Our signature is the safety control. Radiology just asked FDA to make someone prove it works.

Almost every argument for putting AI-drafted text into a radiology report ends the same way: a radiologist reviews it and signs it. We make that argument too. We read and sign reports every day. At ApertureAI we're building Portal, a workflow that places AI-drafted reports in PACS for a radiologist to review, edit and sign. We're still working out how to measure it, so weigh our bias.

Last week ACR and RSNA told FDA that the signature argument is not enough on its own. We think they're right, and we think the details of what they asked for are worth reading closely if you sign reports for a living.

What FDA asked

On Aug 18, 2026, FDA's Digital Health Center of Excellence released a discussion paper, "Considerations for the Regulation of Generative AI-Enabled Medical Devices," and opened docket FDA-2026-N-7874 for feedback through Oct 19 [1]. FDA is explicit that the paper is not draft or final guidance and is not meant to announce policy [1]. It is a set of questions.

Three of them matter most for report drafting.

First, risk. FDA floated a two-axis framework, with device activity on one axis (from non-directive information up to autonomous action) and the consequence of relying on an incorrect output on the other [1].

Second, premarket evidence. FDA described a "competency-based" approach: benchmark the device, then confirm performance clinically [1].

Third, postmarket monitoring. FDA said it is considering whether to accept "greater premarket uncertainty" for generative AI devices in exchange for more reliance on postmarket monitoring. It listed three possible approaches: periodic re-benchmarking, periodic sample-based review by independent clinicians, and monitoring for performance degradation or drift [1].

What ACR said

ACR filed its comments on Oct 19 and summarized them for members on Oct 22 [2][3]. The letter opens by naming our use case directly: "image-informed draft reporting is among the most active areas of development and early clinical evaluation" [2].

The point that stopped us is about human oversight. ACR asks FDA to "explicitly credit qualified end-user oversight, but only if demonstrated." Its reasoning: an output that a qualified user reviews and accepts before it affects care is lower risk than one that acts on its own, but that oversight "should earn credit only when sponsors demonstrate its effectiveness in the intended workflow (e.g., error detection and correction under realistic workload, time pressure, and interface conditions, with residual errors reported after review)" [2].

ACR uses the definition of a qualified end-user from the 2026 ACR–SIIM Practice Parameter for Imaging Artificial Intelligence: a licensed physician who could provide the same service without the AI, and who is trained to judge whether its output is valid [2].

Then comes a line every vendor of drafting tools, including us, should read twice. Personalizing a draft to match a radiologist's phrasing "can lower editing burden, but it makes the AI output read like the radiologists' own, which can lower scrutiny by the reporting physician." ACR asks FDA to test error detection in the personalized setting, not only the default one [2]. That's the honest tradeoff. The feature that makes a draft feel easy to sign is the same feature that makes it easy to sign without really reading.

ACR also asks FDA to "measure the failures that matter clinically," not "generic hallucination rates, nondeterminism, or textual similarity" [2]. For imaging it lists them [2]:

1. Laterality errors.
2. Localization errors: correct finding, wrong site or level.
3. Fabricated or incorrect comparison to a prior, including a prior that doesn't exist.
4. Measurement hallucination, or a correct measurement attached to the wrong lesion.
5. Omission of a clinically significant incidental finding.
6. Wrong study, protocol or technique attribution.
7. Report-image discordance.
8. Failure to communicate critical or actionable results.
9. Erroneous or missing follow-up recommendations, including ones that don't match published guidelines.
10. Anatomic mislabeling.

It adds that safety measures should have minimum thresholds that a strong score elsewhere can't offset [2]. A draft that reads beautifully and gets the side wrong has failed.

On monitoring, ACR says programs should capture every eligible case, including processing failures, missing outputs, abstentions and "raw outputs before clinician editing" [2]. That last item is the one we would underline. If you only keep the signed report, you can't tell which errors the AI made and the radiologist caught, which ones the radiologist introduced, and which ones the AI made and the radiologist missed. ACR notes that its Assess-AI registry already monitors classification tools for hemorrhage, PE and large-vessel occlusion, and that it intends to extend that architecture to generative outputs [2].

What RSNA said

RSNA also filed on Oct 19 [4]. It supports a risk-based, total product lifecycle approach. It argues that stochastic outputs and models that change after deployment "create risks that cannot be addressed through one-time premarket review alone" [4].

