A medical AI tool can look transformative in a demonstration and still be unready for a clinic. The UK’s Medicines and Healthcare products Regulatory Agency is trying to close that gap through the second phase of AI Airlock, a controlled testing environment where developers and regulators examine emerging systems together.
Seven technologies are entering the new cohort. Their proposed uses range from reducing the administrative burden of clinical notes to detecting cancer and inherited eye disease, interpreting complex blood tests, and summarising an entire hospital stay. One could reduce the wait for bowel-cancer test results from weeks to minutes.
The promise is speed, but the programme’s name captures the necessary restraint. Before an AI reaches patients, it enters an environment where performance, limitations, and risks can be exposed without treating ordinary clinical use as an experiment.
Seven Tools Face Real Healthcare Questions
The selected systems address both diagnosis and the flow of information around care. Earlier detection of skin cancer or genetic eye conditions could open treatment options sooner. Faster interpretation of tests could shorten an anxious wait. Automated note-taking and hospital-stay summaries could give clinicians more time while making complex records easier to navigate.
Each application also creates a different safety problem. A diagnostic model must recognise disease without producing an unacceptable stream of false results. A generated summary must preserve the details that matter for the next decision. A note-taking system must accurately capture a conversation while respecting the sensitivity of patient information.
AI Airlock lets manufacturers test those questions within a regulatory sandbox. The evidence will help establish a path toward approval and NHS deployment rather than leaving developers to guess how general medical-device rules apply to adaptive, data-driven systems.
Regulation Develops Alongside the Technology
The first AI Airlock cohort included Philips and OncoFlow, and the MHRA has published four reports on the work. Participants identified unresolved issues such as validating synthetic training data, making model decisions explainable to clinicians, and addressing hallucinations that produce incorrect or nonsensical content.
Those findings are feeding into future rules for AI as a medical device and the National Commission into the Regulation of AI in Healthcare. The Commission includes patients, clinicians, regulators, and technology companies, reflecting the fact that acceptable risk cannot be determined by model developers alone.
Philips representative Yinnon Dolev described weekly contact with the MHRA as a catalyst for development and a rare opportunity for research teams to influence regulatory strategy. That interaction benefits both sides: innovators learn what evidence is required, while regulators encounter technical limitations before writing rules that may shape an entire market.
Trust Must Advance at the Same Pace
The government wants the NHS to become highly AI-enabled, and Health Innovation Minister Zubir Ahmed presented Airlock as a way to test thoroughly while moving quickly. Those ambitions will only reinforce one another if patients and frontline staff believe that speed has not displaced safety.
Sir Andrew Goddard, a consultant gastroenterologist and chair of the programme’s Governance Board, said clinicians are interested in AI but concerned by overpromising and inadequate reassurance. His perspective identifies the practical test for the sandbox: not merely whether a tool earns formal approval, but whether a clinician can understand its role, recognise its limitations, and responsibly use its output.
MHRA chief executive Lawrence Tallon says the UK is the first country to create a dedicated regulatory environment specifically for AI medical devices. The advantage of that early move is learning through concrete systems rather than debating hypothetical risks indefinitely.
AI Airlock is therefore not simply an accelerator. It is a place where excitement meets evidence. If the seven tools emerge with clearer boundaries and credible performance, patients may receive results faster and clinicians may lose less time to administration. Just as importantly, the rules governing the next cohort will be informed by what happened when promising AI was placed under clinical pressure.
