TL;DR
New AI tools heading into the NHS would earn their approval in stages rather than all at once, under recommendations published this morning by a commission of practising doctors. The same report asks for lifetime monitoring of every AI-enabled device, a searchable public record of safety incidents, and sharper enforcement powers for the MHRA. Government has promised a formal reply later.
Approval that arrives in stages
The MHRA convened the National Commission a year ago, asking it to examine how healthcare AI ought to be regulated. Chairing it are Professor Alastair Denniston and Professor Henrietta Hughes, both of whom still see NHS patients.
Its central proposal borrows from the driving test. A model would launch closely supervised, hemmed in by constraints, prove itself on real patients, and only then graduate to unrestricted use. The report calls these “L-plates”.
That reframes what approval means. At present a device clears the bar once and is largely finished with the process; the Commission argues this suits software that keeps changing very badly, and wants scrutiny running for as long as the tool stays in service.
Two further asks have teeth. Patients would be able to look up safety information on individual AI devices, adverse incidents included, modelled on the interactive Drug Analysis Profiles the regulator already publishes for medicines. And the MHRA itself would gain enhanced powers to act where a system falls short of expectations.
Built on unusually broad evidence
Some 12,000 people fed into the work over that year: patients, clinicians, NHS leaders, developers, and contributors from as far afield as Singapore and the United States. The MHRA calls it the biggest consultation this country has run on health technology rules, with separate deliberative research from the Health Foundation out alongside it. Backing for AI in care came with conditions attached — safety standards, human oversight, and being told when a machine is involved in your treatment.
Looking forward
Nothing here binds anyone yet. Ministers and the regulator will respond formally in due course, and that reply is where the actual policy gets settled.
The direction of travel is visible elsewhere already. This morning the same regulator opened a live testing sandbox with Manchester University NHS Foundation Trust, while eight trusts started a year-long surgical AI trial. Three announcements before lunch, all circling one question: whether these tools hold up in a working hospital rather than on a validation set.
Suppliers should read the cost implication carefully. Continuous monitoring is a standing obligation, not a launch expense, and it lands on whoever built the model.