What does Japan’s PMDA expect when AI informs a drug or device decision?
Japan’s medicines and devices regulator says AI may support decisions across a product’s lifecycle, but humans retain responsibility. Its new principles emphasise traceable data, risk-based validation in the real use environment, monitoring and intervention—and are guidance, not a new legal requirement.
The short version
- What happened
- PMDA says AI should support rather than replace human judgement, with ultimate responsibility remaining with people and the interests of patients and the public placed first.
- Who it affects
- Patients should not assume that an AI-supported regulatory decision is automatic: PMDA says people retain responsibility and affected users should receive appropriate information about purpose, use and limits.
- What remains unproven
- The publication is a non-binding Early Consideration and explicitly does not establish new regulatory requirements.
Keep in mind
The publication is a non-binding Early Consideration and explicitly does not establish new regulatory requirements.
It contains no study sample, dataset, comparator, measured performance, patient outcome or implementation evaluation.
PMDA is describing principles for its own regulatory environment; the source is not independent evidence that those principles work.
The English text is a provisional translation, and the Japanese original prevails if wording differs.
Basic research and ordinary medical practice are not the document's primary focus, although regulated products and regulatory evidence can overlap with clinical use.
The framework leaves thresholds and evidence requirements to intended use and risk, so consistent implementation will need more detailed examples and review practice.
Source analysis · 3 linked sources. Primary analysed source date: 2 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 12 min The evidence, methods, implications and limitations
The direct answer: AI can assist, but it cannot own the decision
Japan's Pharmaceuticals and Medical Devices Agency has published a set of principles for using artificial intelligence in regulatory judgements and decisions about medicines, medical devices and related products. The central rule is human-centred: AI may support a judgement, but it must not replace the person or organisation accountable for that judgement. Patient and public interests come first, and ultimate responsibility remains human even when a model generates an assessment, prioritises a signal or recommends an action.
That position matters because the document reaches well beyond an office chatbot. It covers AI, including generative AI, used to inform regulatory work across a product's lifecycle: non-clinical and clinical development, manufacturing, applications and reviews, and post-market activity. The intended audience includes PMDA itself, applicants and marketing-authorisation holders, manufacturers, sponsors, contract research organisations and the developers or providers of AI systems. A model used to help interpret trial evidence, assess manufacturing information or find a safety signal could therefore fall inside the framework when its output informs a regulatory judgement.
The qualification is as important as the scope. PMDA labels the publication an Early Consideration and says it is not intended to establish new regulatory requirements. It is a reference point that may evolve with technology, law and international discussion. It does not approve any product, endorse a model or report that AI has improved regulatory accuracy. The Japanese document is authoritative; PMDA describes the English version as a provisional translation and says the original prevails if wording differs.[1][2][3]
The framework has three layers rather than a single compliance checklist
PMDA organises the principles in three layers. The first is the human-centred foundation: AI should augment human work while preserving accountability and protecting patients and the public. The second contains core principles for deciding whether and how AI should be used. These include a risk-based approach; a clear purpose, context and boundary; suitable data governance; fairness, accountability and transparency; multidisciplinary expertise; and compliance with relevant law, guidance and standards.
The third layer concerns implementation and management. Organisations are expected to address design and development, evaluate a system before use in the environment where it will operate, monitor it through its lifecycle, review it periodically, preserve human oversight and intervention, and develop AI literacy among the people who use or govern it. This structure treats safe use as a continuing process rather than a one-off model score. A system can change in practice because its input data, software, users or operating environment change even when the model name remains the same.
It is not a universal checklist with identical evidence for every use. The document repeatedly points to intended use and risk. An internal drafting aid whose output is always checked should not be treated in the same way as a system whose ranking could influence whether a safety concern is escalated. The practical burden should rise with the possibility and severity of harm, the degree of automation, the difficulty of detecting an error and the consequences of relying on the output.[1][2]
The distinction that matters
Three layers of PMDA's AI principles
The framework moves from a human-centred foundation to decision principles and then to controls that must continue through use.
