|
Issue 23
|
|
The weekly brief The FDA just finished asking the public how to regulate AI doctors. The deadline passed almost unnoticed.This issue walks through what the FDA’s discussion paper on generative and agentic AI medical devices actually proposes, why a separate oversight loosening for diagnostic chatbots passed in January, and the new governance jobs that regulatory uncertainty is already creating inside hospitals. October 26, 2026 · 9 min read
|
Sourced from FDA, ARPA-H, AMA, JMIR, FDA Law Blog, MD Anderson, Duke Health. | No hype We read the FDA’s own discussion paper, not just the press summaries. | Why now Comments closed Oct. 19 — most clinicians never knew the window existed. |
|
If you read nothing else The FDA’s Digital Health Center of Excellence closed its comment period October 19, asking the public to help shape oversight for generative and agentic AI in medical devices — a two-axis risk framework and a physician-style premarket exam, not yet a rule. It follows a January policy that already loosened oversight for ‘glass box’ diagnostic AI, and a record 295 AI-enabled devices cleared in 2025 alone. The shape of ‘AI doctor’ regulation is being drafted in public right now, largely without the public watching. |
|
|
Nova’s signal Discussion papers don’t bind anyone, which is exactly why they matter — they’re where an agency tips its hand before the stakes get real. Miss this stage, and you’re reacting to a rule instead of shaping one. Nova’s call. If you work anywhere near health AI, read the discussion paper’s open questions — that’s next year’s rule, in draft form. Last issue tracked why enterprises keep pulling AI agents back out of production; this one is about the regulator trying to get ahead of the same reliability question before it reaches a hospital. Story 01 has the FDA’s actual framework, 02 the scale of AI medicine it’s trying to govern, 03 the policy that already loosened oversight in January, 04 the government’s own bet on an FDA-approved AI agent, and 05 what patients and doctors actually trust so far.
|
|
This week 5 developments in the quiet rulemaking fight over AI doctors.
|
|
01 / Health AI The FDA’s discussion paper floats a two-axis risk test and a ‘physician exam’ for AI.On August 18, 2026, the FDA’s Digital Health Center of Excellence published Considerations for the Regulation of Generative AI-Enabled Medical Devices, a discussion paper (not binding guidance) under docket FDA-2026-N-7874, asking 26 questions and taking comment through October 19. Its core proposal: calibrate oversight along two axes — how autonomously the system acts, from purely informational to fully autonomous action, and how severe the consequences are if its output is wrong — and evaluate premarket readiness the way a residency program evaluates a new physician, combining non-clinical benchmarking with clinical confirmation, per FDA Law Blog’s analysis. Why it matters. This is the actual blueprint for how ‘AI doctor’ software gets reviewed before it touches a patient — and the comment period, now closed, was the one formal chance to shape it before the FDA turns it into real guidance. The signal. The paper also raises foundation-model device master files and postmarket monitoring scaled to risk — meaning a model that gets updated after clearance, which describes most generative AI, may need its own ongoing oversight track rather than a one-time approval. FDA Law Blog: Regulating a Moving Target →
|
|
02 / Health AI The paper is trying to catch up to roughly 1,500 AI devices the FDA already cleared.The FDA’s own AI-enabled device list has grown to somewhere between 1,450 and 1,600 entries depending on the cutoff date, with a record stretch of new clearances in 2025 — one tracker counts 295 for the year, another puts it near 335 — and 2026 on pace to add hundreds more. Roughly three-quarters of all clearances are radiology tools, and 96% of the whole list went through 510(k), the pathway built for devices ‘substantially equivalent’ to something already on the market, per MedTech Dive’s tracker. Why it matters. A pathway designed for incremental hardware tweaks is now the default route for software that can update its own behavior after release — which is exactly the mismatch the new discussion paper is trying to patch, for generative and agentic systems specifically, after the fact. The signal. The FDA’s own list is explicitly not exhaustive and gets revised with older devices added retroactively — even the regulator’s count of what it has already approved isn’t fully settled. MedTech Dive: tracking new AI devices at the FDA →
|
|
