On 18 August 2026 the U.S. Food and Drug Administration published a discussion paper on how it might regulate medical devices built on generative AI, and opened a public docket (FDA-2026-N-7874) for comments until 19 October. The paper comes from the Digital Health Center of Excellence in the agency’s Center for Devices and Radiological Health (CDRH).
It is not guidance and imposes no new requirements, the FDA’s announcement says. But it is the clearest signal yet of how the agency is thinking about systems whose outputs are open-ended text, which do not fit neatly into the testing regime built for traditional software.
/ 01What the paper proposes
The FDA describes four areas in the paper, each with targeted questions for industry, clinicians, researchers and the public:
- A two-axis framework for risk. Risk assessment for generative AI devices would be organised along two dimensions rather than the single risk class most device sponsors are used to.
- Competency-based premarket evaluation. The agency floats an approach loosely modelled on how physicians are trained and evaluated, in two parts: non-clinical benchmarking of the device, and clinical confirmation.
- Risk-proportionate postmarket monitoring. Several options for how closely a device should be watched once it is in use.
- Foundation models and agentic AI. Considerations specific to devices built on third-party foundation models and to systems that act with some autonomy.
The agency says these devices offer significant potential benefits but may carry risks different from those of traditional software and earlier AI-enabled devices. CDRH Director Michelle Tarver said, “Patients and clinicians deserve a regulatory approach that keeps pace with the rapid innovation.” The ASCO Post notes the paper is intended to start a discussion toward a framework, not to set one.
/ 02Why a competency model is a meaningful shift
Traditional device software is validated by testing defined inputs against expected outputs. A generative system can receive almost any input and produce almost any output, so exhaustive testing is not possible. Treating the device more like a clinician, by benchmarking its knowledge and then confirming its performance in a clinical setting, is a pragmatic answer to that problem.
It also has an implication that sponsors should notice. A clinician is not credentialed once and left alone. They are supervised, audited and re-examined. If the FDA follows the analogy through, generative devices will face expectations of ongoing evidence, not a single clearance. The paper’s emphasis on postmarket monitoring points in that direction.
The inclusion of agentic AI is the other notable element. A device that drafts a note for a clinician to review is one thing. A device that takes multi-step actions, calls other tools, or escalates care on its own is another. The FDA is asking now, before such products arrive in volume, what additional evaluation they should need.
/ 03Evidence chains make competency testable
Benchmarking a generative model on clinical questions measures whether it tends to be right. It does not tell a reviewer why a particular answer was given, or whether it rests on current guidance. For postmarket monitoring, that second question matters more. When a device produces a wrong recommendation, the manufacturer needs to know whether the cause was a bad source, an outdated guideline, a retrieval error or the model itself.
Systems that ground their answers in a structured representation of clinical knowledge, with every recommendation linked to the guideline section, trial or formulary entry it came from, make that diagnosis possible. A knowledge graph of clinical concepts and their relationships can be versioned, so a reviewer can see exactly what the system knew on the day it made a recommendation. That is the kind of evidence a competency-and-monitoring regime will reward.
/ 04What to watch and what to do
- The comment docket. Comments are due 19 October. Submissions from large device makers and health systems will show where industry wants the lines drawn.
- Foundation model responsibility. How the FDA divides accountability between a device sponsor and the third-party model it builds on.
- Agentic devices. Whether autonomy becomes its own risk factor with its own evidence requirements.
- Draft guidance. The discussion paper is a precursor. The draft guidance that follows will be the document that binds.
For developers building clinical AI, the practical step is to design for traceability now: keep answers tied to the sources that support them, and keep those sources versioned. Whatever the final rules say, that record will be easier to show a regulator than a model’s confidence score.
A competency exam tells the FDA that a generative device tends to be right. Postmarket monitoring will need to know why it was wrong on a particular day, and that requires the device to record what it knew at the time.
Represent the clinical knowledge a device draws on as a versioned graph: guideline recommendations, the trials and formulary entries behind them, and the patient criteria that trigger each one. Every output links to the recommendation nodes it used and the guideline version in force. When a device suggests the wrong anticoagulant dose, a reviewer can ask a precise question: did it follow a superseded recommendation, match the wrong patient criterion, or ignore a correct path? That separates source errors from reasoning errors, the distinction a risk-proportionate monitoring plan depends on. Our cardiovascular build works this way, with 145 concepts and 284 recommended actions each tied to the page it came from, and on-prem deployment keeps that evidence trail inside the PHI boundary.
How we’d model it
Device output—citesRecommendation→Guideline recommendationGuideline recommendation—supportedBy→Clinical trialGuideline recommendation—validDuring→Guideline versionAdverse event report—tracesTo→Device output
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices — U.S. Food and Drug Administration fda.gov ↗
- FDA Seeks Public Input Relating to Regulatory Considerations for Generative AI–Enabled Medical Devices — The ASCO Post ascopost.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.