A junior doctor in a busy acute trust opens a large language model on her phone between ward rounds to check a drug interaction she half remembers from a lecture two years ago. Nobody told her to. Nobody told her not to. Her deanery has no guidelines, her trust has no practice policy, and the instrument she is utilising has no clinical responsibility framework at all. This scene, unremarkable and repeated daily across the NHS, is the gap that a newly published international framework has set out to close, and the gap that Britain's own regulators are still working out how to fill.

The Health CARE-AI framework, developed through a three-phase consensus process involving over three hundred clinicians, educators, ethicists and patient representatives worldwide, organises responsible AI use in medicine into four domains: values, competence, accountability and structural equity. Its most striking feature is not the breadth of its ambition but its insistence on specificity. Rather than issuing another statement about AI's potential to transform care, it treats artificial intelligence as a present actor in the clinical encounter, one with its own consent implications, privacy obligations and bias risks, that must be named and managed rather than assumed away. Ninety-six per cent of participants agreed it clearly defines professional expectations. That level of consensus is rare in any field, let alone one moving as quickly as this.

Britain finds itself in an odd position relative to this work. The MHRA's National Commission into the Regulation of AI in Healthcare, chaired by Professor Alastair Denniston and supported by the Patient Safety Commissioner, Professor Henrietta Hughes, has spent the past year gathering evidence with the aim of reporting recommendations this summer, under the oversight of MHRA chief executive Lawrence Tallon. That process matters, but it has been built largely around the logic of product safety: classifying software as a medical device, assessing evidence thresholds, adapting procurement rules. The question that the new framework poses at its core, what happens in the clinician-patient room when an AI tool is subtly involved in the decision-making process, has not yet been answered. This is true whether the tool is an unauthorised reference app that a doctor uses out of habit, an ambient scribe, or a diagnostic aid.

Recent research into how GPs and hospital clinicians actually use these tools has found precisely the inconsistency this framework is designed to prevent. Some integrated care boards have prohibited generative AI outright. Others actively encourage piloting with little shared standard for how errors are documented or how liability is understood when a model produces a confident but wrong answer. Clinicians are left improvising professional judgment in an area where professional judgment has not yet been given a vocabulary. That is an educational failure as much as a regulatory one, and it is where the framework's competence domain, on maintaining critical human judgment as a taught and assessed skill rather than an assumed virtue, has the most to offer.

The timing is not incidental. Yvette Cooper's appointment as Health and Social Care Secretary under Prime Minister Andy Burnham has brought fresh attention to how the department intends to sequence digital reform against the more immediate pressures of maternity safety and social care, the two priorities she named on taking office. A framework that asks trusts to treat AI as a professional discipline rather than a procurement category sits awkwardly against a department already stretched across competing emergencies, and any adaptation of Health CARE-AI's principles will need to be argued for on its own merits rather than assumed to follow naturally from ministerial goodwill.

All of this does not imply that NHS policy should adopt the framework in its entirety. Consensus documents built for global applicability tend to sit above the specific realities of a health system as centralised, as financially constrained and as politically exposed as the NHS. Any UK adaptation would need to reckon with workforce shortages that make additional training burdens unwelcome, and with a public whose trust in NHS data handling has already been tested by earlier disputes over platform contracts. A framework written by consensus is not the same as a framework that survives contact with a struggling acute trust on a winter night.

What the publication does offer, regardless of adoption, is a working demonstration that AI governance in medicine need not be reduced to device classification and information governance policy. It can be written as a professional discipline, taught, supervised and revisited, in the way clinical ethics already is. Denniston's commission has an opportunity to borrow that framing when it reports. Whether it does will say a good deal about whether Britain intends to treat AI in the NHS as a compliance exercise or as the practice-shaping change its own clinicians increasingly say it already is.