PMDA issues AI principles for regulatory judgment and decision-making

The Center for Regulatory Science at Japan's Pharmaceuticals and Medical Devices Agency (PMDA) published "Basic Principles for the Use of AI in Regulatory Activities (Early Consideration)" on November 2, 2026, with an English version the same day. It came as an administrative notice, or jimu renraku, the informal channel PMDA uses to write to trade associations, and it states that the document "is not intended to establish new regulatory requirements," offering instead principles that regulatory stakeholders are encouraged to take into account in common.

The scope is regulatory judgment and decision-making, not every task in a plant

The document primarily applies to settings in which AI, generative AI included, is used in activities involving regulatory judgment and decision-making, across the life cycle of pharmaceuticals, quasi-drugs, cosmetics, medical devices and regenerative medical products, from the development stages, including nonclinical and clinical studies, through manufacturing and post-marketing activities. Manufacturing enters through the regulatory judgments made there, not as a claim over every operation on site. Use in basic research or in general medical practice is not intended to be the primary focus. Development of the AI used for in-scope activities may also fall within the document, including ongoing updates.

The intended audience is PMDA itself, applicants, marketing authorization holders, manufacturers, sponsors and contract research organizations, plus developers and providers of AI models and of systems incorporating them.

An Early Consideration states the thinking at one point in time

Early Consideration is PMDA's label for reference information issued while scientific knowledge and information have not necessarily accumulated sufficiently, to support the practical application of new technologies. The Japanese notice says it presents the thinking at that point in time and may change as new findings emerge and science advances. The document itself will be reviewed and updated as necessary in light of technological developments, practical experience and international harmonization.

Anyone working from the English file should read its first-page footnote. It is marked "Provisional Translation (as of November 2026)" and is provided for reference purposes only; in the event of any inconsistency, the Japanese original prevails.

Three tiers, from the human-centric principle to implementation

The principles sit in three tiers. At the top is the Human-Centric Principle: AI supports rather than replaces human judgment, and ultimate responsibility rests with humans. Below it, six Core Principles cover a risk-based approach, a clear context of use and scope, data governance, fairness, accountability and transparency, multidisciplinary expertise, and adherence to laws, regulations and standards. The third tier, Implementation and Management Principles, covers design and development, evaluation, life cycle management, human oversight and AI literacy.

Data governance splits along the line between using AI and building it. When using AI, the appropriateness of data to be input or processed should be confirmed in light of the intended use. For data used for AI development, characteristics, sources, processing steps, change histories and the basis for decision-making should be documented and managed, with traceability maintained; the Japanese text frames that last point as an effort to be made rather than a flat obligation.

Evaluation is not a one-off either. Performance assessment and validation appropriate to the intended use, scope and estimated risks should be conducted prior to use, and should consider not only AI performance but also its operation in the actual use environment. Monitoring and evaluation continue once AI is in use, with periodic reviews. Outputs should not be relied upon excessively, and operational arrangements should enable intervention. PMDA adds that addressing only the matters in the document is not sufficient: organizational governance and operational frameworks apply whether or not they appear there.

What to settle before an AI output reaches a regulatory decision

Start by listing where AI already touches a regulatory judgment in your own work, then write down the context of use and scope for each, which principle 2.2 asks for. Drafting a submission section or a study record with generative AI falls inside the primary scope once that output feeds a regulatory decision. With vendors, put the disclosure principle in writing: intended use, actual use and limitations, in a form the user can act on, plus notice when a model is updated.

The document cites only two reference works: the Cabinet Office's Social Principles of Human-Centric AI (March 2019) and the Guidelines on the Use of Medical Digital Data for AI Research and Development (March 2024). Computerized system validation and data integrity frameworks appear nowhere in it, so lining these expectations up with an existing validation program is each organization's own call, as is how much of it to raise with PMDA at a GMP inspection or a consultation.