The UK needs new laws for AI in healthcare to keep pace with systems that can learn, change and potentially drift in performance after approval, according to recommendations published on 10 September 2026. The Medicines and Healthcare Products Regulatory Agency (MHRA) is considering how its medical-device framework should evolve as artificial intelligence becomes a more routine part of NHS care.
The central issue is not whether AI will enter clinical practice. It already has: AI-enabled tools can support administrative work, clinical documentation and the interpretation of medical information. The harder question is how to preserve patient safety, professional accountability and public confidence when the software itself may change over time.
Why existing medical-device rules may not be enough
The UK’s established approach to medical-device regulation was developed largely for physical products and more conventional diagnostic tools. That structure can be suitable for relatively fixed software, such as a system trained to identify a defined feature in a scan.
More advanced AI presents a different regulatory problem. As MHRA chief Lawrence Tallon told the BBC, some products may adapt as new data are introduced. A model that changes after authorisation cannot necessarily be treated like a static device whose performance was established once before market entry.
The independent commission behind the report drew on input from more than 12,000 people, including patients and clinicians. Its work produced 44 recommendations intended to modernise the regulatory approach.
The proposals do not suggest that every AI application should be treated alike. A tool that transcribes a consultation raises different risks from one that helps identify a possible disease pattern. Regulation needs to reflect that difference.
The proposed safeguards
Several recommendations address practical weaknesses that can arise when AI moves from development into everyday care.
| Proposed measure | What it would address |
|---|---|
| Ongoing performance monitoring | AI systems could be reviewed after approval, rather than assumed to remain safe and effective indefinitely. |
| Withdrawal of approval where needed | Products that malfunction or lose effectiveness could be removed from authorised use. |
| Patient information rights | Patients would be told when AI is involved in their care and could access understandable information about the relevant product. |
| Penalties for developers | Developers could face sanctions when products fail to meet required standards. |
| Supervised “L plate” trials | New models could be used by healthcare professionals in closely controlled settings before wider deployment. |
These measures would shift attention from a one-time approval decision toward continuous oversight. That is particularly relevant for systems whose real-world performance can change as their data or operation changes.
Transparency is a clinical issue, not merely a communications issue
Patients may reasonably distinguish between an AI tool used to organise paperwork and one that contributes to a decision about their diagnosis or treatment. The report’s emphasis on disclosure recognises that people need to know when a technology has a role in their care.
AI note-taking systems, sometimes called scribes, are reportedly used by 40% of UK-based GPs to record consultations and generate reports. They can reduce administrative burden, but the source material also describes research from the University of Edinburgh indicating that some patients may withhold sensitive information if they know a conversation is being processed by AI.
That finding means healthcare providers should consider how patients who are uncomfortable with AI-assisted note-taking can discuss their concerns and whether an alternative is available.
Professor Henrietta Hughes, a GP involved in the commission, said that some patients welcomed AI use while others opted out. She also stressed that clinicians remain responsible for checking and correcting errors in AI-generated notes.
Human accountability must remain clear
In many current uses, AI is intended to assist rather than replace clinical judgment. A transcription tool may draft a consultation note; a clinician should review it. An image-analysis system may flag a possible abnormality; a qualified professional should determine what the finding means for the patient.
That division of responsibility must be explicit in regulation and healthcare practice. Otherwise, staff may place undue trust in a polished but erroneous output, while developers may argue that the final decision belonged to the clinician.
A credible framework should establish:
- Who reviews AI output before it enters the clinical record or influences care.
- When a clinician may override the system and how that decision is documented.
- How errors are reported and investigated, including errors that do not immediately cause harm.
- What happens when performance changes after deployment.
- How patients can raise concerns about AI-supported processes used in their care.
The aim is not to make healthcare professionals personally responsible for every defect in a product they did not build. It is to ensure that accountability does not disappear into the gap between developer, healthcare organisation and clinician.
Why caution should not become paralysis
The report comes amid substantial optimism about AI’s possible contribution to healthcare and medical research. Professor Alastair Denniston, an ophthalmologist involved in the work, described the technology as a potentially transformational development. Such expectations are understandable, particularly where AI can help clinicians manage growing volumes of information.
But predictions about future breakthroughs are not evidence of present clinical benefit. Claims that AI will rapidly solve major diseases should be treated as forecasts, not as a basis for deploying a product without robust testing and monitoring.
The more immediate case for AI is often practical: reducing repetitive documentation, helping prioritise information for review and supporting clinicians in specific, well-defined tasks. These uses may still create risks, but those risks can be identified and managed more clearly than sweeping promises of autonomous medical decision-making.
What patients should ask when AI is used in their care
Patients do not need technical expertise to raise sensible questions. Where an AI tool is involved, useful questions include:
-
What is the system being used for?
Is it taking notes, analysing an image, prioritising cases or contributing to a treatment decision? -
Will a qualified clinician review its output?
Patients should understand whether a professional retains meaningful oversight. -
Is an alternative available?
This may be especially relevant for AI transcription during sensitive conversations. -
What information is being processed?
Clear explanations help patients decide whether they are comfortable with the arrangement. -
What happens if the system is wrong?
The provider should be able to explain the route for correction, reporting and follow-up.
What healthcare organisations and developers should prepare for
If the MHRA’s recommendations lead to regulatory change, healthcare organisations will need more than a procurement checklist. They will need governance processes that assess how a tool behaves in their own environment, how staff use it and whether it creates inequitable outcomes.
Developers, meanwhile, may need to demonstrate not only pre-market performance but also a capacity to monitor their systems after deployment, respond to reported failures and communicate material changes clearly.
The test will be whether regulation earns trust
The MHRA’s recommendations acknowledge a basic reality: AI in healthcare cannot be governed effectively through a framework built chiefly around stable, physical devices. Systems that change need monitoring; patients need meaningful information; clinicians need clear duties; and developers need enforceable standards.
The UK has not yet found a final answer to these problems, and the MHRA itself recognises that no country has fully resolved them. That uncertainty is not a reason to delay governance. It is a reason to build rules that can respond to technology without asking patients to bear the risk of regulatory delay.
Sources
- UK needs new laws for AI in healthcare, says watchdog — Published 2026-09-10
Meta: MHRA-backed recommendations call for UK AI healthcare rules covering monitoring, patient transparency, developer penalties and supervised trials.