Or: How the United Kingdom reinvented medical residency for AI and forgot to invent the boards
Congratulations are in order. The United Kingdom just matched its first resident class of algorithms.
Nobody used that word. The National Commission into the Regulation of AI in Healthcare, convened by the MHRA to design the future regulatory framework for medical AI, speaks in regulatory prose: staged authorisation pathways, sandboxes, lifecycle assurance, post-market monitoring. But read the architecture instead of the vocabulary and something familiar comes into focus. A surgeon would recognize it anywhere. It is a residency program.
The Curriculum They Accidentally Wrote
Walk through their framework and translate as you go.
A regulatory sandbox where the system practices in a controlled environment before touching anything real. We call that simulation lab. Staged authorisation that deploys the device within a tightly limited scope, under extra risk controls and reporting, after review of initial evidence. We call that supervised practice. Deployment only through healthcare providers who can demonstrate AI readiness: the knowledge, capability, and capacity to run the thing safely. We call those teaching hospitals. Privileges that expand once prespecified performance and safety thresholds are met. We call that graduated responsibility. A status that is explicitly temporary, with a defined path to full authorisation. We call that board eligibility. A central conclusion that a one-time pre-market assessment is not enough and that assurance must continue across the entire lifecycle. We call that maintenance of certification. (Passing the boards once was never a license for life.) And processes for escalating performance degradation. We call that a fitness-to-practice referral.
The report even opens with a commitment that AI should “augment, rather than replace healthcare professionals.” Regular readers may hum along.
Let me say clearly: this is the right instinct. Medicine’s training pipeline is the most battle-tested competence-production system civilization runs, debugged over a century by the least forgiving quality-assurance process there is: dead patients. If you need to make an adaptive, context-dependent, occasionally overconfident intelligence safe to practice medicine, of course you rebuild residency. We already solved this problem. For humans.
The Two Organs They Forgot
But this residency program is missing two organs, and they are the two that make residency work.
No examining board. In a real residency, graduation runs through an independent, standardized examination written by people with no stake in whether this particular resident passes. In staged authorisation, the graduation criteria are evidence thresholds agreed between the manufacturer and the regulator, device by device. That is a residency where each resident negotiates their own board exam with their program director. Ask any program director how that would go.
And the evidence those thresholds would lean on is worse than thin. The published benchmarks for autonomous medical AI share a quiet flaw: the AI-alone arm is never alone. A physician is hidden inside every vignette. Someone took the history, performed the exam, and decided which findings mattered enough to write down. Clinical data acquisition is an active loop of asking, looking, touching, and smelling. The benchmark assumes that loop away, then grades the machine on the leftovers. Meanwhile, when we tested a leading medical AI open-book, on board-level questions derived from articles it confirmed it could access, it could not pass. Access is not mastery. A residency graduates on mastery.
No revocable license. The report includes escalation processes for when performance degrades. What I could not find is the thing escalation is supposed to lead to: suspension. Revocation. The regulatory equivalent of the medical board pulling your ticket. Accountability in medicine is not a paragraph in a framework. It is three load-bearing parts: someone owns the consequences, holds a credential that can be revoked, and answers to an independent adjudicator. Remove any one and you have liability theater.
To its credit, the Commission gets closer than anyone to date. It states plainly that clinicians should not be asked to carry accountability for risks they cannot plausibly mitigate. That is the right principle, and I have waited years for a regulator to write it down. Now finish the sentence. If accountability moves off the clinician, it must land on a credential that can be taken away.
How to Finish the Program
If AI wants physician privileges, give it the whole physician pathway, not the flattering half.
An examining board that is independent of the manufacturer, testing against validated content under confirmed conditions, with questions the system has never seen because they did not exist yesterday. (Regenerable, content-derived assessment: the lab-grown diamonds I described last time. The instrument for this residency already exists. Medicine built it.) A credential that expires, gets re-earned, and can be revoked, version by version and scope by scope, the way my license is state by state and my certification is specialty by specialty. And an adjudication body that can remediate the resident, fail the resident, or expel the resident, and publishes when it does.
The Bottom Line
The Commission conceded the deepest point in this debate, probably without noticing. You do not make an intelligence safe to practice medicine with a product approval. You make it safe with a training pipeline: supervised scope, graduated privileges, independent examination, continuous recertification, and a license that can die. Medicine is the only industry on earth that already runs that pipeline at scale.
So the invitation writes itself. Regulators do not need to invent the boards. They need to call the people who have been writing them all along.
The machine matched. Now somebody has to fail it on purpose, at least once, to prove the program is real.
