AI medical devices in Japan, Great Britain and the EU: where each regulator stands in October 2026

Since July, three regulators have published something about AI-enabled medical devices. Each publication had a different legal status. Japan put out a draft for comment. Great Britain made a set of government commitments. The EU changed the law. If you are an AI or software device company choosing your first markets, those differences matter more than any headline.

This piece compares the three and ends with an order of work. It is written for regulatory leads at device companies and for founders who are about to make their first filing outside their home market. Every fact below was checked against the authority's own document, and we say so where we could not do that.

Why these three. Japan and Great Britain chose themselves: both published in October. For the third we picked the market with the most recent binding instrument specific to AI devices. That was the EU. Regulation (EU) 2026/1744, adopted on 8 July 2026, sets the date from which the AI Act's high-risk rules apply to medical devices. The US was the alternative. FDA's most recent AI-device instrument is a final order of 11 September 2026, but it covers one device type, and FDA's general AI lifecycle guidance is still the draft it published in January 2025. Korea's generative-AI review guideline dates from January 2025. The US is still the reference point for most AI device companies, so we bring it in where it helps.

The short version

	Japan	Great Britain	European Union
Newest AI instrument	Draft review points for generative-AI devices (PMDA, dated 14 September 2026)	Government response to the National Commission on AI in healthcare (6 October 2026)	Regulation (EU) 2026/1744, the Digital Omnibus on AI (in force 27 July 2026)
Legal status	Draft. Comments close 17 March 2027.	Commitments. No change to the law yet.	Binding law.
Planned changes after approval	IDATEN change plans, available since 2020, with extra documents required for AI	Draft PCCP guidance promised by December 2026	Pre-determined changes are not a "substantial modification" under the AI Act (Article 43(4)). In practice this applies to devices from 2 August 2028, when they first count as high-risk.
Who holds the authorisation	A Japanese marketing authorisation holder	The manufacturer. A non-UK manufacturer needs a UK Responsible Person.	The manufacturer. A non-EU manufacturer needs an authorised representative.
Does a foreign approval help?	No reliance route. MDSAP audit reports are accepted for the QMS inspection.	CE marks accepted until 30 June 2030 (MDR). Reliance routes based on FDA approvals are proposed, not in force.	No.

Japan — a first draft for generative AI, and a change-plan route that already exists

What changed. PMDA's Center for Product Evaluation published draft "Review Points for Medical Devices Utilizing Generative AI". The Japanese draft is dated 14 September 2026 and opened for comment on 18 September. Comments close on 17 March 2027. An English provisional translation followed on 14 October. The draft covers devices built on large language or vision-language models that answer open questions in natural language. It does not cover task-specific AI such as a single-finding image classifier.

PMDA is open about how early this is. The draft says that no generative-AI device has yet been reviewed for approval by PMDA. It calls its content provisional and says it is an invitation to comment. It is not a rule.

Read in full, the draft does five things a developer should note.

• It accepts that review before approval cannot cover everything. PMDA proposes a "Total Life Cycle Approach". Approval would rest on performance under defined conditions, as long as the applicant runs a system after launch that actively looks for incidents and responds to them.
• It suggests starting narrow. PMDA suggests getting a first approval for a narrowly defined intended use and widening it later. It does not expect an all-purpose clinical assistant to be validated in one go.
• It expects you to report hallucinations after launch. Applicants are expected to collect data on performance in use, on adverse events and on hallucinations, and to report periodically. Where the device acts with some autonomy (agentic AI), PMDA also expects execution logs to be monitored.
• It treats an update to the underlying model as an open question. When a third-party model behind a cloud API is updated, the draft asks whether the device is still the same device. It expects the applicant's contracts and service-level agreements with the model provider to be part of how the risk is managed.
• It is written about Japanese patients and the Japanese language. The draft cites Japanese-language benchmarks. It warns that external data, giving the example of health information about Western populations, can produce wrong answers for Japanese patients. That is our reading of what PMDA will expect, not a stated requirement: plan for evaluation in Japanese and on Japanese data.

