A 44-point plan for AI in healthcare
The UK needs a new regulatory framework for artificial intelligence used in the NHS and wider healthcare system, according to the country’s medicines and medical devices watchdog. The Medicines and Healthcare Products Regulatory Agency (MHRA) has published 44 recommendations designed to keep pace with AI tools that can learn, adapt and change after receiving approval. Its warning comes as the technology moves from trials and back-office software towards everyday patient care. An independent commission developed the proposals with input from more than 12,000 people, including patients and clinicians. The central challenge is clear: embrace some of healthcare’s most promising new gadgets without allowing safety, transparency or public trust to fall behind.
MHRA chief Lawrence Tallon told the BBC that AI would soon become a familiar part of NHS services. “What I would expect is that patients will… increasingly see AI as part of the way that normal NHS healthcare is delivered,” he said. “That should happen in a way that they can maintain their trust and their confidence in what’s happening.”
Rules built for software that keeps evolving
Britain’s existing medical devices regime was largely designed for physical products with relatively predictable behaviour. That approach can handle a hip replacement, a stethoscope or a plaster rather well. It is far less suited to an algorithm that receives fresh data, adjusts its internal calculations and potentially performs differently months after deployment.
“The medical devices regulatory framework predominantly dates from a period where we were thinking about things like hip replacements and knee replacements, or smaller things like stethoscopes and plasters,” Tallon said. He added that current guidance may be adequate for straightforward AI trained to identify known symptoms in scans, but more sophisticated systems create a different regulatory problem. “Unlike most of the medical products we’re used to regulating, these products continue to change after the point of authorization,” he said. “As new data gets fed in, they learn, they adapt, they drift.”
The MHRA’s proposed safeguards would put continuous monitoring at the heart of approval. Regulators could withdraw a product from the market if it malfunctioned or became less effective over time, rather than treating initial authorisation as the end of the process. Patients would also have the right to know when AI is involved in their care and receive accessible information about the system being used.
Developers failing to meet required standards could face penalties, creating a stronger incentive to test thoroughly and report problems quickly. Another recommendation is an AI “L plate” system, allowing new models to be trialled by healthcare professionals under close supervision before receiving broader approval. It is, in effect, a controlled test drive for technology that may eventually operate across busy hospitals and GP surgeries.
Tallon cautioned that Britain is not trying to solve a purely domestic puzzle. “I don’t think at this moment in time we can point to a single country, a single regulatory framework, and say that they have absolutely cracked it,” he said. That leaves the UK with an opportunity to help establish a credible model for regulating medical AI that learns in the real world.
Patients must be able to opt out
AI is already entering the consultation room through digital scribes. These note-taking systems, powered by large language models, can listen to conversations, produce clinical records and generate reports. They are reportedly used by 40% of UK-based GPs, making them one of the most visible healthcare applications of generative AI.
Convenience, however, does not guarantee comfort. A recent University of Edinburgh study found that patients could be less willing to disclose personal information, including a history of substance abuse, when they knew AI was processing the conversation. Privacy is only part of the issue; some patients simply do not want a machine participating in deeply personal medical discussions.
Professor Henrietta Hughes, a GP who served on the report commission, said many of her patients accepted AI during consultations, although others chose to opt out. “Some say, ‘I don’t want to talk to a robot’, and that is also fine,” she said. Hughes also acknowledged that automated scribes can produce inaccurate notes, but stressed that doctors remain responsible for checking and correcting them. For these tools to earn a permanent place in the clinical toolkit, human oversight cannot be treated as an optional extra.
Extraordinary potential, serious risks
The medical possibilities extend well beyond transcription. AI systems may help clinicians detect disease in images, prioritise urgent cases, accelerate drug discovery and identify patterns across enormous research datasets. Professor Alastair Denniston, an ophthalmologist who also worked on the commission, described the technology as “an exceptional opportunity” for healthcare “likely to rank alongside step-changes such as antibiotics and MRI”.
Technology leaders are making still bolder predictions. In September, Rene Haas, chief executive of British chip designer Arm, told the BBC he believed AI would develop a cure for cancer “within our lifetime”. That prospect helps explain the excitement surrounding medical AI, from hospital imaging platforms to research systems capable of scanning billions of molecular interactions.
Yet the same adaptive machinery that makes these products powerful can also make them unpredictable. Models trained on biased patient data may produce worse decisions for underrepresented groups, while AI chatbots can deliver incorrect medical advice with unnerving confidence. A regulatory system fit for this technology must therefore test not only whether an algorithm works, but for whom it works, how it changes and what happens when it gets things wrong.
Why it Matters
AI could make the NHS faster, sharper and more personalised, but its success will depend on far more than processing power. Patients must understand when algorithms influence their care, clinicians need reliable tools rather than opaque black boxes, and regulators must be able to intervene after approval when software begins to drift. If Britain can protect those principles while allowing safe innovation, its new framework could become a blueprint for healthcare systems worldwide; get it wrong, and mistrust could stall one of medicine’s most transformative technological shifts.
