FDA tightens AI governance

FDA tightens AI governance The U.S. Food and Drug Administration is drawing a sharper line around artificial intelligence in drug regulation. If a company uses an artificial intelligence model to generate information that could influence a decision about a drug’s safety, effectiveness, or quality, the agency now wants that model treated as part of the regulated evidence chain, not as a back-office convenience. (fda.gov) That shift comes from a draft guidance the agency issued on January 6, 2025, titled *Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products*. The notice was published in the *Federal Register* on January 7, 2025, and the public comment window ran through April 7, 2025. As of April 8, 2026, the document remains listed by the Food and Drug Administration as a draft, Level 1 guidance with non-binding recommendations. (fda.gov) The document is notable because it is the Food and Drug Administration’s first guidance focused on artificial intelligence used to support regulatory decision-making for drug and biological product development. The agency says the use of artificial intelligence in drug development and in regulatory submissions has increased exponentially since 2016, which helps explain why it is now formalizing expectations. (fda.gov) The core idea is “model credibility.” In plain terms, the agency is asking a simple question: when an artificial intelligence system helps answer a regulatory question, why should reviewers trust that answer for that specific job? The guidance says credibility depends on the model’s “context of use,” meaning the exact question the model is being used to answer and how its output will be used in practice. (fda.gov) That matters because artificial intelligence can be used in very different ways across the drug product life cycle. The Food and Drug Administration points to examples such as predicting patient outcomes, identifying predictors of disease progression, and processing large datasets including real-world data and data from digital health technologies. A model used to summarize routine information for internal efficiency is one thing; a model used to support a conclusion in a submission to regulators is something else entirely. (fda.gov) The guidance therefore introduces a risk-based credibility assessment framework. Sponsors are expected to assess how much risk attaches to the model’s role in a decision and then match that risk with appropriate credibility activities to show that the model’s output is reliable for the stated context of use. In other words, the more an artificial intelligence output could sway a safety, effectiveness, or quality judgment, the more thoroughly a sponsor should be prepared to explain and validate it. (fda.gov) This is where governance gets more concrete. Companies will need to do more than say an artificial intelligence tool “worked.” They will need to describe what the model is supposed to do, what data it was built or tested on, what its limitations are, and what human review stands between the model’s output and a regulatory action. Those expectations are reflected both in the agency’s framing of credibility and in industry analyses of the draft guidance’s practical demands. (fda.gov) For drugmakers, that has immediate consequences in safety operations. Artificial intelligence is already used in areas such as signal detection, case triage, and large-scale review of post-market information. If those outputs feed decisions about whether a safety issue is real, urgent, or reportable, the model is no longer just a productivity layer. It becomes part of the logic that regulators may inspect. That is an inference from the guidance’s scope, but it is a strong one because the document applies whenever artificial intelligence produces information intended to support regulatory decision-making on safety, effectiveness, or quality. (fda.gov) The same logic reaches into manufacturing. Artificial intelligence models are increasingly used to monitor processes, flag anomalies, and analyze quality trends. When those outputs inform decisions about product quality, batch disposition, or process understanding, sponsors should expect the Food and Drug Administration to ask how the model was defined, tested, and overseen. The guidance does not ban those uses; it makes them explainable obligations. (fda.gov) One practical implication is that companies should inventory informal artificial intelligence use before it shows up in a submission or an inspection. In many organizations, teams adopt models locally for literature review, adverse event sorting, document drafting, coding support, or trend analysis without a central record. Under the Food and Drug Administration’s framework, the key question is not whether a tool feels low-profile internally. The key question is whether its output influences a regulated judgment. If it does, governance has to catch up. That conclusion is an inference from the agency’s stated scope and emphasis on context of use and credibility. (fda.gov) Another important feature is who issued the document. The draft guidance was produced through a cross-agency effort involving the Center for Drug Evaluation and Research, the Center for Biologics Evaluation and Research, the Center for Devices and Radiological Health, the Center for Veterinary Medicine, the Oncology Center of Excellence, the Office of Combination Products, and the Office of Inspections and Investigations. That breadth suggests the agency wants consistent expectations across review, product oversight, and inspection functions rather than isolated rules in one office. (fda.gov) The Food and Drug Administration has also continued building out its broader artificial intelligence posture since this draft appeared. In December 2025, the agency announced deployment of agentic artificial intelligence capabilities for agency employees, and in January 2026 it published *Guiding Principles of Good AI Practice in Drug Development*. Those moves do not replace the 2025 draft guidance, but they show the agency is trying to pair internal adoption of artificial intelligence with clearer external standards for companies. (fda.gov) For sponsors, the message is straightforward. If artificial intelligence helps produce evidence or analysis that could affect a regulatory conclusion, the Food and Drug Administration expects a sponsor to explain the model well enough that reviewers can understand why its output should be trusted in that exact use. In practice, that means artificial intelligence governance is moving closer to validation, documentation, and traceability disciplines that regulated companies already know from other critical systems. (fda.gov) The story is less about the agency suddenly embracing or rejecting artificial intelligence than about changing the burden of proof. A year ago, many companies could still treat artificial intelligence as a smart assistant wrapped around existing workflows. Under the Food and Drug Administration’s framework, once the model influences a decision on safety, effectiveness, or quality, it becomes something closer to an auditable instrument. And