THE MED-AI BRIEF
{{current_date_full}} · 5-MINUTE READ · Read online
Good morning, {{first_name| }}. Two issues running, this brief counted the months since the National Commission's report on AI in healthcare was due. On 10 September the count ended: 44 recommendations, a staged authorisation model borrowed from the driving test, and a promise that patients will be told when AI is treating them. None of it is law yet.
In today's Med-AI Brief:
🩺 Deployment: the NHS's biggest AI scribe rollout runs on the honour system
⚖️ Regulation: L-plates for AI, and a promise to tell you
🌍 Use cases: Saudi Arabia already built the database the UK is proposing
🛠️ Practice: find your own disclosure gap
USE CASES
🌍 Saudi Arabia already built the database the UK is proposing
The brief: The public searchable device database the Commission wants already exists elsewhere. Saudi Arabia's SFDA runs a Medical Device National Registry, public and searchable through the government's my.gov.sa portal, and a device without an MDNR number cannot legally be marketed or imported.
The details:
The SFDA has run AI/ML-specific device guidance, MDS-G010, since January 2023, aligned with FDA and IMDRF frameworks and built for devices that keep changing after they reach the market, the same problem the Commission's lifecycle model is trying to solve.
Registration through the MDNR is not an innovation programme. It is the baseline condition for selling a device in the Kingdom at all.
The UK's equivalent, the MHRA's AI Airlock sandbox, has run since 2024 on £3.6m of funding through 2029 and its findings are feeding directly into the framework the Commission wants built.
Why it matters: What the UK is proposing as a forward step, Saudi Arabia treats as an entry requirement that predates this newsletter. The two systems will not arrive at the same transparency by the same route: Saudi built the database first and made lifecycle change control a condition of market access, while the UK consulted 12,000 people and is now waiting on government to decide whether to legislate. Worth knowing which model your employer answers to before you assume "searchable database" means the same thing in both places.
REGULATION
⚖️ L-plates for AI, and a promise to tell you
The brief: On 10 September the National Commission into the Regulation of AI in Healthcare published its recommendations, chaired by Professor Alastair Denniston with Professor Henrietta Hughes as deputy chair, after what it calls the largest public engagement on health-technology regulation the UK has run.
The details:
Over 12,000 patients, clinicians and technologists fed into the 44 recommendations.
Centrepiece: staged authorisation, nicknamed "L-plates" - a new AI model launches under close supervision and tight constraints, has to prove itself on real patients, and only then graduates to unrestricted use. Approval becomes a lifecycle, not a single sign-off.
Patients get a stated right to know when AI is used in their care, with a push for records to track which AI device touched a given case. A public, searchable database would carry safety information and adverse incidents on named devices.
Commission chair Denniston: "People are open to AI improving their care, but only if it is safe, overseen by humans, and if they know when it is being used." MHRA chief executive Lawrence Tallon welcomed the report, saying AI's opportunities need "a modern, dynamic regulatory framework."
Why it matters: This is a recommendation, not a rule. DHSC and the MHRA say they will "carefully consider" the findings, with a formal response due in due course and no date attached. That gap between recommended and required is the same one issue 004 flagged when OpenEvidence withdrew from the UK citing regulatory uncertainty. Until DHSC responds, treat every line of this as "proposed", not "in force", and watch the response date more closely than the report itself.
DEPLOYMENT
🩺 The NHS's biggest AI scribe rollout runs on the honour system
The brief: NHS Midlands has completed the largest ambient voice technology procurement the NHS has run, covering 1,239 GP practices and more than 70,000 clinicians across 15 trusts, with Heidi as preferred provider. It is already the live test of what "tell the patient" looks like at scale.
The details:
Under NHS England's ambient scribing guidance, developed with the ICO and the National Data Guardian, clinicians do not need explicit consent to use a scribe for individual care. They do have to tell the patient at the start of the session and let them object; if a patient dissents, the tool comes off.
That is disclosure at the point of care. It is not the same as a record that says which AI device was used, which is what the Commission is now recommending nationally.
Healthwatch's response to the Commission, published a day later, asks for exactly that gap closed: patient records that log which AI device touched a case, an explicit route to decline scribing for sensitive consultations such as sexual health or domestic abuse, and a clearer way for patients and staff to report AI safety concerns.
Why it matters: 70,000 clinicians are already running the "tell them" half of the Commission's transparency recommendation. The "record it" half does not exist yet, in this rollout or any other. If DHSC takes up the recommendation, this is the deployment large enough to show whether logging which device was used is a data-entry problem or an integration one.
PRACTICE
🛠️ Two minutes: find your own disclosure gap
Pick the next three patients you see. For each one, ask yourself honestly: would they know, right now, if an algorithm had a hand in their care? An ambient scribe, most trusts now tell them. A triage tool, an imaging prioritisation flag, or an AI-drafted discharge letter: most do not.
The Commission's recommendation is built on the assumption that disclosure already happens somewhere close to universally. It does not, and the gap is usually invisible to the clinician, because nobody has to declare a tool they were never told to declare. Three patients and two minutes will tell you whether your own practice matches the standard about to be proposed for everyone.
QUICK HITS
📰 Everything else
Response awaited: Healthwatch's reply to the Commission, 11 September, backed the recommendations and pushed further: a named route for patients to decline AI scribing in sensitive consultations, and better reporting when AI gets something wrong.
Trials: IQVIA launched Predictive Clinical Development on 3 September, an AI platform it says cuts trial start-up time by a third. Vendor-reported figures, no independent replication yet.
Funding: The MHRA's AI Airlock sandbox, running since 2024, now has £3.6m committed through 2029 to test voice tools, large language models and cancer diagnostics under regulatory supervision. Its findings shaped this week's report.
Still counting: The Commission's 44 recommendations now sit with DHSC and the MHRA, who say a formal government response will follow "in due course." No date given. We will keep counting.
How was today's brief?
✉️ One favour: hit reply and tell me whether your trust or clinic would pass the three-patient disclosure check above.
Until next week,
Saeed
Know a colleague who'd use this? Send them your referral link.
Nothing here is clinical or legal advice. Check any tool against your own information-governance rules before it touches patient data.