- 24-hour Emergency Line: 0300 30 32 442
Dr AI is now firmly part of the consultation room, whether we invite it in or not. In the UK, both patients and clinicians are quietly using large language models such as ChatGPT, Claude, Gemini, Copilot and other chatbots to make sense of symptoms, treatment options and clinical dilemmas, often without a clear view of the health and medico legal consequences.
This note is aimed at UK healthcare professionals and organisations. It complements existing duties and should be read alongside local policies and UK GDPR obligations.
For years, people have turned to search engines before contacting their GP, but large language models have changed the tone and confidence of online advice. Instead of lists of websites and articles, patients now receive fluent, personalised narratives that feel uncomfortably close to professional guidance. Recent UK focused research has shown that these tools perform less reliably when interacting with real people describing messy, real world symptoms, creating risks of misplaced reassurance or inappropriate urgency that can distort help seeking behaviour.
A typical scenario is a person describing chest pains and anxiety to a chatbot at two in the morning and being reassured that this is most likely stress. If that individual delays seeking help for an evolving myocardial infarction because of this, the medico legal question becomes whether any duty of care was owed, by whom, and how that interacts with the clinician’s later duties when the patient eventually presents.
Emerging surveys suggest that a majority of British users who experiment with artificial intelligence chatbots will at some point ask about a health issue, often as a way to sense check before bothering a doctor. Patients use large language models to decide whether to contact NHS 111, to interpret NHS test results and letters, or to challenge clinical decisions they do not understand. This can support shared decision making, but the provenance, quality and limitations of the underlying models are opaque to most users.
Clinically, one visible effect is the patient who arrives with a fully formed diagnosis and management plan generated by a chatbot and expects the clinician to justify any deviation. This can lengthen consultations, encourage defensive practice and generate tension when the clinician’s advice appears more conservative than that of the chatbot, placing a premium on clear explanation, shared decision making and robust documentation.
Healthcare professionals are themselves experimenting with artificial intelligence tools for summarising notes, drafting correspondence and, in some cases, exploring diagnostic possibilities. UK regulators have not banned the use of such tools, but their core standards continue to apply. Clinicians must recognise and work within the limits of their competence, obtain informed consent, protect confidentiality and keep clear, accurate records.
The legal position is clear on one key point. Reliance on artificial intelligence does not dilute professional accountability. If a UK clinician bases a clinical decision materially on the output of a general purpose large language model, any resulting negligence claim will be judged against the usual legal tests and relevant professional standards, not against the performance of the model itself. Using a non medical grade large language model as a primary diagnostic tool for a patient is likely to fall below the standard of care, especially where red flag symptoms are present. These systems should therefore be treated as tools to support, but never replace, structured clinical reasoning and evidence based practice.
When harm occurs after artificial intelligence influenced diagnosis or self diagnosis, the legal analysis in England and Wales still turns on familiar concepts of duty, breach, causation and damage, applied against the backdrop of professional guidance and case law. For clinicians, the key risk lies in allowing artificial intelligence advice to dilute the rigour of history taking and examination, or in failing to probe a patient’s reliance on external information.
If a general practitioner learns that a patient delayed attending for a red flag symptom because an artificial intelligence tool downplayed it, they should explore this explicitly, document it and correct misunderstandings. If the clinician then fails to act appropriately on those disclosed symptoms, the presence of artificial intelligence in the background will not shield them from liability.
General purpose large language models that explicitly disavow medical use and are not marketed as diagnostic tools currently sit outside the strictest tier of UK medical device regulation, but that position may change. As outputs increasingly resemble personalised medical advice, regulators such as the MHRA are likely to scrutinise how these tools are presented and used, particularly where vulnerable patients are involved.
Consent in this context extends beyond clinical consent to digital consent about how data are used. Clinicians who paste identifiable or easily re-identifiable patient information into public interfaces for large language models risk breaching UK GDPR and common law confidentiality duties, as well as local NHS information governance policies. A simple organisational rule that no patient identifiable data are ever entered into public chatbots is a powerful and proportionate control.
The following measures can help clinicians and organisations manage risk within existing professional and legal frameworks.
Clinicians should avoid entering identifiable or easily re-identifiable patient information into public artificial intelligence tools, in line with UK GDPR, common law confidentiality and NHS information governance requirements. Local policies should make this explicit and be reflected in induction and ongoing training.
If a large language model is used to help draft clinic letters or reports, the clinician should review and edit the content as if it had been written by a trainee, ensuring it accurately reflects the encounter and contains no fabricated or distorted information. Where artificial intelligence use is material to a recommendation, there is a growing ethical and emerging legal argument for explaining that use to patients as part of shared decision making.
Normalise disclosure with questions such as “Many people look things up online or use artificial intelligence tools before seeing us. Have you done anything like that”. Clinicians should correct inaccuracies, explain uncertainty, signpost to trusted UK specific resources such as NHS guidance and record the discussion.
Clinical governance frameworks should explicitly address generative artificial intelligence, including policies on acceptable use, bans on entering identifiable data into public tools, incident reporting pathways and training on issues such as hallucinated content. Indemnity providers and medico legal advisors are likely to expect evidence of such governance when assessing organisational risk.
Use micro practices to reduce risk:
Simple habits can make a difference. Before closing a consultation, clinicians can ask “Is there anything you read online or from an artificial intelligence tool that we have not addressed today”. When safety netting, they can anticipate artificial intelligence use. For example “If you later look this up online or ask an artificial intelligence tool and it seems different from what we have discussed, please contact the practice or NHS 111 so we can clarify”.
Brief entries such as “Patient consulted online artificial intelligence tool, advised probable muscular pain. I have explained the need to exclude cardiac causes and given clear red flag advice” can be crucial if the condition deteriorates. They demonstrate that the clinician recognised the influence of Dr AI, corrected it where needed and provided appropriate safety netting.
Clinicians need not discourage patients from using large language models to generate questions or understand terminology, but should emphasise that artificial intelligence can be confidently wrong. When advice from a chatbot contradicts NHS guidance or professional advice, patients should be told to treat this as a trigger to seek clarification, not as a reason to choose whichever answer feels more reassuring.
Patients are unlikely to stop using artificial intelligence for health questions, and clinicians will continue to explore digital tools that promise efficiency and support. For now, the safest stance is to recognise Dr AI as a powerful but unreliable narrator whose role must be carefully bounded by professional judgment, regulatory standards and robust governance. Clinicians remain accountable for their own decisions, organisations must explicitly integrate generative artificial intelligence into their governance systems, and patients need clear, practical guidance on how to use and not use these tools in relation to their health.
Looking for medical protection? MDS offers membership plans tailored to a wide range of healthcare professionals, including access to expert medico legal services. Visit our membership page to see if we have the right plan for you. If you have any questions about the medical protection and support we provide as a Medical Defence Organisation, our membership team is here to help.
Not a Member Yet? Join MDS now for immediate access to expert medico-legal advice, statement & employment support.
Call us on 0300 30 32 442 
Email:[email protected].
