What actually happened
On August 18, 2026, FDA's Digital Health Center of Excellence, part of the Center for Devices and Radiological Health, issued a discussion paper titled "Considerations for the Regulation of Generative AI-Enabled Medical Devices," opened under docket FDA-2026-N-7874. It is explicitly framed as a discussion paper, not draft guidance: it sets no current regulatory expectations, and its purpose is to surface stakeholder input on risk assessment, premarket evaluation, postmarket monitoring, and considerations specific to foundation models and agentic AI systems. The public comment window closed October 19, 2026.
As reported by Radiology Business, two of the field's largest professional societies — the Radiological Society of North America (RSNA) and the American College of Radiology (ACR) — filed comments by that deadline. Their input matters for a different reason than the paper itself: it's the clinical community that will be expected to work inside whatever framework eventually results, and their asks preview what "good" is likely to mean in practice long before any rule is final.
What RSNA told the FDA
RSNA's comments, submitted on the October 19 deadline, set the tone plainly: "patient safety should remain the central organizing principle, and regulatory flexibility should be paired with clear, enforceable safeguards that promote transparency, accountability, clinical validation, and continuous real-world monitoring," as quoted by Radiology Business.
Three specific points stand out in RSNA's letter:
Models change after they ship
RSNA argued that generative AI models have the potential to evolve after deployment, and that the resulting risks "cannot be addressed" through a one-time premarket review alone. It wants regulators to develop a plan for continuously monitoring deployed genAI products — covering model performance, human-AI interaction, workflow effects, and appropriate use.
Accountability sits with the manufacturer
RSNA holds that AI manufacturers, not radiologists, must remain accountable for a device's safety and effectiveness — a direct statement about where liability for a flawed output should land as these tools scale.
Model changes need to be disclosed, fast
The society wants the regulatory process to require timely, detailed communication whenever a generative AI model or its system configuration changes — so a clinical user isn't the last to know the tool behind their reports was updated.
What ACR told the FDA
ACR's comments, filed by the same deadline, pushed on a related but distinct point: regulate the whole clinical function of a tool, not just the narrow feature named on its label. Per Radiology Business's coverage, ACR urged FDA to regulate the complete clinical function of an AI tool and any associated risk, to let the technology continue to evolve rather than freezing it at clearance, to balance premarket flexibility with real-world performance monitoring, and to "measure the failures that matter clinically" — not just technical accuracy on a benchmark. ACR said it welcomes continued partnership with FDA and the radiology community as the framework develops.
Read side by side, the two letters converge on the same underlying worry: that a generative AI tool cleared once, against a narrow test set, can keep changing in ways that matter clinically — and that premarket review alone can't catch that drift. Both societies want the regulatory system built around continuous, real-world accountability rather than a single pass/fail gate.
The buyer's version of these asks
None of this is binding yet, and it may be years before it is. But an imaging center evaluating an AI reporting vendor doesn't have to wait for a final rule to apply the same logic RSNA and ACR just put in writing. Each safeguard they asked FDA to eventually require has a direct, askable equivalent in a vendor conversation today.
| What the societies asked FDA to require | What to ask a vendor today |
|---|---|
| Enforceable safeguards for transparency, accountability, and clinical validation (RSNA) | Ask for the validation evidence behind the product — and whether it was tested against outcomes, not just internal benchmarks. |
| Risk from models that evolve post-deployment "cannot be addressed" by one-time premarket review (RSNA) | Ask how often the underlying model changes, and whether you are notified before a change reaches your reports. |
| Manufacturers, not radiologists, remain accountable for safety and effectiveness (RSNA) | Check what the contract actually says about liability for a flawed AI output. |
| Regulate the complete clinical function of a tool and its full risk, not just the labeled feature (ACR) | Ask what the tool actually touches across the reporting workflow — not only what its clearance language covers. |
| "Measure the failures that matter clinically," balanced against real-world monitoring (ACR) | Ask for ongoing, real-world performance monitoring after go-live — not a one-time accuracy number from launch. |
The point isn't that these questions are new. It's that they now carry the weight of formal professional-society input on the exact rulemaking that will eventually govern this category — which makes them a reasonable baseline for vendor diligence rather than a nice-to-have.
Where xAID fits
The workflow both societies are describing — a tool that assists, with a human clinician accountable for the final clinical decision — is the model AI CT reporting already runs on. xAID's AI drafts a structured report, xAID's in-house radiologist reviews every preliminary, and it arrives ready-to-sign: your reading radiologist signs the final. Model updates, validation evidence, and post-deployment performance are exactly the kind of questions worth asking any AI reporting vendor now — before a final rule makes them mandatory.
Frequently asked questions
What did RSNA tell the FDA about regulating generative AI?
In comments submitted October 19, 2026 on FDA's generative-AI discussion paper, RSNA wrote that "patient safety should remain the central organizing principle, and regulatory flexibility should be paired with clear, enforceable safeguards that promote transparency, accountability, clinical validation, and continuous real-world monitoring." RSNA also argued that generative AI models can evolve after deployment in ways a one-time premarket review cannot address, that device manufacturers — not radiologists — must remain accountable for safety and effectiveness, and that any changes to a model or its configuration should be communicated promptly and in detail.
What did the American College of Radiology recommend to the FDA?
ACR urged FDA to regulate the complete clinical function of a generative AI tool and the risk it carries, not just the narrow feature named in its clearance. It asked the agency to let the technology keep evolving while balancing premarket flexibility with real-world postmarket monitoring, and to "measure the failures that matter clinically" rather than relying only on technical accuracy metrics. ACR said it welcomes continued partnership with FDA as the framework develops.
Is FDA's generative AI discussion paper an actual regulation?
No. The paper, "Considerations for the Regulation of Generative AI-Enabled Medical Devices," was issued August 18, 2026 by FDA's Digital Health Center of Excellence under docket FDA-2026-N-7874, with public comment closing October 19, 2026. FDA has described it explicitly as a discussion paper meant to prompt stakeholder input, not draft guidance — it sets no binding requirements, and there is no published timeline for when, or whether, it becomes enforceable policy.
Should an imaging center wait for final FDA rules before adopting AI reporting tools?
Not necessarily. The safeguards RSNA and ACR are asking FDA to eventually require — documented clinical validation, post-deployment performance monitoring, prompt disclosure of model changes, and clear accountability for a flawed output — are reasonable due-diligence questions a buyer can put to any AI-reporting vendor today, regardless of how long rulemaking takes. Waiting for a final rule delays those questions; it doesn't answer them.
Source: Radiology Business, "Radiology societies weigh in on FDA's push to regulate generative AI" (2026). FDA discussion paper: FDA Digital Health Center of Excellence, docket FDA-2026-N-7874. Figures and quotes are reported as published; rounded where noted.