What sponsors need to know about FDA and EMA principles for AI in drug development

The FDA and EMA AI principles, published jointly on January 14, 2026, set out how the two largest drug regulators expect artificial intelligence (AI) to be used when it helps generate or analyze evidence for a medicine. The document contains 10 high-level principles that apply across the full development lifecycle, from laboratory studies and clinical trials to manufacturing and safety monitoring after approval. The principles are not legally binding, but both agencies have stated they will inform future guidance in their regions.

For sponsors, the practical question is what this means for the next filing: an Investigational New Drug (IND) application to begin human testing in the United States, a New Drug Application (NDA) or Biologics License Application (BLA) to market a product, or a marketing authorization application (MAA) in the European Union. When AI touches the evidence in any of these, regulators will expect a clear account of what the AI did, how much a decision relied on it, and how its reliability was confirmed.

What are the FDA and EMA AI principles?

The document, titled Guiding Principles of Good AI Practice in Drug Development, was developed by the U.S. Food and Drug Administration (FDA), with officials from its drug and biologics review centers, and the European Medicines Agency (EMA). It defines AI as system-level technology used to generate or analyze evidence across the drug product lifecycle.

The principles are short and deliberately general. In plain terms, they ask sponsors to ensure that AI:

• Serves people first, with foreseeable harms considered from the design stage.
• Receives validation and oversight in proportion to its risk.
• Follows the legal, ethical, scientific, cybersecurity, and quality standards that already apply, including GxP, the family of "good practice" quality rules such as Good Clinical Practice (GCP) and Good Manufacturing Practice.
• Has a clearly defined context of use, meaning a written statement of the specific job the AI performs and the boundaries of that job.
• Is developed and used by teams that combine technical AI expertise with clinical, regulatory, and quality expertise.
• Is supported by traceable data, with documented data provenance (where data came from and how it was processed) and protection of sensitive information.
• Is built using sound model design and software engineering, with attention to whether its outputs can be explained.
• Is evaluated as a complete system, including how people interact with its outputs.
• Is monitored across its lifecycle, including checks for model drift, the gradual loss of accuracy that can occur when real-world data changes.
• Is described in plain language, including its purpose, performance, and limitations.

None of these ideas is entirely new; what is new is a shared FDA and EMA vocabulary for AI in drug development. Sponsors already using AI in study design and patient recruitment can expect both agencies to evaluate that use through a similar lens, reducing the risk of building separate evidence packages for each region.

How the principles connect to existing FDA and EMA guidance

The joint principles sit on top of guidance each agency had already published. In the United States, the most detailed document is an FDA draft guidance issued in January 2025 on using AI to support regulatory decision-making for drugs and biological products. As of October 2026, that guidance remains a draft, which means it reflects FDA's current thinking but is not final.

The FDA draft introduces a seven-step credibility assessment framework. Credibility, in this context, means trust in an AI model's performance for a specific use, established through evidence. The steps run from defining the question and context of use, through assessing model risk and testing, to deciding whether the model is adequate for the job.

Model risk is the central idea. FDA describes it as a combination of two factors: model influence, which is how much the AI output contributes to a decision compared with other evidence, and decision consequence, which is how serious the harm would be if that decision were wrong. An AI model that alone decides which trial participants need overnight safety monitoring would carry high risk. The same model backed by an independent check would carry less, because its influence on the final decision is lower.

In Europe, EMA published a reflection paper on AI in the medicinal product lifecycle in September 2024. It uses two related ideas, high patient risk and high regulatory impact, and states that AI used in clinical trials should meet Good Clinical Practice requirements. Those requirements were updated in the international guideline ICH E6(R3), which took effect in the European Union in July 2025 and was published by FDA as final guidance in September 2025. Its data integrity and audit trail expectations apply to AI tools too, as explained in this overview of how ICH E6(R3) changes sponsor oversight.

Both agencies have also accepted AI-generated evidence when this discipline is followed. In March 2025, EMA issued its first qualification opinion for an AI-based tool that helps pathologists assess liver biopsy images in trials for metabolic dysfunction-associated steatohepatitis (MASH), a progressive form of fatty liver disease, and FDA qualified the same tool in December 2025. The tool's role was tightly defined, the model was locked, and a pathologist remained responsible for the final interpretation.

What regulators will expect to see in your next submission

The principles do not prescribe a submission template, but read alongside the FDA draft guidance and the EMA reflection paper, they point to consistent expectations. For each AI use that affects safety, effectiveness, or quality evidence, sponsors should be ready to provide or explain:

• An inventory of every AI tool used in the program, from protocol design and site selection through data analysis, safety monitoring, and manufacturing.
• A context of use statement describing the question each tool answers, how its output is used, and what other evidence sits alongside it.
• A documented risk assessment explaining how much the decision relies on the AI and what would happen if the AI were wrong.
• Credibility evidence proportionate to that risk, such as a description of the model, the data used to train and test it, and its measured performance.
• Data governance records that trace where data came from, how it was processed, and how patient privacy was protected.
• A description of human oversight, including who reviews AI outputs and how they can override them.
• A lifecycle plan covering performance monitoring, change control, and the triggers for revalidation.

