Healthcare · · 7 min read
What FDA competency-based evaluation means for generative AI devices
Acumen Editorial · Updated November 21, 2026
On August 18, 2026, FDA's Center for Devices and Radiological Health published a discussion paper on generative AI-enabled medical devices. It is not guidance and sets no policy. The comment period closed on October 19 under docket FDA-2026-N-7874.
It still matters. The paper is the clearest signal yet of the evidence FDA expects to see for devices built on large language models. And it moves away from the input-output testing that most device teams know.
On November 10, the American Medical Association added its own marker. Its House of Delegates adopted policy that AI cannot be licensed to practise medicine, and that diagnostic and prescribing tools need "transparent, auditable evidence" of safety and efficacy.
If you lead clinical AI, regulatory strategy or model evaluation, here is what the proposal asks for and how we would prepare.
What is FDA proposing for generative AI devices?
Three things, per FDA's announcement and law firm summaries from Cooley and Haynes Boone.
A two-axis risk framework. One axis is how much the device does on its own. The other is how severe the harm would be if a user relied on a wrong output. Risk rises toward the corner where both are high. Patient-facing information may count as higher risk than the same content delivered to clinicians.
A competency-based premarket model. CDRH says traditional input-output testing may be impractical, because the range of inputs and outputs is too large. Instead it proposes something loosely based on how physicians are trained and assessed. Cooley calls it a "credentialing-style" assessment and a significant departure from 510(k) substantial equivalence.
Risk-proportionate postmarket monitoring. Re-benchmarking on a schedule and after triggering events such as model changes, sample-based review by independent clinicians, and drift monitoring.
The paper also floats a voluntary master file route for foundation model developers and asks about agentic functions.
What would device benchmarking test?
The paper names three competency categories, plus additional ones for agentic functions. Haynes Boone summarises them as follows.
| Competency | What FDA describes | What a test set needs |
|---|---|---|
| Safety | Escalating safety-critical states in time, recognising requests outside intended use, conveying confidence and limits | Cases with planted red flags, out-of-scope requests, and rubric items for escalation and uncertainty |
| Clinical proficiency | Accurate knowledge, sound reasoning through ambiguous information, correct reading of numerical data, clear communication | Clinician-written cases with reference answers and graded reasoning, by specialty |
| Generalizability | Consistent safety and proficiency across foreseeable variations in input, user and runtime | The same cases rephrased, reordered, shortened, or posed by patients vs. clinicians, plus repeat runs |
| Agentic functions | Additional competencies, not yet specified | Multi-step tasks with checkpoints, tool use and stopping rules |
Benchmarking runs on the device in its deployed or representative configuration, against thresholds set during premarket review. Those same thresholds then govern postmarket re-benchmarking. That makes the first benchmark your long-term contract with the agency.
Two gaps stand out. The summaries we reviewed say nothing about test-set independence, contamination or sequestration. And no threshold values are given. Expect both to be argued in pre-submission meetings.
What counts as clinical confirmation?
Benchmarking is nonclinical. Clinical confirmation shows the device performs as intended with real use, and its rigour scales with risk. Haynes Boone lists four approaches from the paper:
Retrospective evaluation on prior patient inputs, compared with a reference standard or expert adjudication, including independent multi-clinician consensus.
Standardised patient interactions with trained actors, scored against prospectively defined criteria.
Blinded clinician adjudication of real cases, where a clinician decides the right response without seeing the device output.
Clinical trials, where the risk warrants it.
Three of the four depend on clinicians producing reference judgments. That is the operational centre of the proposal.
Who grades the outputs, and can AI do it?
The paper leaves open how clinician adjudicators would be credentialed and governed, Cooley notes. It also asks whether supervisory AI agents could take on some monitoring tasks.
The research says automated grading is possible but not off the shelf. A preprint by Buckley, Manrai and colleagues, "Scaling Clinical Judgment to Evaluate Medical AI," built GRAND-ROUNDS, a benchmark of 9,217 physician-scored responses from 160 clinicians across seven studies. The authors report that frontier models used as judges often disagreed with physicians and with each other. A model fine-tuned on a small number of physician examples scored consistently at physician level, and reproduced headline findings from five published studies without new human grading.
