UK Funds AI Device Rules

Britain just put new money behind a problem every hospital using artificial intelligence faces: software can change after launch, but a medical device is supposed to keep behaving the way regulators approved it. The Medicines and Healthcare products Regulatory Agency said on April 8 it had secured £3.6 million over three years to expand its “AI Airlock” work on rules for artificial intelligence used as medical devices. (gov.uk) That agency is the British regulator for medicines and medical devices, and the “AI Airlock” is its test bed for companies building software that helps diagnose, triage, or guide treatment. The programme started in spring 2024 as the United Kingdom’s first regulatory sandbox for artificial intelligence as a medical device. (gov.uk) A regulatory sandbox is basically a practice field with referees watching every move. Instead of letting a company guess what evidence regulators want, the company and the regulator work through real edge cases before the product spreads through the health system. (gov.uk) The edge cases are unusually hard with medical artificial intelligence because the software may learn from new data, rely on hospital workflows, or produce different risks in different patient groups. The Medicines and Healthcare products Regulatory Agency has said its reform programme covers the full life cycle, from classification before sale to monitoring after the product is already in use. (gov.uk) Britain has been building these rules for several years, not starting from scratch this week. The agency’s Software and Artificial Intelligence as a Medical Device Change Programme, first published in 2021 and updated afterward, laid out work on qualification, risk classification, pre-market evidence, change management, and post-market surveillance. (gov.uk) The first Airlock pilot ran from April 2024 to March 2025 with the National Health Service and the Department of Health and Social Care involved. The pilot report later highlighted concrete problems like synthetic training data, large language model hallucinations, explainability trade-offs, and continuous real-time monitoring after deployment. (gov.uk) Those examples show what regulators are really trying to pin down. If an imaging tool is trained on artificial scans instead of real ones, or if a chatbot-style clinical tool occasionally invents an answer, the question is no longer just “does it work in a lab,” but “how do we keep checking it safely in the wild.” (gov.uk) The new funding also changes the tempo. The government said the Department of Health and Social Care will provide about £1.2 million a year from 2026 to 2029, which lets the programme operate beyond one-year budget cycles and scale after a second phase. (gov.uk) That second phase was designed around three recurring headaches: how to regulate systems that evolve, how to assess artificial-intelligence diagnostics, and how to run strong post-market surveillance once hospitals start using the tools. The Medicines and Healthcare products Regulatory Agency said phase 2 testing would run until March 2026. (gov.uk) For companies selling imaging software, this points to a tougher market than the old “get cleared once and start selling” model. The products that win are more likely to be the ones with audit trails, monitoring plans, validation data, and a clear process for handling updates after release. (gov.uk) For the National Health Service, that means trust will increasingly come from governance paperwork as much as from glossy demo results. Britain is trying to turn that paperwork into a formal rulebook before artificial-intelligence medical devices become routine across radiology, triage, and clinical decision support. (healthcare-management.uk)