Authorization is still a gate. Selective autonomy is still a bet.
Who decides what AI may do without a human in the loop, and what patients already believe before they walk in. Four signals from FDA and the journals.
This pack is about who gets to decide what AI is allowed to do without a human in the loop, and what patients already believe before they walk in.
1. FDA: radiology CAD and CADt still need 510(k)
FDA denied Harrison.ai’s petition to exempt certain radiology computer-aided detection and/or diagnosis devices from premarket notification. Even with a prior clearance and a post-market plan on the table, the agency kept authorization as the gate.
Marketing can sound finished. The regulatory file is not. If you sell or buy radiology CAD, treat 510(k) as the floor, not a paperwork nuisance you outgrow after one clearance.
Source: Federal Register notice
2. Nature Medicine on selective autonomy for on-premise clinical AI agents
A new Nature Medicine paper maps selective autonomy for clinical AI agents that run on-premise. The useful idea is not raw accuracy. It is decision-time reliability: which tasks the agent may handle alone, and which must defer to a clinician. The evidence base includes MIMIC benchmarks. That is not bedside proof.
Autonomy without stop rules is a demo, not a deployment. Boards should ask for the deferral map before they ask for the accuracy slide.
Source: Nature Medicine
3. Lancet Digital Health: patient-facing GenAI shapes understanding before the visit
Lancet Digital Health looks at how patient-facing generative AI can shape what people think they know before they see a clinician. Fluent explanations can act like informal second opinions. Most system-level CDS evaluation frameworks miss that interpretive influence. If you cite this piece in formal materials, disclose DeepMind co-authorship.
The visit does not start when the door opens. It starts when the patient finishes chatting with a model. Clinicians inherit that framing, for better or worse.
Source: Lancet Digital Health
4. Galleri: FDA advisers vote on September 23 (advice, not approval)
FDA’s Molecular and Clinical Genetics Panel met on September 23, 2026, on GRAIL’s Galleri PMA, after NHS-Galleri missed its primary late-stage endpoint.
An advisory vote is advice. It is not approval. Multi-cancer early detection stays a serious evidence question. Watch the panel’s questions, not the press-release draft.
Source: FDA advisory committee calendar
Authorization gates, selective autonomy, and patient-facing GenAI keep asking the same question: who owns the decision when the model sounds sure? Keep clinician judgment in the loop.
Ask DR GPT™: If a patient arrives with a GenAI-written “second opinion,” what is your clinic’s first move: reframe, document, or decline to engage the model’s framing at all?
Educational and futurist context only. Not clinical advice. Not a diagnostic service.
Originally published on my Substack. Planning a board session on clinical AI governance? Explore my healthcare AI keynotes.