In December, Washington asked two opposite questions about medical AI

Two federal requests for comment were open in December, and they point in different directions.

The first asks how anyone would know whether an AI-enabled medical device is still working once it is installed in a hospital. The second asks what is slowing AI adoption down and what should be removed to speed it up.

Both are legitimate questions. Read together, they describe a government trying to do two things at once with a workforce that, by its own auditor's account, is not sized for either.

Question one: how do you measure a device after it ships

The FDA opened docket FDA-2025-N-4203, a request for public comment on measuring and evaluating artificial intelligence-enabled medical device performance in the real world [s1].

This is the genuinely hard problem in the field. A conventional device performs the same way in year three as in year one. A model does not necessarily. Its inputs shift — a new scanner, a new patient mix, a changed referral pattern — and its performance can degrade without anything visibly breaking.

The American Hospital Association filed its comments on 1 December [s1]. Its recommendations are specific and, notably, mostly about instrumentation rather than restriction.

The AHA asked the FDA to pursue risk-based post-deployment measurement and evaluation standards, with adverse-event reporting mechanisms that capture AI-specific factors — algorithmic stability, and detection of drift between training data and the real-world population a model is serving [s1]. Existing adverse-event reporting tools, it argued, have no fields for any of that [s1].

It also asked that post-market evaluation be synchronised with existing clearance pathways rather than layered on top of them. More than 96% of AI devices reach market through the 510(k) route [s1]; the AHA's suggestion is that manufacturers submit a post-market monitoring plan inside a single application instead of returning repeatedly [s1].

And it raised a distributional point that rarely surfaces in AI policy: rural and safety-net hospitals do not have the resources to run AI governance programmes, and will need training, technical assistance and grant funding if post-deployment monitoring becomes an expectation [s1]. Otherwise monitoring quality tracks institutional wealth, and the hospitals least able to detect a failing model are the ones most likely to be running one unexamined.

Question two: what is slowing adoption down

On 23 December, the HHS Office of the Deputy Secretary, working with ASTP/ONC, published a request for information on accelerating the adoption and use of AI as part of clinical care [s2]. Comments close on 23 February 2026 [s2].

It follows the HHS Artificial Intelligence Strategy issued on 4 December, and is framed against OMB Memorandum M-25-21 and America's AI Action Plan [s2].

The RFI organises its inquiry under three headings — Regulation, Reimbursement, and Research & Development — and poses ten numbered questions [s2]. Under Regulation, it asks how current rules affect adoption while protecting patients and maintaining public trust, and describes the goal as a regulatory posture on AI that is well understood, predictable and proportionate [s2]. Under Reimbursement, it argues that legacy fee-for-service payment creates barriers to adopting innovation, and invites proposals for modernised payment structures [s2].

Several of the ten questions are ones a sceptic would also want answered: real-world AI performance outcomes, evaluation methods and how HHS could support them, patient and caregiver perspectives on benefits and concerns, and research priorities including literature on costs, benefits and clinical impacts [s2].

But the document's stated purpose is acceleration, and that framing shapes what comes back. STAT reported on the same day that the administration is shifting agencies from setting protective guardrails toward promoting rapid deployment, with policy experts arguing that combining rapid uptake with reduced regulation requires the public to accept the risks of unchecked experimentation [s5].

The staffing problem underneath both

Whoever answers either question has to be employed to do it.

On 16 December, STAT reported the departure of two senior FDA officials in the relevant space: Jessica Paulsen, acting deputy director of the Digital Health Center of Excellence and a 15-year agency veteran who had led the centre since summer 2024, and David McMullen, director of the Office of Neurological and Physical Medicine Devices, who left to join Neuralink [s4].

The centre's leadership has turned over repeatedly. Paulsen had replaced Sonja Fulmer, who left for Mayo Clinic; Fulmer had succeeded Troy Tazbaz, who returned to Oracle [s4].

In the same reporting, lawmakers directed the FDA to report within 90 days on its authorities for regulating AI medical devices, and within 180 days on its engagement with AI in drug development [s4]. The Government Accountability Office found that insufficient staff limit the FDA's ability to conduct oversight activities [s4].

That is the sequence: more work assigned, capacity flagged as inadequate, senior expertise leaving.

Meanwhile, in practice

None of this is prospective for clinicians. AMA data show 66% of physicians reported using some AI tool in practice in 2024, against 38% in 2023 [s3]. Sixty-eight per cent saw at least some advantage to AI in their practice, up from 63% [s3]. The AMA's chief executive, John Whyte, put it directly in December: AI is not the future of the practice of medicine, it is happening now [s3].

Adoption is not waiting for the answer to either question.

What to watch

Three things. Whether the FDA's real-world performance work produces anything with teeth — specifically, whether adverse-event reporting gains AI-specific fields, which is the single concrete change the AHA asked for [s1]. Who is appointed to lead the Digital Health Center of Excellence, and whether the post is filled permanently. And what the HHS RFI's comment file looks like when it closes in February [s2] — in particular whether patient and clinician organisations file, or whether the record is written mostly by vendors.