On 2 November 2026, Japan’s Pharmaceuticals and Medical Devices Agency (PMDA) published new principles for the responsible use of artificial intelligence in regulatory activities. The document explicitly includes generative AI and covers regulatory judgment and decision-making throughout the pharmaceutical lifecycle, including manufacturing and post-marketing activities.
The publication, titled “Basic Principles for the Use of AI in Regulatory Activities (Early Consideration)”, was developed by PMDA’s Center for Regulatory Science.
Its intended audience includes regulators, pharmaceutical manufacturers, marketing authorisation holders, applicants, sponsors, contract research organisations and developers of AI systems used in relevant regulatory activities.
The scope covers AI applications that support regulatory judgments and decisions from product development through manufacturing and post-marketing activities.
Regulatory status: The document is an official PMDA Early Consideration publication. It explicitly states that it does not establish new regulatory requirements.
It should therefore be understood as a statement of regulatory principles, not as binding legislation or a new GMP guideline.
The English version is a provisional translation. In the event of inconsistency, the Japanese original prevails.
Three levels of AI governance
PMDA presents its principles in a three-level structure.
Level 1 - The human-centric principle: AI is intended to support rather than replace human judgment and decision-making. Patient and public interests take priority, and ultimate responsibility remains with humans.
Level 2 - Six core principles:
- Risk-based approach.
- Clear context of use and scope.
- Data governance.
- Fairness, accountability and transparency.
- Multidisciplinary expertise.
- Adherence to applicable laws, regulations and standards.
Level 3 - Five implementation and management principles:
- AI model or system design and development.
- AI model or system evaluation.
- Lifecycle management.
- Human oversight.
- AI literacy.
Together, these principles provide a structured approach for assessing whether AI systems are suitable for decisions that may affect regulated medical products.
Validation must reflect the actual operating environment
One particularly important element for pharmaceutical companies concerns model evaluation.
PMDA states that performance assessment and validation should be appropriate to the intended use, scope and estimated risks.
Crucially, these assessments should consider not only model performance but also how the AI system operates in its actual use environment.
This distinction matters for GMP applications.
An AI model may achieve impressive results during controlled testing but perform differently when exposed to real manufacturing data, unusual process conditions, incomplete records or changes in operating practices.
For example, companies evaluating AI-assisted deviation analysis, manufacturing trend interpretation or quality-control decision support should consider whether testing adequately represents the intended operational context.
These are potential GMP applications of the principles, rather than specific use cases mandated by PMDA.
Data governance and traceability
PMDA emphasises that AI systems depend on appropriate data governance.
The principles identify the need to maintain reliable data foundations, assess whether input data are suitable for the intended use and manage information about data sources and processing.
For AI development, PMDA highlights the documentation and management of:
- Data characteristics and sources.
- Data processing steps.
- Change histories.
- The basis for decisions.
- Traceability of relevant information.
For pharmaceutical quality professionals, this reinforces the importance of data integrity and controlled data pipelines when AI contributes to regulated decisions.
A model cannot provide trustworthy decision support if its input data, preparation methods or limitations are insufficiently understood.
Lifecycle management after deployment
PMDA’s principles extend beyond initial testing.
They call for ongoing monitoring and evaluation following deployment, together with periodic reviews to support model and system performance.
The publication also addresses human oversight and the ability to intervene where AI outputs are inappropriate.
These considerations are particularly relevant to AI systems whose behaviour may be affected by changing data, model updates or evolving operating conditions.
A practical pharmaceutical AI lifecycle strategy should therefore consider performance monitoring, periodic reassessment and controlled responses when performance deteriorates.
Although the PMDA document recommends these measures, it does not itself establish a new legally binding GMP monitoring requirement.
What pharmaceutical quality and compliance teams can do now
A useful response is to assess existing or planned AI applications against the PMDA principles.
The following questions provide a practical starting point:
- Is the intended context of use sufficiently specific, including the decisions influenced by AI?
- Have the risks to product quality, patient safety and data reliability been assessed?
- Are training, input and evaluation data appropriate and traceable?
- Has performance been assessed under conditions representative of actual use?
- Are model updates, monitoring and periodic reviews adequately governed?
- Are human responsibilities, review activities and intervention arrangements clearly defined?
- Do the personnel overseeing the system understand its capabilities and limitations?
These questions are consistent with a risk-based quality-management approach and can help organisations identify gaps before expanding AI use in regulated processes.
Significance for global GMP compliance
The PMDA publication is notable because it addresses AI-supported regulatory decision-making across the lifecycle rather than focusing exclusively on drug discovery or clinical research.
For multinational pharmaceutical companies, it provides another authoritative reference when developing common AI governance principles across regions.
There are clear conceptual connections with established expectations for risk management, data integrity, validation, human accountability and lifecycle control.
However, the publication must not be confused with a harmonised international GMP standard.
It does not replace applicable Japanese GMP requirements, FDA CGMP requirements or European GMP provisions. Nor does it establish that a particular generative AI application is automatically acceptable for a critical GMP decision.
AIforPharma assessment
The principal value of the PMDA publication is its combination of human accountability, context-specific validation and ongoing lifecycle management.
It reinforces an important direction in pharmaceutical AI governance: evaluating an AI system requires more than testing its technical accuracy.
Companies also need to understand the decisions it supports, the quality and provenance of its data, the behaviour of the system under operational conditions, and the mechanisms available for human control.
For GMP and pharmaceutical quality functions, these principles provide a useful framework for designing risk-proportionate controls before AI becomes more deeply embedded in manufacturing and quality-management processes.
Official sources and publication dates
- PMDA - Basic Principles for the Use of AI in Regulatory Activities (Early Consideration). Published 2 November 2026; provisional English translation:
https://www.pmda.go. … /files/000283394.pdf - PMDA - Official regulatory science project page linking to the new AI principles (publication listed as 2 November 2026):
https://www.pmda.go. … -sb-std/rs/0015.html - PMDA - Official What’s New listing confirming publication on 2 November 2026:
https://www.pmda.go.jp/english/0006.html
For other AI regulatory references relevant to GMP and pharmaceutical quality, visit AIforPharma’s collection of key AI documents:
https://www.aiforpha … ry/key-ai-documents/