Two of its points speak to the signature question. On risk, RSNA says the framework should account for reviewability: "A radiology output that cannot be independently verified without specialist expertise may warrant greater oversight even if it is labeled non-directive" [4]. On accountability, it says manufacturer responsibility "should remain primary" and that "Clinicians should not become the default bearers of liability for errors arising from AI systems they cannot reasonably oversee or control" [4].

RSNA also asks for things that make oversight possible in practice. It wants model and version identifiers, timely notice before material updates, transparency about the prompts and configurations that govern device behavior, and point-of-care information on which images, series or text the output actually used [4]. For agentic systems, it wants reviewable audit logs of prompts, retrieved information, actions, overrides and human interventions [4].

Why one PE study belongs in this conversation

Two days after the comment period closed, npj Digital Medicine published a retrospective study from Emory [5]. The investigators evaluated two FDA-cleared PE algorithms from a single commercial platform across a 17-facility academic system. One triaged PE on 30,678 dedicated CTPAs. The other looked for incidental PE on 37,191 routine contrast-enhanced CTs [5].

The CTPA model had 86.8% sensitivity and 99.1% specificity. Sensitivity was 99.3% for saddle emboli and 72.9% for subsegmental PE, and 89.7% for acute versus 65.3% for non-acute PE. The incidental PE model had 73.5% sensitivity and 99.8% specificity. Both had lower sensitivity than their FDA clearance benchmarks and higher specificity [5].

This is not a generative AI study, and we don't want to stretch it. Labels came from radiology reports through a validated LLM pipeline (97% accuracy), and the paper we read is the accepted preview [5]. Secondary coverage notes that radiologists could see the AI output during their reads, so the labels weren't fully independent, and that the deployed software version may not match the cleared one. We haven't confirmed those limits in the full text. The authors' own conclusion is the useful part: they call for "standardized post-market surveillance of AI-enabled medical devices" [5].

If a narrow classifier with a well-defined target behaves differently in the field than in its clearance file, a system that writes free text will too. And free text is harder to score.

Where we land

We'll say plainly what we think, and where we're unsure.

We think ACR's framing is correct. Radiologist review is a real safeguard, and it should count. It should count based on evidence, not because we assume it works. The evidence that matters is what reaches the patient chart after review, under real worklists, at 3 PM on a Tuesday with 40 studies waiting.

For a drafting tool, that means the record has to keep three things apart: what the model drafted, what the radiologist changed, and what errors survived the signature anyway. Without the first two, you can't measure the third. That's the bar we're holding Portal to, and it's harder than it sounds. We wrote about why we chose audit-traceable design in July [6].

We don't know if or when FDA will turn this discussion paper into guidance. Nothing in the paper is binding yet [1]. We also don't know how much residual error after review is acceptable for each failure type. We doubt anyone does yet. That's the work ACR offered to help define, and we'd rather radiologists define it than have it defined for us.

What we do know is that "a radiologist signed it" is a claim about our judgment, and it goes out under our names. We'd like the numbers behind it.

If your department runs AI-drafted text today, do you keep the raw draft next to the signed report? If not, how would you know what your review is catching?

References

1. U.S. Food and Drug Administration, Digital Health Center of Excellence. Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. Aug 18, 2026. Docket FDA-2026-N-7874; feedback due Oct 19, 2026. https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request (PDF: https://www.fda.gov/media/194242/download)
2. American College of Radiology. Comments on Docket FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices. Oct 19, 2026. https://edge.sitecorecloud.io/americancoldf5f-acrorgf92a-productioncb02-3650/media/ACR/Files/Advocacy/ACR-Comments-FDA-GenAI-Device-Oversight.pdf
3. American College of Radiology. ACR Weighs In on FDA GenAI Device Oversight. Oct 22, 2026. https://www.acr.org/News-and-Publications/2026/acr-comments-genai-device-oversight
4. Radiological Society of North America. Comments on Docket FDA-2026-N-7874. Oct 19, 2026. https://www.rsna.org/-/media/Files/RSNA/government-relations/RSNACommentsFDAGenAI101926.pdf
5. Mansuri A, Chavoshi M, Dapamede T, et al. Multi-site real-world performance of commercial AI for pulmonary and incidental pulmonary embolism detection. npj Digit Med. Published Oct 21, 2026. doi:10.1038/s41746-026-03418-3
6. Ruffolo A. Why we chose audit-traceable over foundation model. Radiologist.com. https://radiologist.com/insights/audit-traceable-vs-foundation-model.html