03 / Health AI Nine months before this consultation, the FDA already loosened the rules for one slice of AI medicine.In January 2026 the FDA revised its clinical decision support guidance to apply enforcement discretion to software that gives a single recommendation a clinician can independently review — a ‘glass box’ standard that can cover some generative AI features, provided the tool doesn’t process medical images or signals directly, per KevinMD’s clinician-facing explainer. It partially reverses a 2022 guidance written specifically because regulators worried clinicians would defer to AI recommendations under time pressure. Why it matters. This is the closest thing to a precedent for how the FDA resolves a hard AI-oversight call, and it leans toward enforcement discretion over blanket regulation when a human stays in the review loop — a strong hint at where the generative-AI framework may land, too. The signal. The carve-out is narrower than most headlines suggested: the FDA says it still claims full authority over opaque models, time-critical tools, and anything that substitutes for a clinician’s own judgment rather than informing it. KevinMD: what clinicians need to know about the 2026 guidance →
|
|
04 / Agentic AI A federal agency is already racing to build the first FDA-authorized AI agent — on a 2-year clock.ARPA-H’s ADVOCATE program is funding up to $62.7 million over four years to build an agentic AI system for heart-failure and heart-attack patients that can autonomously adjust medications, appointments, diet, and exercise plans as a “24/7 digital member of the care team.” A team including Tempus AI, UpDoc, and Atman Health — backed by Microsoft, OpenAI, and Nvidia — is required to submit an actual FDA authorization package within two years, with Johns Hopkins’ Applied Physics Lab validating performance and safety independently, per ARPA-H’s own program page. Why it matters. This turns the discussion paper’s abstract risk-tiering into a live test case with a filing deadline — ARPA-H says it’s working directly with FDA leadership on the regulatory framework alongside the technology, so this program may end up writing the first real-world precedent for agentic AI in medicine before any formal rule exists. The signal. The program is also funding a separate supervisory AI “overseer” to monitor the clinical agent in real time — oversight-of-the-AI-by-AI is already part of the design, previewing a question any future FDA guidance on agentic systems will have to answer: who approves the overseer? Fierce Healthcare: ARPA-H’s $63M cardiovascular AI initiative →
|
|
05 / Health AI 81% of doctors now use AI at work. Most patients still don’t want it running their diagnosis.The AMA’s 2026 physician survey (1,692 respondents, fielded January–February) found AI use has more than doubled since 2023, to 81%, and most doctors say it’s improving their ability to care for patients — but 88% are also worried about skill erosion, especially among physicians with under 10 years of experience, per the AMA’s own release. On the patient side, a JMIR Formative Research survey of 524 US adults found 62.8% trust a physician’s own expertise most for a diagnosis, versus just 7.8% who’d trust an AI-first model, per ASCO Post’s coverage. Why it matters. Whatever risk framework the FDA settles on has to coexist with this gap: the clinicians using AI the most are also the most anxious about what it’s doing to their own skills, and the patients on the other end of the exam table still want a human making the call. The signal. The same JMIR study found that simply telling a patient their doctor used AI to help diagnose them lowered trust and willingness to seek care — regardless of whether the AI made the diagnosis more accurate, an awkward fact for any future rule that leans on disclosure as its main safeguard. AMA: more than 80% of physicians now use AI professionally →
|
|
Signal & chatter A few more things worth knowing.CDRH already has the next two documents on its list. The FDA’s FY2027 guidance agenda names two follow-ups in progress: evidentiary recommendations specifically for generative-AI mental-health chatbots, and a broader risk-assessment guidance for generative-AI devices — proof this discussion paper is step one of a longer pipeline, not a one-off. Explore → | One device maker didn’t wait for the comment period to end. Radiology AI company Mosaic Clinical Technologies reportedly filed a citizen petition the same day the discussion paper dropped, asking the FDA to clarify how AI vision-language tools get classified — a sign that even companies already in the market find the current rules too unclear to sit on. Explore → |
|
|
