. Health information is special category personal data and requires additional protection under data protection law.
Uploading that material into a public AI tool without an approved contractual and security framework is difficult to defend.
The recommended system should operate under a written data-processing agreement that defines the respective responsibilities of the organisation and the technology provider. The provider should state where information is processed, how long it is retained, who can access it and whether any input is used to train or improve wider models. The ICO expects contracts with AI providers to identify controller and processor responsibilities and describe the processing being undertaken.
Case data should not be used to train a general model. Access should be role-based, transmissions and stored files should be encrypted, and each case should have an enforceable deletion and retention policy.
Where full identifiers are unnecessary, documents should be pseudonymised before processing. The ICO’s guidance requires organisations to consider data minimisation, security and data protection by design when deploying AI systems. A data protection impact assessment may be required where the proposed processing is likely to present a risk to individuals.
These are procurement questions, not optional settings to be considered after the software has been purchased.
The system should preserve an audit trail.
A defensible platform should record what documents were uploaded, what instructions were given to the system, what text it produced and what the expert changed.
That does not mean preserving every experimental sentence indefinitely. It means retaining enough information to explain how the report was prepared and to investigate an error if one is later discovered.
Version control is particularly important. The expert should be able to distinguish the original AI draft from the text ultimately approved and signed. Any amendment following further records, written questions or an experts’ discussion should be identifiable.
The current Civil Procedure Rules remain centred on the expert rather than the software. The expert’s duty is to help the court, and that duty overrides any obligation to the party instructing or paying them. The report must be the independent product of the expert and contain the expert’s own opinion.
An AI system cannot assume that duty. It cannot be cross-examined and it cannot sign the statement of truth.
What the AI should never be allowed to decide.
The system may draw attention to differences between the claimant’s account and the records. It should not label the claimant unreliable or dishonest.
It may display recognised diagnostic criteria. It should not decide that those criteria have been satisfied.
It may identify earlier symptoms or accidents. It should not determine whether the index event caused, aggravated or merely coincided with the current condition.
Those decisions require clinical knowledge, examination findings and professional reasoning. The GMC requires experts to give objective and unbiased evidence, remain within their competence, identify the facts and assumptions supporting their opinion and address information that weakens it.
A system that generates a completed causation and prognosis section from a questionnaire and a bundle may save time. It also creates a serious risk that the expert will review the wording rather than form the opinion independently.
There is a difference between editing one’s own reasoning and approving reasoning produced by a machine. The finished prose may look similar, but the intellectual process is not.
Human review must be more than proofreading.
The named expert should review the original relevant records, conduct or supervise the clinical assessment and form the medical opinion before approving the final report.
Human review does not mean correcting spelling, changing a few phrases and adding a signature. It requires checking every factual statement, considering omitted material and rewriting any section that does not represent the expert’s own analysis.
The expert should also be able to produce the report without the system. Otherwise, the software has become the true author and the clinician merely its approver.
The Civil Justice Council is presently examining whether specific rules are needed for the use of AI in court documents, including expert reports. The position may therefore become more prescriptive.
That possibility should not delay sensible safeguards now.
The system I would recommend is a secure, closed and source-linked drafting assistant with strict access controls, an audit trail and mandatory expert approval. It should undertake the clerical work that machines perform well while leaving diagnosis, causation, prognosis and professional judgement where they belong.
The expert may use AI to prepare the report. The AI must never become the expert.