Accountability for AI does not transfer to a vendor. When a sponsor uses a technology provider's model, or a contract research organization (CRO) deploys AI on the sponsor's behalf, the sponsor remains responsible for the evidence in the submission. The questions a CRO asks when it validates an AI recruitment tool offer a useful starting point for sponsor-level vendor review.

Early engagement is the other clear signal. FDA encourages sponsors to discuss AI plans in formal meetings, such as a pre-IND meeting, while EMA offers early discussion through its Innovation Task Force and scientific advice for higher-impact uses. For AI that could influence a pivotal efficacy or safety conclusion, raising it before the data are generated is generally easier than defending it after submission.

Where does AI-assisted patient recruitment fit?

Where recruitment AI sits on the risk spectrum depends on how it is used. The FDA draft guidance excludes AI used for operational efficiencies that do not affect patient safety, drug quality, or the reliability of study results. AI that helps identify and prioritize potential candidates, with qualified staff confirming eligibility before enrollment, generally fits a lower-influence profile. Neither agency has published guidance specific to recruitment AI, however, so sponsors should document their reasoning rather than assume the use is out of scope.

Risk rises when the AI's role expands. If an algorithm alone makes final enrollment decisions, assigns treatment, or influences dosing, its outputs begin to shape safety and efficacy evidence directly. EMA's reflection paper specifically identifies treatment assignment and dosing in early-phase trials as potential high patient risk uses.

Two failure modes deserve attention. The first is referral risk, where an AI tool sends unsuitable candidates forward, increasing screen failures and site workload. The second, and less visible, is auto-exclusion risk, where the tool silently filters out people who would have qualified. Auto-exclusion can narrow the enrolled population and weaken how well results represent the people who will eventually use the treatment.

Human review is the most practical control for both. DecenTrialz is a clinical trial recruitment and pre-screening platform that combines AI-assisted participant matching with registered nurse (RN)-led pre-screening and structured referrals to research site teams. In this model, AI narrows the field and a nurse reviews each potential participant's information before referral, which keeps the AI's influence on any final decision limited. A closer look at what nurse review adds to AI matching shows how the two layers work together.

Final eligibility determination, informed consent, the study walk-through, and enrollment remain with the authorized research site and study team.

Steps sponsors can take before the next submission

The principles carry no compliance deadline, so the more useful frame is readiness: what a sponsor would want in place if a reviewer asked about AI in the next filing. Practical steps include:

• Building a single AI register across the development program, including tools used by CROs and other service providers.
• Classifying each tool by influence and consequence, and focusing documentation effort on the highest-risk uses first.
• Writing context of use statements now, while the people who selected and configured each tool are available to explain it.
• Aligning AI oversight with the existing quality management system rather than creating a parallel process.
• Planning regulator engagement for any AI use that could affect a pivotal conclusion.
• Tracking guidance in both regions, since the FDA draft could be finalized and EMA has signaled further AI guidance for clinical development.

Ongoing monitoring of AI performance fits naturally within risk-based quality management, which already asks sponsors to identify what matters most to trial quality and focus oversight there.

Frequently asked questions about the FDA and EMA AI principles

Are the FDA and EMA AI principles legally binding?

No. They are non-binding recommendations that both agencies intend to inform future guidance, so they are best read as an early indication of what binding expectations will likely cover.

Is FDA's guidance on AI in drug development final?

As of October 2026, the January 2025 FDA guidance on AI to support regulatory decision-making for drugs and biological products remains a draft. Sponsors should confirm its status before a filing.

Does the EU AI Act apply to AI used in drug development?

The EU AI Act is a separate, cross-sector law. Most AI tools used solely in medicines research and development are not listed among its high-risk categories, although the full scope has not been settled in every case. AI that is itself a medical device or diagnostic may fall under high-risk rules, which amendments adopted in 2026 delayed to December 2027 and August 2028. Sponsors should confirm the position for specific tools with legal counsel.

Do sponsors need to disclose AI used for patient recruitment?

Neither agency has issued a recruitment-specific disclosure requirement. Sponsors should still document how the tool was used and how qualified staff reviewed its outputs, and raise it with regulators if it could meaningfully shape who enters the trial.

Preparing for AI questions in the next regulatory review

The FDA and EMA AI principles do not change what a sponsor must submit today. They change what reviewers in both regions are likely to ask. Sponsors that can show a defined role for each AI tool, validation scaled to risk, traceable data, and accountable human oversight will be better positioned as binding guidance follows.

For sponsors evaluating recruitment support built on AI-assisted matching followed by nurse-led pre-screening, the DecenTrialz team can walk through how the process works and where the research site's role begins.