The question is often framed as though there are only two answers: AI is too dangerous for confidential records, or modern systems are secure enough to make the concern outdated. Neither is satisfactory. Medical records can be processed through an AI system safely, but only where the organisation knows what the system does with the information, who can access it, where it is processed and how it will be removed. Convenience is not evidence of compliance.
Medical information requires greater care.
Health information is special category personal data under UK data protection law. Its use requires a lawful basis for processing personal data and an appropriate condition for processing special category data. The fact that records were provided for a medico-legal assessment does not automatically permit every later use. The ICO also advises that processing likely to create a high risk may require a data protection impact assessment.
The expert must identify the purpose. Is the system summarising records, extracting dates, suggesting headings or helping draft an opinion? Is the supplier processing the material only on the user’s instructions, or may it retain prompts and documents for testing or model development? Those are different activities.
GMC guidance states that patients are entitled to expect their information to be treated confidentially. Doctors responsible for records must ensure that information is stored, transferred, protected and disposed of in accordance with data protection law. Passing records to a technology provider remains the responsibility of the professional or organisation making the disclosure.
A chatbot is not a secure filing cabinet.
A common mistake is to assume that an AI service is suitable because it has a password, an established brand or a statement that communications are encrypted. Those matters are relevant, but insufficient.
Before records are uploaded, the organisation should know the contracting entity, whether the supplier acts as a processor, which subcontractors are involved, whether prompts are used to train models, how long information is retained and whether deletion can be verified. It should also understand what happens to backups, logs and support access. The ICO expects organisations buying third-party AI systems to consider security and data minimisation during procurement and to map where personal information is processed.
Whether the platform is Medqon or another specialist report-writing or case-management system, the product name proves nothing by itself. The relevant evidence lies in the contract, technical controls, retention arrangements, access permissions and documented information flows.
A system designed for professional use may offer better controls than a public-facing chatbot. It is not made safe merely by being described as “medical,” “legal” or “compliant.”
Use only what the task requires.
Data minimisation is easily overlooked when an entire medical bundle can be uploaded with one click. ICO guidance requires personal data to be adequate, relevant and limited to what is necessary for the stated purpose. AI does not create an exception.
If the task is to extract dates from physiotherapy records, the system may not need the claimant’s full general-practice history, photographs, financial documents and unrelated family information. Names, addresses, NHS numbers and third-party details should be removed where they are unnecessary.
Pseudonymisation can reduce risk, but it is not the same as anonymisation. A reference number attached to a detailed medical history may still allow the claimant to be identified. Pseudonymised records should therefore continue to be treated as personal data. The safest upload is often the smallest useful one.
Where is the information going?
Cloud systems may involve international access even when the user is sitting in a UK office. The ICO explains that making information accessible to a separate organisation outside the UK can amount to a restricted transfer. The legal location of the provider and its subcontractors may matter as much as the location of a server.
Organisations should map the flow of information rather than rely on phrases such as “UK hosting” or “European data centre.” They need to know which legal entities can access the records, including technical-support teams and sub processors.
Where a restricted transfer is involved, an appropriate transfer mechanism and any required risk assessment must be in place. The ICO makes clear that transfer rules can apply even to small or infrequent transfers.
A DPIA should come before routine use.
Processing detailed medical records through AI may create a risk to individuals, particularly where it occurs at scale, introduces new technology or influences decisions about a claimant. A data protection impact assessment allows the organisation to identify the purpose, necessity, risks and safeguards before use becomes routine.
It should address the actual workflow: who uploads records, what information is selected, what the AI produces, who reviews it, where outputs are stored and how errors or breaches are handled.
Supplier assurances should be tested. Certificates and policy documents do not replace questions about role-based access, multi-factor authentication, audit logs, deletion, incident reporting and staff training.
AI output must be checked.
Confidentiality is only one part of the risk. An AI-generated chronology may omit a consultation, confuse two conditions or present an inference as a recorded fact. ICO guidance distinguishes data-protection accuracy from the statistical accuracy of an AI system. Organisations should account for the possibility that an AI-generated inference may be incorrect and for the effect of relying upon it.
The expert must return to the source records. AI may assist with organisation, but it cannot decide which entry is reliable, whether an apparent inconsistency is material or how a previous condition affects causation. Those are professional judgements.
The final report remains the expert’s work. Saying that the software produced an error is no answer when the expert approved it. Meaningful human review requires more than briefly reading the generated text before accepting it.
The sensible boundary.
Confidential records should not be uploaded to an AI system merely because the technology is available or manual review is slow. They may be uploaded where the purpose is clear, the processing is lawful, the supplier has been assessed, the minimum necessary information is used, and meaningful human checking remains in place.
There should also be a defensible answer to a simple question from the claimant: what happened to my records after they were uploaded? If the organisation cannot explain whether the information was retained, used for training, accessed abroad or permanently deleted, it did not understand the system well enough to use it.
AI can assist medico-legal work without weakening confidentiality. That requires controlled systems, documented decisions and professionals who remember that outsourcing a task does not outsource responsibility.