Human-centred foundation
Use AI to support, not replace, judgement
Patients and the public come first; people retain responsibility
Core decision principles
Define purpose, risk, data, fairness and transparency
Evidence and controls should match intended use and possible harm
Lifecycle management
Validate, monitor, review and allow intervention
Changes and failures remain visible after deployment
Traceable data are part of the decision record
The principles make data governance central because a regulatory conclusion can only be understood if the evidence entering the system can be traced. PMDA calls for organisations to document data characteristics, sources, processing and change history, along with the basis for reasoning and decisions. That creates a chain from the original material through transformation and model use to the judgement a person finally makes. A polished output without that chain may be difficult to audit, reproduce or challenge.
For pharmaceutical and device work, the risk is not limited to obviously incorrect text. Training or evaluation data can omit important populations, reflect a different clinical or manufacturing context, or encode historical practices that are inappropriate for a new product. Data can also drift after deployment. A post-market system trained on one reporting pattern may behave differently when coding, reporting behaviour or the mix of products changes. Traceability should make those differences visible rather than treating all records as interchangeable.
The document also connects data practice with fairness and transparency. Organisations should consider whether outputs differ across relevant groups or settings and should provide information about the purpose, actual use and limits of AI to users, patients, the public or healthcare professionals where appropriate. The level and form of disclosure will depend on context, but transparency cannot be reduced to naming a model. People affected by a decision need an understandable account of what role AI played, where its evidence came from and who can review or correct the result.[1][2]
Validation must resemble the real use environment
PMDA says performance and validation should be appropriate to the intended use and risk, and that evaluation should consider the actual use environment rather than model performance alone. This is a meaningful distinction. A benchmark can show how a model handled a fixed dataset, yet say little about how it behaves with incomplete submissions, unusual products, changing terminology, time pressure or interactions with a human reviewer. A single accuracy number can also hide costly errors if the rare missed case is the one that threatens patient safety.
A defensible evaluation should therefore begin with the decision the AI is meant to support. Developers and users need a comparator, a relevant dataset, prespecified outcomes and an error analysis tied to consequences. For a signal-detection tool, that might include missed signals, false alerts, time to escalation and analyst workload, not merely overall classification accuracy. For a generative assistant used in a review, evaluation should test factual fidelity, source traceability, omissions and the reviewer's ability to identify a plausible but unsupported statement.
The document provides no sample, dataset, measured effect or preferred threshold. It is policy guidance, not a methods study. It also does not establish that human review automatically makes a system safe. People may over-trust fluent output, face alert fatigue or lack the time and expertise to reconstruct a model's reasoning. Validation must test the combined human-and-system workflow, including whether oversight catches consequential failures and whether the system adds work as well as speed.[1][2]
Monitoring and intervention are required after deployment
The lifecycle approach continues after a system begins operating. PMDA calls for ongoing monitoring and periodic review, with arrangements for human oversight and intervention. That implies defined triggers for investigation, correction, restriction or withdrawal. It also requires version control: when a model, prompt, knowledge base, data pipeline or surrounding process changes, the organisation needs to know whether prior validation still applies.
Human intervention must be practical rather than ceremonial. A reviewer needs authority, information and time to challenge the output; the organisation needs a route to stop or roll back unsafe use; and responsibility cannot be transferred to a vendor's terms of service. Contracts should support incident reporting, change notification, audit access, security and continuity. If a provider updates a model without enough notice or cannot explain a material shift, the regulated organisation still owns the consequences of using it in a decision.
AI literacy is therefore a governance control. Regulatory scientists, clinicians, statisticians, manufacturing experts, quality staff, data specialists, lawyers and patient-facing professionals may all need different knowledge. Not everyone must become a model engineer, but people must understand the intended boundary, common failure modes, data restrictions, escalation route and evidence required before relying on an output. Multidisciplinary review can expose a technical success that fails a clinical, legal or operational test.[1][2]
What changes for patients, reviewers and product teams now
For applicants and manufacturers, the immediate task is to map where AI already influences regulated decisions. That inventory should distinguish low-consequence administrative help from uses that shape evidence selection, quality assessment, benefit-risk reasoning, manufacturing control or safety surveillance. Each material use needs a named owner, intended purpose, data lineage, validation record, monitoring plan and route for a person to intervene. Calling a tool 'assistive' does not remove the need to examine how much its output affects the final judgement.
For PMDA and other reviewers, the principles create a public benchmark for internal use as well as submissions from industry. Consistency matters: a regulator asking companies for traceability and human accountability must be able to demonstrate comparable controls in its own AI-supported processes. The document does not say that every model detail will be public, and legitimate confidentiality or security constraints remain. But people affected by regulation should be able to understand the role and limits of AI well enough to contest a material error.
