FDA cleared 1,450 AI devices

The FDA’s own list now puts the U.S. market for AI-enabled medical devices above 1,450 products authorized through the end of 2025. That number matters less as a bragging point than as a clue to where healthcare AI actually is: not in pilot mode, but already embedded in clinical workflows, imaging systems, software tools and decision-support products. (fda.gov) (theregreview.org) Michael W. Craige, writing in *The Regulatory Review* on August 6, said the mismatch is no longer between innovation and regulation in the abstract, but between deployment at scale and uneven governance around how these tools are described, tested and monitored. (nature.com) A separate 2026 developer survey points to the same gap from the other side: many people building medical AI tools do not know the relevant regulatory frameworks well, even though they say they should be held responsible for what they build. (fda.gov) ### Why does the 1,450-device figure matter beyond the headline? (fda.gov) The FDA says its AI-enabled medical device list is meant to identify products authorized for marketing in the United States and to give providers, patients and developers more transparency about which devices use AI. The agency also says the list is not comprehensive, because it is built largely from public summaries and classification information, and some authorized products may appear only in later updates. (theregreview.org) Congressional Research Service analysts said in a June 10, 2026 brief that most of those authorized products were cleared through existing device pathways and are concentrated in radiology, cardiology and neurology. (nature.com) That means the regulatory question is not whether AI has arrived in medicine. It is where oversight applies, where it does not, and how much outsiders can tell about the systems already in use. (fda.gov) ### What exactly is the governance gap people are talking about? Craige wrote that by the end of 2025 the FDA had authorized more than 1,450 AI-enabled devices, including 295 in 2025 alone, while fewer than 2% were supported by randomized clinical trials. (fda.gov) He also cited estimates that large language models used for clinical decision support can produce hallucinations at rates of 8% to 20%. (congress.gov) January 2026 added another layer. Craige wrote that the Department of Health and Human Services proposed an HTI-5 rule that would scale back disclosure requirements for AI used in certified health IT “decision support interventions.” He said the earlier framework required 31 source attributes on how tools were built, validated and maintained, along with intervention risk management obligations. (fda.gov) (theregreview.org) ### Why does developer unfamiliarity with regulation matter? A 2026 study in *npj Digital Medicine* found that developers of medical AI tools often lacked familiarity with the regulatory frameworks governing their work. (theregreview.org) At the same time, the study found most developers believed they should be accountable for the systems they create. That combination suggests a practical problem for hospitals and buyers. (theregreview.org) If developers accept responsibility but do not know the governing rules well, health systems may need to ask harder questions about intended use, evidence, validation, monitoring and whether a tool is regulated as a medical device at all. That is an inference from the survey findings and the FDA framework, not a direct quote from either source. ### Where does FDA authority start and stop? The Congressional Research Service said the FDA regulates only a subset of AI-enabled healthcare applications — specifically software functions that meet the legal definition of a medical device. (nature.com) The 21st Century Cures Act excluded some administrative software, electronic records functions and certain clinical decision support functions from that device definition. The FDA says its list covers AI-enabled devices that met applicable premarket requirements, but it also notes that public decision summaries are not all-inclusive. (congress.gov) So a product can be influential in care delivery without giving outsiders a full picture of training data, testing conditions or maintenance plans. That reading follows from the FDA’s description of the list and Craige’s account of the disclosure rollback debate. ### What should readers watch next? The FDA said it plans periodic updates to the AI-enabled device list and is exploring ways to tag devices that use foundation models, including large language models and multimodal systems. (congress.gov) The HHS HTI-5 proposal, cited by Craige, is the nearer policy milestone because it would determine whether federal transparency requirements for AI inside certified electronic health records are narrowed or preserved. (fda.gov) (theregreview.org)