The US Food and Drug Administration has rejected a petition from Harrison.ai that would have created a partial exemption from 510(k) premarket review for several categories of AI-enabled diagnostic and detection software, including computer-aided detection used in radiology. The proposal was published in the Federal Register on 29 December 2025, with a public comment period that ran until 27 February 2026, before the agency closed the docket this spring.

The FDA's reasoning emphasised two points. First, premarket clearance of one device from a developer does not amount to evidence that subsequent products from the same developer will perform safely or effectively, particularly when the underlying model architectures or training data shift. Second, the agency expressed concern about leaning on manufacturers' own post-market monitoring as the primary safety mechanism when independent FDA oversight has been weakened upstream.

The decision sits awkwardly alongside the FDA's January 2026 guidance loosening oversight of certain general-wellness AI tools and clinical-decision-support software where clinicians can "independently review" the underlying recommendation. Together the two moves draw a line: low-risk consumer- and clinician-facing tools where humans remain in the loop are now further outside FDA reach, while AI tools that directly diagnose or that produce outputs clinicians cannot meaningfully audit remain firmly inside.

Neither decision addresses how adaptive AI and post-clearance model updates will be regulated.