Hospitals want clearer rules for generative AI in medicine, and they want the companies selling the tools to carry the monitoring burden. On 19 October 2026, the American Hospital Association (AHA) submitted feedback to the Food and Drug Administration (FDA) on regulatory considerations for generative AI-enabled medical devices, responding to the agency's discussion paper and request for feedback. Writing to Dr Michelle Tarver, director of the FDA's Center for Devices and Radiological Health, the association, which represents nearly 5,000 member hospitals, health systems and other healthcare organisations, said it supports a flexible, risk-based framework that protects patients without impeding responsible innovation. Becker's Hospital Review, reporting the letter on 20 October, framed the core message the same way: hospitals often lack the technical visibility and resources to evaluate AI tools on their own, so the FDA should make developers primarily responsible for monitoring safety and performance.
Who Should Own Monitoring
The letter's sharpest point is about what happens after a tool is cleared. The AHA says hospitals and health systems have broadly built multidisciplinary governance teams to review AI use cases, technologies and value over time, and it cites its own Health IT supplemental survey finding that 74% of hospitals have multiple teams responsible for evaluating predictive AI, including senior leaders, department leaders and IT staff. But it argues that the black-box nature of many AI systems makes it harder for hospitals to identify flaws that could affect the accuracy and validity of an AI tool's analyses and recommendations. Its answer is that post-market measurement and evaluation standards should be developed for vendors, covering performance metrics, thresholds that trigger further evaluation, and communication requirements to end users about ongoing performance. The association also notes that over 96% of AI-enabled medical devices are approved through the 510(k) clearance process, which caps the number of indications per application, and it asks the FDA to explore clearance pathways where vendors provide detailed post-market evaluation plans, including additional special controls, to streamline clearances under fewer applications.
Fixing the Risk Heuristic
The AHA broadly supports the FDA's risk-based direction but wants its measurement tool improved. The FDA's discussion paper proposes a heuristic with two variables: the activity level of the device, from providing informational but non-directive outputs to fully autonomous action, and the consequences of its actions, from limited to severe. The AHA recommends adding variables the current version misses, including the likelihood or probability of consequences, reversibility, downstream safeguards and output traceability. It illustrates the gap with triage tools: citing a recent Nature study, it notes that certain commercial AI tools under-triaged true emergencies at a rate of 51.6% while over-triaging 64.8% of non-urgent cases, and it argues that a strongly worded, patient-facing recommendation about seeking emergency care can therefore carry consequences as serious as a supervised action-taking tool. Activity level alone, it says, does not predict risk in a straight line, so likelihood of error and negative outcome should weigh more heavily than whether an activity is informational or action-directing.
Labelling and Scope
Two further asks in the letter concern what hospitals and consumers can actually see on the label, and which tools fall inside FDA oversight at all. The AHA encourages the FDA to update labelling guidance, in collaboration with the Federal Trade Commission, describing how to distinguish informational from action-taking devices, with suggested disclaimer language for informational products and clarity about how much a clinician is in the decision loop and what kind of clinician that is. On scope, it asks the FDA to clarify that generative AI clinical decision support and administrative tools are not subject to FDA oversight, pointing to the 21st Century Cures Act exclusion for tools that support clinical decision-making where the clinician independently reviews recommendations. It also raises the position of smaller providers directly: some hospitals, particularly rural, critical access and other safety-net hospitals, may lack the staff or resources for governance structures and ongoing measurement, and many have not deployed AI tools at all, a gap the AHA says could widen a digital divide unless cross-agency training, technical assistance and grant funding help. Finally, it supports transparency through AI model cards, noting that hospitals have found the Office of the National Coordinator for Health IT's model-card transparency criterion helpful for procurement, and that it opposed the proposed removal of those criteria in the HTI-5 proposed rule.
What This Means for Hospitals Buying AI
For hospitals, health systems and clinics evaluating AI vendors, this letter is best read as a procurement checklist rather than a rule change: nothing in it is FDA policy yet, since it is feedback on a discussion paper. The practical questions it raises are the ones buyers should already be asking vendors: what performance metrics will you report after deployment, at what threshold will you re-evaluate, and how will you tell us when performance changes? Organisations with limited IT and governance capacity should treat the AHA's rural and safety-net warning as a reason to favour vendors that supply model-card-style transparency and documented monitoring plans, rather than tools whose inner workings cannot be examined. And because the letter asks for clinical decision support and administrative tools to stay outside FDA oversight, practices should not assume that every AI product they buy carries the same regulatory status. The intended use, the labelling and the degree of clinician review are what determine the category, so vendor claims need checking against the label, not the marketing page (see https://medclinicx.com/solutions/ai-patient-portal).
Analysis: Transparency Is the Procurement Question
Analysis (MedClinicX): The AHA's argument is less about slowing AI down than about moving the cost of proving it works from buyers to sellers. Hospitals already staff governance teams. What they lack is access to model weights, training data and performance drift data, and no governance committee can audit what it cannot see. If the FDA adopts developer-owned post-market standards, the practical effect for buyers would be contractual: performance reporting, re-evaluation triggers and drift disclosure would become terms to negotiate, not favours to request. The triage example in the letter is the cautionary case, an informational recommendation can still send a patient to the wrong place, and it is a reminder that the label category matters less than the real-world consequence. Until the FDA acts, the safest posture for a healthcare organisation is to require the transparency the AHA is asking for in writing, before signing.
Sources and References
- American Hospital Association, AHA Provides Feedback to FDA on Regulatory Considerations for Generative AI-Enabled Medical Devices, 19 October 2026: https://www.aha.org/2026-10-19-aha-provides-feedback-fda-regulatory-considerations-generative-ai-enabled-medical-devices
- Becker's Hospital Review, Hospitals can't see inside AI tools, AHA tells FDA, 20 October 2026: https://www.beckershospitalreview.com/healthcare-information-technology/ai/hospitals-cant-see-inside-ai-tools-aha-tells-fda/
- EIN Presswire, AHA details recommendations to FDA on regulatory frameworks for GenAI-enabled medical devices, 20 October 2026: https://business.einnews.com/pr_news/950915567/aha-details-recommendations-to-fda-on-regulatory-frameworks-for-genai-enabled-medical-devices
Reporting based on the sources above, accessed 25 October 2026. The AHA letter was read directly in full, and the Becker's report was read directly in full and cross-checked against it. MedClinicX did not conduct original reporting or interviews for this article. Corrections: contact us via https://medclinicx.com/contact