What already exists, and matters more for most AI companies. Japan has had a route for planned changes since 2020: IDATEN, a change-plan confirmation procedure. PMDA confirms the change plan during review. Changes made within that plan are then filed as a notification, not a partial-change application. PMDA states it checks within 30 days that the data and test results match the plan. Under the PMD Act Enforcement Regulations (Article 114-45-2), a plan for a device that uses AI must also include the procedures for drawing up and carrying out the plan. Changes that need new clinical data, such as a new indication, are outside the route.

Who it affects. Any AI device company considering Japan. The generative-AI draft affects only the small group building on large models. IDATEN affects everyone who expects to retrain.

What to do. If you build on a large model, read the draft and send comments by 17 March 2027. Few texts like this exist, and the final version will be shaped by the comments it receives. If your AI is a conventional classifier, ask PMDA's Office of Software as a Medical Device about an IDATEN plan at your first consultation. Choose your Japanese marketing authorisation holder carefully, because that company, not you, holds the approval.

Great Britain — 44 commitments, and nothing in force yet

What changed. On 6 October 2026 the government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare, in a 42-page response produced by the MHRA. It states plainly that the current device regulations are "not fit for purpose" for AI. Most of its commitments say "further details to be shared by Spring 2027". Three have dates and matter to a manufacturer:

• Change control. The MHRA will issue draft PCCP guidance for public comment by December 2026. It proposes letting manufacturers set boundaries for the changes they may make, rather than listing each specific modification in advance. That would go further than the FDA model. It explicitly includes adaptive AI.
• Classification. The MHRA intends to consult by Spring 2027 on qualification and classification of software and AI. It will then use secondary legislation for a new device definition, which may include carve-outs for some administrative, wellbeing and decision-support software. It will also introduce a new classification approach.
• General-purpose models. The MHRA will look at "master files" for general-purpose AI models, possibly starting with a pilot. It will also set out what a submission must say about a device's dependencies on such a model. This is the same problem PMDA raises. Two regulators now name it as a gap.

The response also commits to recognition and reliance arrangements for AI devices with "recognised international regulators". It gives no route and no date.

AI Airlock phase 3 is open. The MHRA's regulatory sandbox for AI devices is taking applications. Apply before 31 October to be considered in the current review wave. The government says the first wave will be selected in November 2026, and a further shortlisting is planned for Spring 2027. Testing plans usually run 6–12 months, and the phase is funded until April 2029. Taking part does not grant an approval.

Two October items that are not news. The MHRA's "Good machine learning practice" page was updated on 5 October 2026, but the change note says only that a contact email address changed. The ten principles date from 2021. Their international version is IMDRF's final document N88, issued in January 2025. The digital mental health technology collection, updated on 8 October, added resources and no new requirement.

The rules today. Great Britain still classifies software under UK MDR 2002, which uses the old EU Directive rules. In the MHRA's own software guidance, software for diagnosis is Class IIa when it allows direct diagnosis or monitoring of vital physiological processes. Software that does not drive a device and does not fall under another rule is Class I, which the manufacturer self-declares. EU MDR Rule 11 puts most software that informs a diagnostic or treatment decision at Class IIa or above. So some AI software that needs a notified body in the EU is Class I in Great Britain today. The MHRA has said it will consult on changing that, so treat this as a window that will close, not a strategy. CE-marked devices under the MDR are accepted in Great Britain until 30 June 2030. The reliance routes in the draft 2026 Regulations are not law. They would send software devices to a review by a UK approved body, which would also check whether any PCCP is appropriate. The MHRA says they will start "from 2027".

What to do. If Britain is on your list and your product raises a question the rules do not answer, apply to the Airlock by 31 October. Read the draft PCCP guidance when it arrives in December. It may be the most flexible change-control model of the three, and you can comment on it before it is fixed. Do not plan a UK filing on the assumption that FDA reliance will exist by a particular date.

European Union — the AI Act moved to 2028, and a proposal would take devices out of it

What changed. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It moves the date from which the AI Act's high-risk requirements (Chapter III, Sections 1–3) apply. For AI systems that are high-risk because they are, or are part of, products under Annex I, the new date is 2 August 2028, a year later than before. Medical devices and IVDs are in Annex I. AI systems that are high-risk under Annex III move to 2 December 2027. The same regulation moved machinery out of direct AI Act application, but it left the MDR and IVDR where they were.