The lesson for a sponsor: an automated grader is only as credible as the physician labels it was calibrated on. You will need those labels anyway. For more on this, see LLM-as-a-judge for medical AI.
How should teams prepare now?
- Place each function on the two axes. Note where multi-turn use can drift from informational into action-directing.
- Map existing evidence to the three competencies. Most teams have proficiency data and little on safety escalation or generalizability.
- Build a private, held-back benchmark from clinician-written cases with reference answers and rubrics. Keep it out of training and vendor tuning.
- Cover your specialties and users. Recruit graders from the specialties in your intended use, and include patient-facing phrasing if patients will use it.
- Write perturbation sets for generalizability: rewordings, missing data, conflicting values, piecemeal case presentation.
- Measure grader agreement before you trust any score, human or automated.
- Design blinded adjudication for clinical confirmation: clinicians answer without seeing the output, then compare.
- Prespecify thresholds and re-test triggers, including silent foundation-model updates from your supplier.
- Document provenance: who wrote and graded each item, their credentials, conflicts and dates.
Report results by competency, not as a single number. The InclusionMed preprint, posted November 16, found that a model's rank changed with the role or capability measured, and that its top ten models all scored higher on patient communication than on clinical decision-making. One headline score hides that.
Does this apply if we are not a device maker?
Often, yes. Health systems buying clinical AI will ask vendors for evidence in the shape regulators favour. The AMA now wants physician-led oversight of AI in credentialing, privileging and peer review. Its delegates also directed advocacy for vendor liability when a tool used as intended causes harm. Both push buyers to demand competency evidence whether or not a product is cleared.
How Acumen helps
We build the expert-graded test sets this proposal assumes. Our medical model evaluation service supplies private benchmarks written and graded by practising clinicians, organised by safety, clinical proficiency and generalizability, with documented grader credentials and agreement.
We also run blinded adjudication of real or retrospective cases and re-benchmark on every model change.
FAQ
What is FDA's competency-based evaluation for generative AI devices?
It is a proposal in CDRH's August 18, 2026 discussion paper. Instead of exhaustive input-output testing, devices would be benchmarked on safety, clinical proficiency and generalizability, then shown to work through clinical confirmation scaled to risk. It is loosely modelled on how physicians are trained and assessed. The paper is not guidance and sets no binding policy.
Is the FDA discussion paper binding?
No. It is a discussion paper with a request for feedback, and FDA states it proposes no new policy yet. Comments closed on October 19, 2026 under docket FDA-2026-N-7874. It is still the best current signal of the evidence FDA expects for LLM-based devices, so sponsors are using it to shape pre-submission plans.
What are the three competency categories?
Safety, clinical proficiency and generalizability, with additional competencies proposed for agentic functions. Safety covers escalation, recognising out-of-scope requests and conveying uncertainty. Proficiency covers knowledge, reasoning, numerical interpretation and communication. Generalizability asks whether performance holds across foreseeable variation in inputs, users and runtime conditions.
What does clinical confirmation require?
It depends on risk. The paper describes retrospective evaluation against expert adjudication, standardised patient interactions with trained actors, blinded clinician adjudication of real cases, and clinical trials for some devices. A prospective study may not be needed in every case. Three of these four approaches rely on clinicians producing independent reference judgments.
Can an LLM grade our device instead of clinicians?
Not without calibration. A 2026 preprint found frontier models used as judges often disagreed with physicians and with each other. A model fine-tuned on a small set of physician examples reached consistent, physician-level scoring. Either way, credible automated grading starts from physician-labelled data, and FDA has only asked whether AI agents could help with monitoring.
What happens after a device is on the market?
The paper proposes re-benchmarking against the premarket thresholds on a schedule and after triggering events such as model changes, sample-based review by qualified independent clinicians, and monitoring for drift. It also asks about shared roles for health systems, payors and professional societies. How adjudicators would be credentialed is left open.