For patients and healthcare professionals, no approval or label changes because this paper was published. It does not cover ordinary clinical practice as its primary focus, although medical software or evidence used in regulatory work can overlap with clinical settings. Its practical promise is procedural: decisions supported by AI should remain attributable, reviewable and connected to traceable evidence. Whether that promise improves safety or timeliness will depend on implementation and published outcomes, not the existence of principles alone.[1][2][3]
What would strengthen—or weaken—the assessment
The next useful evidence would show how the principles are applied to real regulatory tasks. PMDA could publish case studies or evaluation protocols describing the intended use, comparator, dataset, human role, error categories, subgroup performance and decision consequences without exposing confidential product information. Consistent reporting of incidents, overrides, performance drift and withdrawn uses would help distinguish effective oversight from paperwork. External or cross-agency evaluation would add independence to evidence produced by the organisations operating the systems.
The assessment would become more favourable if risk-based validation catches consequential failures, traceability allows reviewers to reproduce decisions, human intervention works under operational pressure and monitoring detects changes before patients are harmed. It would become less favourable if 'human in the loop' is used as a label without time or authority to challenge outputs, if vendors change systems without revalidation, if important groups are absent from data, or if efficiency claims are reported without missed cases and correction workload.
International alignment will also matter because drug and device evidence crosses borders. The principles resemble a broader move toward lifecycle governance, risk proportionality, data quality and human accountability, but this report does not claim that another regulator has adopted PMDA's wording or that one jurisdiction's validation will satisfy another. For now, the document is a clear Japanese reference point: broad enough to cover generative AI and the full product lifecycle, cautious about replacing judgement, and explicit that its effectiveness remains to be demonstrated.[1][2][3]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring patient safety: compare the finding and its evidence limits.
Is operating-room AI ready for clinical use?
Not on the published evidence yet. A peer-reviewed scoping review screened 3,020 records but found only one completed feasibility study with five analysed patients; four larger prospective studies had no results posted.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients should not assume that an AI-supported regulatory decision is automatic: PMDA says people retain responsibility and affected users should receive appropriate information about purpose, use and limits.
- Regulatory and product teams need enough time, authority and multidisciplinary skill to challenge an output, document the evidence trail and stop unsafe use; nominal human review is not sufficient.
- Healthcare professionals see no immediate change to an approval or label. The publication changes expectations for governance, not the legal status of a medicine or device.
Global context
Medicines and devices move through international development, manufacturing and surveillance systems, so firms may face several regulators' expectations for the same AI-supported process. PMDA's emphasis on lifecycle management, risk proportionality, data governance and human accountability fits a wider regulatory direction, but the exact legal force, submission requirements and acceptable evidence differ by jurisdiction. This Japanese guidance should not be presented as an international rule or mutual recognition decision.
What the evidence does not yet show
- The publication is a non-binding Early Consideration and explicitly does not establish new regulatory requirements.
- It contains no study sample, dataset, comparator, measured performance, patient outcome or implementation evaluation.
- PMDA is describing principles for its own regulatory environment; the source is not independent evidence that those principles work.
- The English text is a provisional translation, and the Japanese original prevails if wording differs.
- Basic research and ordinary medical practice are not the document's primary focus, although regulated products and regulatory evidence can overlap with clinical use.
- The framework leaves thresholds and evidence requirements to intended use and risk, so consistent implementation will need more detailed examples and review practice.
What to watch next
- PMDA case studies or detailed expectations for particular regulatory tasks
- Published validation protocols with relevant comparators, error categories and subgroup analysis
- Evidence on human overrides, performance drift, incidents and uses withdrawn after monitoring
- Rules for vendor model changes, audit access, security, continuity and revalidation
- Whether patients and professionals receive understandable disclosure when AI materially informs a decision
- International convergence or divergence in evidence accepted for the same AI-supported process
Living evidence record
Impact record IAI-14QN8Q1
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
3 November 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 3 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
How Japan's PMDA expects organisations to govern AI used in regulatory judgement and decision-making across the pharmaceutical and medical-device lifecycle
Topics: Medicines regulation · Medical devices · AI governance · Patient safety · Data provenance · Human oversight
Regions: Japan · Asia
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.