EMA and FDA set 10 principles for good AI in the medicines lifecycle
EMA and the U.S. FDA have agreed on ten principles for good AI practice across the medicines lifecycle. The guidance spans early research, clinical trials, manufacturing, and safety monitoring - a full-stack view of how AI should be built, validated, and maintained in regulated pharma.
For engineering and data teams, this is a clear signal: AI used in drug development will be held to disciplined, auditable standards. The principles will inform future guidance in each jurisdiction and push stronger cross-border collaboration among regulators, standards bodies, and industry.
Why this matters to engineering and data teams
AI use in medicines has grown fast, and EU pharmaceutical legislation now makes room for AI in regulatory decision-making and controlled testing environments. EU guideline development is underway, building on EMA's 2024 AI reflection paper, with continued EU-US collaboration following the April 2024 FDA-EU bilateral meeting.
Ethics and safety remain non-negotiable. Expect more structure around data quality, validation, documentation, and monitoring - not as red tape, but as the cost of deploying AI where patient outcomes are on the line.
"The guiding principles of good AI practice in drug development are a first step of a renewed EU-US cooperation in the field of novel medical technologies. The principles are a good showcase of how we can work together on the two sides of the Atlantic to preserve our leading role in the global innovation race, while ensuring the highest level of patient safety." - European Commissioner for Health and Animal Welfare, Olivér Várhelyi
What "good AI practice" likely looks like in delivery
The announcement doesn't publish the list, but the direction is familiar to anyone shipping ML in regulated settings. Expect emphasis on the following themes:
- Clear problem framing and clinical relevance tied to a defined context of use.
- Data governance: lineage, quality controls, representativeness, and bias checks.
- Model transparency appropriate to risk, plus human-in-the-loop oversight.
- Validation beyond internal test sets: multi-site, temporal, and population shifts.
- Security and privacy by design, including threat modeling for AI-specific risks.
- Lifecycle monitoring: drift detection, triggers for retraining, and rollback plans.
- Traceable documentation: decisions, datasets, model versions, and audit trails.
- Quality systems integration (GxP), change control, and incident management.
- Third-party tool/vendor diligence and reproducibility of builds.
- Clinical safety risk management linked to performance claims and limits.
Practical checklist for your ML pipeline
- Define the context of use and risk level before model design; set acceptance criteria up front.
- Lock down data contracts, lineage tracking, and data quality gates in your ETL.
- Standardize experiment tracking, model registries, and immutable artifacts.
- Run pre-specified validation across sites, time windows, and subgroups; document failures.
- Build explainability where it helps clinical decision-making; avoid superficial dashboards.
- Establish monitoring with thresholds, alerting, and controlled rollback paths.
- Map your pipeline to GxP and CSV expectations; treat models as change-controlled configuration.
- Run security reviews for training and inference (supply chain, prompts, data exfiltration, poisoning).
- Codify human oversight: who reviews, what gets escalated, and how it's recorded.
- Keep a single source of truth for documentation: model cards, test reports, risk logs, and SOPs.
Policy signals to watch
- EU guidance updates that build on the 2024 EMA AI reflection paper.
- FDA-EU coordination on pilots and regulatory sandboxes for AI methods.
- Greater alignment with international technical standards to simplify cross-border submissions.
Bottom line
Regulators are setting expectations that mirror strong engineering practice: clarity, validation, and accountability. Teams that bake these into their MLOps now will ship safer systems, move faster in review, and avoid expensive rework later.
Resources
Level up your team
If you're building regulated ML systems and need structured upskilling across data, validation, and MLOps, explore our role-based learning paths.
Your membership also unlocks:
