FDA issues AI guidance

The Food and Drug Administration has started drawing a brighter line around how drugmakers may use artificial intelligence in submissions that ask the agency to trust a result. In a draft guidance released on January 6 and published in the Federal Register on January 7, 2025, the FDA laid out its first formal recommendations for using AI to generate information meant to support decisions about a drug’s safety, effectiveness, or quality. The document is narrow on purpose: it is not about AI as a general productivity tool, but about AI that helps produce evidence the FDA might rely on when judging a medicine. (fda.gov) (federalregister.gov) That changes the conversation from “Are companies using AI?” to “Can the FDA see exactly what the model did, where it could fail, and why its output should be trusted in this use?” The agency’s answer is a risk-based credibility framework. A company first has to define the model’s specific job, which the FDA calls its “context of use,” then show that the model is credible for that job with testing, documentation, and controls scaled to the risk of getting the answer wrong. (fda.gov 1) (fda.gov 2) The examples make the point concrete. An AI model might be used to predict patient outcomes, sort through real-world data, flag patterns in disease progression, or help assess manufacturing quality. Those are very different tasks. A model that helps explore a dataset early in development does not carry the same regulatory weight as one that supplies evidence in a submission tied to whether a product should be approved, how it should be labeled, or whether a production process is reliable. The FDA says the amount of validation should rise with that weight. (fda.gov 1) (fda.gov 2) This is not the agency’s first encounter with the technology. The FDA said in its January 2025 announcement that use of AI in drug development and regulatory submissions has grown exponentially since 2016, and trade coverage at the time described the guidance as a response to a rising flow of AI-assisted filings. The draft guidance also urged sponsors to talk with the agency early, before a model becomes baked into a pivotal study or a manufacturing claim that will be hard to unwind later. (fda.gov) (pharmexec.com) The timing matters because the FDA is building this framework while trying to widen its reach elsewhere. On April 7, 2026, Politico reported that the administration’s fiscal 2027 budget request asked Congress for new authority to treat a drug ad as misbranded if it lacks fair balance and creates a misleading impression about what the FDA actually approved. The same request sought more flexibility over how advisory committees are composed and when they meet. Those proposals are separate from the AI guidance, but they point in the same direction: an agency asking for more room to decide what counts as reliable evidence and what counts as a misleading claim. (politico.com) (hhs.gov) The FDA is also not acting alone. In January 2026, the FDA and the European Medicines Agency published ten shared principles for “good AI practice” across the medicines lifecycle, from research and trials to manufacturing and safety monitoring. The agencies described those principles as a base for future guidance, which makes the January 2025 FDA draft look less like a one-off memo and more like the first brick in a larger regulatory architecture. (ema.europa.eu) (fda.gov) For industry, the practical shift is plain. If a company wants to lean on AI in a filing, it will need more than a strong model and a polished slide deck. It will need a paper trail that shows what data went in, what the model was meant to do, how it was tested, where it breaks, who is accountable for it, and why the FDA should trust it in that exact setting. The agency has moved the burden from AI’s promise to AI’s paperwork, one submission at a time. (fda.gov) (federalregister.gov)