FDA rejects AI deregulation

The Food and Drug Administration just said no to a push that would have let some new radiology artificial intelligence tools reach hospitals without the usual premarket review. The rejected idea came from Harrison.ai, which asked for partial exemptions from the 510(k) clearance process for certain imaging tools. (beckershospitalreview.com) That matters because 510(k) is the main gate many medical devices use before sale in the United States. A company usually has to show its product is substantially equivalent to one already on the market, instead of just promising it will watch for problems later. (fda.gov) The petition was not about all artificial intelligence in medicine. It focused on radiology software that spots lesions, analyzes scans, or flags urgent cases for faster review, which means software that can change who gets called back first and what a doctor looks at next. (federalregister.gov) Harrison.ai’s pitch was that companies with at least one cleared product should be able to launch some follow-on tools under extra conditions like post-market monitoring and transparency rules, instead of filing a new 510(k) each time. The Food and Drug Administration opened that petition for comment on December 29, 2025. (regulations.gov) Radiologists pushed back before the decision landed. The American College of Radiology told the agency on February 26, 2026 that any shortcut should be narrowed, tied to stronger transparency, and backed by ongoing registries that track real-world performance. (acr.org) The Food and Drug Administration’s answer fits the path it has been building for years. Instead of saying artificial intelligence should get a lighter touch, the agency has been writing guidance that asks companies to explain how the software works, how it was tested, and how future updates will be controlled. (fda.gov) That is what a predetermined change control plan is for. It is basically a pre-approved renovation permit: a company tells the Food and Drug Administration in advance what kinds of software changes it wants to make later, how it will validate them, and what guardrails will keep the device safe and effective. (fda.gov) The agency finalized that change-control guidance in August 2025, which gave developers one route to update artificial intelligence tools without reopening the whole case every time. The rejected petition asked for something broader: fewer front-end reviews for whole categories of new products. (fda.gov) The Food and Drug Administration is also moving in the opposite direction on transparency. Its artificial intelligence device list says the agency wants better ways to identify products that use foundation models, including large language models and multimodal systems, so hospitals and patients can see when those features are present. (fda.gov) So the message from this decision is not “no artificial intelligence.” The message is that if software helps detect cancer, sort urgent scans, or shape a clinical workflow, the Food and Drug Administration still expects paperwork, testing, and a clear audit trail before it treats that tool like just another app. (fda.gov) That logic does not stop at radiology. Drug safety teams are already using artificial intelligence to tag case reports, rank incoming alerts, and route possible side effects to human reviewers, and this decision signals that “the model seems useful” will not be enough in regulated work without documented controls and traceable performance. (fda.gov)