Stale AI approvals: Medical AI keeps changing; its safety checks haven't kept up
Before a hospital or clinic can use an AI tool, the tool is tested and approved. That happens once. Afterwards, the rules mostly wait for someone to report that something went wrong. But an AI tool does not stay the way it was tested. It gets updated. It behaves differently with real patients than it did in the test. And it can work well in one place and less well in another. So a single test, done once before the tool went into use, cannot tell you whether it is still safe where you are treated today. What is missing is a way to keep testing it while it is in use, to update it carefully, and to help hospitals and clinics see for themselves how it is doing.
Healthcare provider governance teams · Healthcare professionals using AI · Patients receiving AI-supported care
Why now
The Medicines and Healthcare products Regulatory Agency’s report, published on 10 September 2026, warns that reliance on one-off pre-market assessment is insufficient and calls for more continuous monitoring after deployment.
Why it matters
Existing assessment approaches suit relatively static products, while continuing safety depends on both the product and its clinical setting. A small team could begin with measuring changes in performance for a single clinical use across deployments, working with a provider to connect those measurements to review and escalation responsibilities.
From public reporting
Framed from public reporting. Here is what it was drawn from, so the framing can be checked against it:
Problem Radar listens to public reporting, from institutions, research and the press, for problems worth solving, and frames each one to the same standard. Every problem says where it came from, so the framing can be checked. Nothing on the Index comes from a user, and what you bring to Ainna is never published.
Caring for an ageing parent means medications, appointments, home visits, bills and forms, shared between siblings in different cities and often different countries. It runs on a group chat, a shared folder and whoever remembers. Things fall through, one person quietly carries most of it, and the professionals involved each see a fragment of a picture nobody holds whole.
Resistance is as much a data problem as a drug problem. Surveillance depends on cultures, and much of the world's prescribing happens where there is no microbiology laboratory within reach, so the clinician makes a reasonable guess, and what happens next is rarely recorded anywhere that would improve the next one. The map of what still works is drawn from the minority of cases that reach a lab.
Loss happens at the handovers: field to packer, packer to distributor, distributor to shelf. Every party measures only its own side of each one. So each set of numbers looks defensible while the total is enormous, and the interventions that would work are precisely the ones nobody can see, because they sit between two businesses rather than inside one.
