Cleared once, learning forever

A medical device is tested, cleared and then trusted, with regulators watching for trouble after launch. That premise is comfortable only because it assumes the device stands still. On September 9th a British commission handed the MHRA, the country's drug and device regulator, 44 recommendations that attack the premise head-on, urging it to stop clearing artificial intelligence in medicine once and to supervise it instead across its working life.

The problem that forces the rethink is structural, not bureaucratic. Medical-device approval assumes a locked algorithm: a diagnostic tool is shown to work at the moment of authorisation, and its behaviour is thereafter assumed constant. An adaptive AI is different. It learns from its own outputs after it ships, so its weights — and therefore what it actually does to a patient — drift as it meets new data. A model cleared in one hospital may behave differently in another, in one population versus the next. A device that changes its own rules is, functionally, a new device every time it updates. Safety and liability can no longer be settled at the point of clearance, because there is no stable "it" to clear.

The commission's remedy is a lifecycle-based regime. It proposes staged authorisation of new models to accommodate their evolution, and a new system under which the hospitals that deploy the tools report their real-world performance, including cases where the AI misbehaves or inhibits care. The ambition shows in the examples officials reach for, such as scanning millions of eye images to track diabetes-linked retinopathy in NHS patients, a task too large to validate against a frozen snapshot of a model. The commission is chaired by Alastair Denniston, a clinician-researcher, and builds on the MHRA's "AI Airlock" sandbox, which has been probing where standalone AI medical devices trip over the existing rules.

Two regulators, two wagers

Across the Atlantic, the other leading regulator has made a different bet. The FDA, unsure how to evaluate generative-AI devices, has admitted four companies into a pilot called TEMPO that lets them launch products without full marketing authorisation, gathering real-world data while the agency settles the standards; two of the four, Cadence and Limbic, use generative AI. The FDA's own officials concede they do not yet know how they will evaluate the class. Britain proposes to regulate the thing it cannot fully know by supervising it; America proposes to launch the thing it cannot fully know and learn on the move. Both are improvisations against the same collapse of the once-for-all clearance.

To be sure, the instinct that "reform means faster access, therefore buy health-tech" is half right and half dangerous. Faster access is the intent. But watch who pays for continuous trust. A lifecycle regime with mandatory provider-side reporting turns every developer into a permanent evidence machine and every hospital into a monitoring station. That is a fixed cost: trivial for diversified incumbents with regulatory departments, onerous for the small pure-play startups that populate AI health-tech. Regulation of learning machines is not a neutral gateway. It is a mechanism that widens the moat around the GE HealthCares of the world, which already carry the burden, and raises the wall that entrants must climb.

The cost of trust

For the investor the live question is not whether AI medicine gets approved, but what approval will cost, and to whom. The direction of travel — continuous oversight, real-world evidence, developer and provider accountability — is a regulatory overhang on any adaptive-AI thesis, a recurring operating expense to be priced in beside the algorithm itself. AInvest's aggregate signal still labels GE HealthCare, a diversified device maker, a Buy, but such labels are expectations to test, not evidence. The safer conclusion is narrower. In medical AI the gatekeeper has become a recurring cost and a competitive verdict, and whoever holds a durable regulatory path holds a durable rent. That, not the cleverness of the model, is where value will be fought over.