EMA’s Final Reflection Paper Sets a Risk-Based Approach to AI Across the Medicine Lifecycle

The reflection paper sets out EMA’s current thinking on the use of artificial intelligence and machine learning across the lifecycle of human and veterinary medicines. EMA published a draft for consultation in July 2023, received a large volume of comments (trade press reported 1,342), and published the final version in September 2024 after adoption by the CHMP and CVMP. The document reference is EMA/CHMP/CVMP/83833/2023.

The paper addresses AI use in drug discovery, non-clinical development, clinical trials, product information, manufacturing and post-authorisation activities such as pharmacovigilance. Law firm commentary notes that the final version does not introduce significant shifts compared with the draft, but adds more precise definitions, an expanded glossary and more detailed technical explanations, and adapts wording to align better with the EU AI Act.

Responsibility rests with the regulated parties. Summaries identify the marketing authorisation applicant, the marketing authorisation holder, the clinical trial sponsor and the manufacturer as the parties responsible for the AI tools they use and for the data and content those tools produce.

At its core the paper recommends a risk-based approach to the development, deployment and performance monitoring of AI and machine learning tools. The degree of risk depends on the AI technology itself, the context of use, the degree of influence the AI has on decisions, and the stage of the lifecycle. To avoid confusion with the definition of high-risk AI in the EU AI Act, the paper uses the categories of high patient risk and high regulatory impact.

Applicants and marketing authorisation holders who foresee that AI or machine learning could affect a product’s benefit-risk balance are encouraged to interact early with regulators. The paper accepts that some model architectures, including so-called black box models, cannot always be fully explained, and summaries note that such models may be acceptable in some circumstances where more transparent models perform less well, provided their use is justified and safeguards are in place.

For medical writers and regulatory teams, the section on product information is directly relevant. It covers AI and machine learning applications used for drafting, compiling, editing, translating, tailoring or reviewing information to be included in a medicine’s product information, such as the Summary of Product Characteristics. Because responsibility for submitted content rests with the applicant or marketing authorisation holder, text produced with such tools should be reviewed by qualified staff before submission.

On 5 September 2024 EMA and the Heads of Medicines Agencies published guiding principles on the use of large language models by staff across the European medicines regulatory network. They cover handling data safely, critically checking outputs and knowing whom to contact about concerns, and ask agencies to define governance, permitted use cases, training and risk monitoring. EMA describes them as a living document under its multi-annual AI workplan to 2028.

The large language model principles also stress that compliance with data protection rules must be ensured at all stages of the lifecycle of these tools, and recommend that staff report incidents or severely biased or erroneous outputs to the appropriate team. Although written for regulators, they indicate the standards that EU agencies will apply when they use such tools to handle submissions.

In January 2026 EMA and the US Food and Drug Administration published ten joint guiding principles of good AI practice in drug development. They are voluntary and are presented as a basis for future guidance.

EMA’s Methodology Working Party workplan for 2026 to 2028 lists guidance on AI in clinical development and on AI in pharmacovigilance, building on the reflection paper. Summary notes of an HMA and EMA industry stakeholder meeting in October 2026 state that a concept paper on AI in clinical development was being finalised, with public consultation planned. Pharmacovigilance guidance is to be developed as a question-and-answer document. At the time of checking, EMA had not published a revision of the reflection paper itself.

Sponsors and marketing authorisation holders should inventory where AI tools are used in activities that generate evidence or content for regulatory submissions, classify each use by patient risk and regulatory impact, and document validation and performance monitoring in proportion to that risk. Where AI could influence benefit-risk conclusions, the paper recommends early interaction with regulators.

Medical writing and regulatory operations teams using AI to draft or translate product information or other submission documents should keep human review in place, maintain records of which tools were used and how outputs were checked, and ensure that data protection rules are respected. This brief summarises public EMA documents for information and is not affiliated with EMA.