Who it captures. Under Article 6(1), an AI system in a medical device is high-risk only if the device needs a third-party conformity assessment, meaning a notified body. Because Rule 11 places most decision-support software at Class IIa or above, most clinical AI software is captured. The notified body that certifies the device under the MDR would also check the AI Act requirements (Article 43(3)). The AI Act also gives the EU its nearest equivalent of a PCCP, which in practice reaches devices when they first count as high-risk in 2028. For a system that keeps learning after it is placed on the market, changes the provider pre-determined at the initial conformity assessment are not a substantial modification (Article 43(4)). The MDR itself has no equivalent mechanism. A significant change to certified software still goes back to the notified body. That is our reading of the MDR; the AI Act wording is quoted from the Regulation.

The proposal that could change this again. The Commission's targeted revision of the MDR and IVDR, COM(2025) 1023 of 16 December 2025, would move both regulations from Section A to Section B of the AI Act's Annex I. Devices would then fall only under the limited provisions in Article 2(2), and any AI-specific requirements would come through implementing and delegated acts under the MDR and IVDR. As at 19 October 2026 that file (2025/0404(COD)) is awaiting a committee decision in Parliament, with an indicative plenary date of 14 December 2026, and the Council has still to act. It is a proposal, not law.

Who it affects. Every AI device company that needs a notified body for the EU, and through the CE mark, for Great Britain until 2030.

What to do. Certify under the MDR now. Do not wait for the AI Act question to be settled. Build your quality system so it can take on AI Act obligations by 2 August 2028 if the proposal fails. The Commission's FAQ on how the MDR, IVDR and AI Act interact (MDCG 2025-6, June 2025) is the place to start. Plan against the Regulation in force, not against the proposal.

What carries across, and what does not

What carries across. Three things, and together they are most of the work:

• Good machine learning practice. The ten IMDRF principles (N88, January 2025) are the shared vocabulary of all three regulators and FDA. If your design history can be mapped to them, every reviewer can follow it.
• Evidence built to the strictest common standard. FDA's final order for cardiovascular machine-learning notification software (21 CFR 870.2380, 11 September 2026) shows what a regulator means by good evidence. It requires a test set independent of training data, from at least three geographically diverse sites that were not used in training. It also requires performance reported by subgroup and a human-factors assessment of how users may misread the output. Build to that standard once, and you will not need to redo the work for each reviewer.
• One change-control plan, written once. FDA's PCCP guidance (final; current version August 2025) sets out the parts: a description of the modifications, the protocol for making them, and an assessment of their impact. IDATEN in Japan, AI Act Article 43(4) and the coming MHRA guidance all ask for the same substance in different formats. That is our comparison; no regulator has said the documents are interchangeable.

What does not carry across.

• Classification. The same algorithm can be Class I in Great Britain and Class IIa or higher in the EU.
• Who holds the licence. In Japan the marketing authorisation holder holds your approval. If that is your distributor, changing distributor means moving the approval. Pick an independent holder unless you have a reason not to.
• Local data. For a large-language-model device in Japan, expect to show performance in Japanese.
• Reliance. None of the three has a reliance route in force that accepts your FDA clearance today.

An order of work for an AI device company

This is our recommendation for a company that already has, or is about to get, a US clearance. It is not regulatory advice for any particular product.

1. Narrow the intended use before anything else. It drives classification in Britain and the EU and the scope of review in Japan. PMDA explicitly suggests starting narrow.
2. Write the change-control plan into your first submission. Use the FDA PCCP structure. It is the only one of these models already final, and it is the easiest to adapt to the others.
3. Take Europe next through the CE mark, if Europe is on the plan. One notified-body certification opens the EU and, until 30 June 2030, Great Britain. If your software is Class I under Great Britain's current rules, a UKCA self-declaration is cheaper and faster. That only makes sense if Britain on its own is worth it to you, and the classification consultation may remove the option.
4. Do the AI Act gap analysis by 2027, not 2028. Notified bodies will be busy near the deadline. If the December vote moves devices to Section B, you have lost little. Most of the work overlaps with MDR technical documentation you will already have.
5. Treat Japan as a project of its own. There is no reliance route, you need a local holder, and you may need local data. In return, IDATEN can make retraining cheaper after approval than it is in the EU. Start with a PMDA consultation, not a filing.

What our own data can and cannot tell you

We tried to answer two questions from our registration corpus, as at 19 October 2026. The answers show where public registration data stops.

How many AI devices are registered in each market? We could not answer this honestly. No registry in these markets records whether a device uses AI. In the US, FDA has product codes for the software functions where most AI devices sit, such as triage, detection and automated image processing. Twelve of those codes cover 636 clearances from 267 companies in our US corpus. FDA says it had authorised over 1,600 AI-enabled devices as of September 2026. Many of the rest sit under broad codes that mostly contain software that is not AI. In our Japanese corpus, 760 registrations name themselves as software (プログラム) in the product name or the Japanese generic name. Only 26 distinct product names among them say "AI". Our British data records device types, not products, and no row names AI.

Which US-cleared AI devices are also registered in Japan, Britain or the EU? We can answer this at company level for the EU. We cannot answer it at device level anywhere. Of the 267 companies, 157 hold at least one EU registration and 73 hold an EU registration for software. Only 19 appear in Japan. That number is an artefact. A Japanese registration usually names the Japanese marketing authorisation holder, or writes the foreign maker's name in Japanese script. Either way, a foreign AI developer's approval does not match on its own company name. No identifier links a 510(k) to the same product's registration in another market. Anyone who tells you a specific AI device is "not registered in Japan" because a name search came back empty is reading the data wrongly.

The calendar

• 22 October — MHRA AI Airlock phase 3 webinar
• 31 October — apply by this date for the current AI Airlock review wave
• November 2026 — first phase 3 Airlock wave selected
• December 2026 — MHRA draft PCCP guidance due for comment
• 14 December 2026 — indicative European Parliament plenary on the MDR/IVDR revision, including the AI Act move
• 17 March 2027 — comments close on PMDA's generative-AI review points
• Spring 2027 — MHRA classification consultation and the UK implementation roadmap
• 2 August 2028 — AI Act high-risk requirements apply to AI medical devices, unless the proposal is adopted first
• 30 June 2030 — last day Great Britain accepts MDR CE marks

Sources

• PMDA, Review Points for Medical Devices Utilizing Generative AI (Draft), 14 September 2026; English provisional translation, October 2026; public comment page (comments 18 September 2026 – 17 March 2027).
• PMDA, Overview of the IDATEN System (SaMD seminar material, 18 September 2026), citing PMD Act Enforcement Regulations Article 114-45-2.
• MHRA, Government Response to the National Commission's Recommendations on the Regulation of AI in Healthcare, 6 October 2026; AI Airlock Phase 3 application, 6 October 2026.
• MHRA, Good machine learning practice for medical device development: guiding principles (2021; page updated 5 October 2026); IMDRF/AIML WG/N88 FINAL:2025.
• MHRA, Medical device stand-alone software including apps (including IVDMDs), v1.10g.
• Regulation (EU) 2026/1744 of 8 July 2026 (OJ L, 24.7.2026); Regulation (EU) 2024/1689, Articles 6(1) and 43; MDCG 2025-6 (June 2025).
• European Commission, COM(2025) 1023 final, 16 December 2025; European Parliament procedure file 2025/0404(COD), read 19 October 2026.
• FDA, final order classifying cardiovascular machine learning-based notification software, 21 CFR 870.2380 (FR Doc. 2026-18612, 11 September 2026); PCCP for AI-enabled device software functions, final guidance (docket FDA-2022-D-2628, August 2025); AI-enabled device software functions lifecycle guidance, draft (docket FDA-2024-D-4488, January 2025).

A regulator's position is easy to read. Deciding what it means for your products across markets is harder: where you file first, who holds each licence, and which evidence you can reuse. That is the question our market assessments are built to answer.