From the community Clinician commentary · KevinMD and physician trade coverage, not a vendor survey. The reaction to January’s CDS loosening split along predictable lines: analysts who called it a sensible fix for a 2022 guidance that was already pushing hesitant clinicians to ignore useful AI suggestions, versus critics who argue it dodged the actual hard question — what happens once a tool moves from informing a decision to effectively making it. Both camps agree on one thing: the FDA still hasn’t drawn a bright line between those two, and the discussion paper that just closed comment is the agency’s attempt to finally draw it. KevinMD: FDA loosens AI oversight — what clinicians need to know →
|
|
FOWL prediction 23 By June 30, 2027, the FDA will issue its first binding generative-AI-specific device guidance — and it will keep the discussion paper’s two-axis risk tiering rather than replace it with one flat rule.Every stakeholder response I could find — legal counsel, device makers, radiology’s own trade group — argued for risk-tiering over a blanket standard, and ARPA-H’s ADVOCATE teams are on a two-year clock to actually file for FDA authorization of an agentic system, giving the agency a concrete, funded test case pushing it toward a workable framework well before any legislative deadline would force one. The honest counter-case: a single high-profile AI-related patient harm before mid-2027 could push the FDA toward a stricter, less risk-tiered rule instead, the way a bad outcome has reset medical-device rules before. Review date: June 30, 2027.
|
|
Put it to work 3 things to do this week.| 01 | If you work in health, health tech, or hospital IT, skim the FDA’s discussion paper’s 26 questions directly — they’re the actual outline of next year’s rule, not just a comment-period formality. | | 02 | If a next career move is on your mind, look at the new ‘AI governance’ titles hospitals are hiring for right now — MD Anderson and Duke both have open roles, and several don’t require a clinical degree. | | 03 | If you or someone you know uses an AI chatbot for health questions, check whether it tells you upfront that it’s AI — per this issue’s JMIR study, that one disclosure changes how much people trust the answer, accurate or not. |
|
Career radar AI Governance Lead (Health Systems)This is the job regulatory uncertainty itself created: someone inside a hospital or health system who owns the model inventory, runs the AI governance committee, and signs off before any tool that touches a patient encounter goes live. MD Anderson Cancer Center has two open roles along these lines — Associate Director, Institutional AI Governance and a related Associate Director, AI Governance — and Duke Health’s AI Evaluation & Governance Program is hiring an Evaluation Lead for its clinical decision-support oversight initiative. A place to start. Read one real AI governance job description end to end before assuming it needs a clinical background — several of these roles are explicitly looking for policy, risk, or compliance experience, not an MD. Explore AI careers → |
|
|
The opportunity board Know where to look.General AI-training platforms (Outlier, Mercor, DataAnnotation) still have the volume, but health-specific evaluation work is where domain expertise actually pays a premium right now — clinicians and health-policy people rating AI outputs for accuracy and safety, not just fluency. Outlier: current AI evaluation projects → On the hiring side, hospitals and research institutes are staffing up for exactly the oversight this issue describes — A*STAR’s Principal Scientist posting in Singapore covers post-deployment surveillance and model-drift detection for clinical AI, the kind of work any FDA postmarket-monitoring rule will eventually require someone to do. A*STAR: Principal Scientist, Computing & Intelligence →
|
|
A discussion paper with no binding force can still be the most important document in a regulatory cycle — it’s the one moment an agency tells you what it’s thinking before the thinking turns into a rule you just have to follow. Whether anyone outside a law firm’s client-alert list actually read this one is a separate question from whether it mattered. Until next Monday,
TYB
Data scientist & founder, FOWL AI
|
|
One question for you Would you want your doctor’s AI tools regulated more like a drug, or more like a search engine?Hit reply. I read every response, and it helps shape what we cover next.
|
|
Online · Vibe Code Saturdays Read about the future.
Then build it.A free, hands-on Saturday to build a website, newsletter, portfolio, or useful tool with AI. Beginners welcome. Live on Zoom. Next date coming soon — apply and we’ll email your Zoom link and setup videos once it’s set.
|
Know someone who’d find this useful?Share this briefing → Forwarded this email? Get the free weekly briefing.
|
|
|
|