Clinical software that gives one answer is no longer a device, the FDA says
Revised guidance reverses the 2022 position that a single recommendation made software a device. The agency's own criterion — that a clinician can independently review the basis — now carries the weight.
On 6 January the FDA issued revised final guidance on Clinical Decision Support Software [s1]. It was announced not at a medical meeting but at CES in Las Vegas, alongside a parallel revision to the agency's general wellness policy [s3].
The substantive change is narrow to state and broad in consequence. Software that gives a clinician one recommendation is no longer, by that fact, a regulated medical device.
What the 2022 position was
Non-device clinical decision support is defined by four criteria drawn from Section 3060 of the 21st Century Cures Act. Software falls outside device regulation if it does not acquire, process or analyse medical images or signals from diagnostic devices; is intended for displaying, analysing or printing medical information; supports or provides recommendations to a health care professional about prevention, diagnosis or treatment; and enables that professional to independently review the basis for the recommendation, such that the intent is not for the clinician to rely primarily on the tool [s1].
Under the 2022 guidance, the agency read the fourth criterion as effectively requiring software to present multiple options. A tool that surfaced a list of possibilities, with the reasoning attached, left the clinician doing the deciding. A tool that produced one answer did not, and was therefore a device.
What changed
The FDA now exercises enforcement discretion for tools producing a singular output where only one recommendation is clinically appropriate [s1] — a direct reversal of the earlier position [s2].
The agency also extended enforcement discretion to a specific and consequential use case: software that analyses a radiologist's findings to generate a proposed report summary containing diagnostic recommendations, provided a health care provider remains in the loop and the tool does not analyse the underlying images [s1].
Two further adjustments matter. Time-critical decision-making has been relocated from the third criterion to the fourth, though the regulatory approach itself is unchanged [s1]; FDA no longer automatically excludes software supporting time-critical decisions from non-device status [s2]. And the guidance places stronger emphasis on usability and on drawing from well-understood and accepted sources such as clinical guidelines and peer-reviewed literature [s1].
The load-bearing criterion is now "independently review"
Strip the changes back and the entire non-device category rests on one thing: whether a clinician can independently review the basis for a recommendation.
This is where the FDA's own framing gets interesting, because the guidance explicitly invokes automation bias. It directs attention to the time-sensitive nature of clinician decisions, so that clinicians can genuinely evaluate recommendations independently — particularly with algorithmic and AI-driven tools requiring immediate clinical action [s2].
That is the agency naming the failure mode correctly. Independent review is not a property of software. It is a property of a person, working in a particular setting, with a particular amount of time, at a particular point in a shift. A tool can display its reasoning perfectly and still be relied on primarily, because the clinician has ninety seconds and eleven more patients.
The guidance handles this by attending to whether the deployment permits review. Whether that translates into practice is not something guidance can determine.
The AI gap
The most notable thing about a guidance the FDA framed as an innovation measure is what it does not address.
The document is silent on AI and provides no specific policy for AI-enabled CDS products [s1]. That leaves uncertainty around clinician-facing medical chatbots and AI-generated clinical communications [s1].
This is a strange place for the regulatory perimeter to sit. Generative systems are precisely the category where "independently review the basis for the recommendation" is hardest to satisfy — a model can produce a fluent, cited-looking rationale that is not the actual basis for its output. The criterion assumes reasoning is inspectable. For a class of systems now entering clinical workflows, it may not be.
What the announcement was for
The venue tells you the intent. Commissioner Marty Makary framed the broader package around predictability for developers, saying the agency wants companies to know that if their device or software is simply providing information, they can do that without FDA regulation, with the stipulation being claims of medical-grade performance [s3]. He also indicated the agency would not act on isolated inaccuracies [s3].
Read as industrial policy that is coherent. Read as patient safety policy it relocates responsibility rather than reducing it. Software that once required a premarket submission now requires a purchasing health system to evaluate it — for accuracy, for population fit, for how it behaves when the underlying data drifts.
Large academic centres have the informatics capacity to do that. Most hospitals do not, and nothing in the guidance gives them a standard to buy against.
What to watch
Three things over the coming year. Whether FDA issues the AI-specific CDS policy this guidance leaves out. Whether the radiology report-summary carve-out expands to other specialties. And whether health systems begin publishing their own validation of tools that no longer arrive with any.
Sources
- [s1] 5 Key Takeaways from FDA's Revised Clinical Decision Support (CDS) Software Guidance. Covington & Burling LLP, January 2026. https://www.cov.com/en/news-and-insights/insights/2026/01/5-key-takeaways-from-fdas-revised-clinical-decision-support-cds-software-guidance
- [s2] Key Updates in FDA's 2026 General Wellness and Clinical Decision Support Software Guidance. Faegre Drinker Biddle & Reath LLP, January 2026. https://www.faegredrinker.com/en/insights/publications/2026/1/key-updates-in-fdas-2026-general-wellness-and-clinical-decision-support-software-guidance
- [s3] FDA announces regulatory exemptions for 'non-medical grade' devices. Medical Device Network, 7 January 2026. https://www.medicaldevice-network.com/news/fda-announces-regulatory-exemptions-for-non-medical-grade-devices/
Sources
- 5 Key Takeaways from FDA's Revised Clinical Decision Support (CDS) Software Guidance — Covington & Burling LLP , January 6, 2026
- Key Updates in FDA's 2026 General Wellness and Clinical Decision Support Software Guidance — Faegre Drinker Biddle & Reath LLP , January 6, 2026
- FDA announces regulatory exemptions for 'non-medical grade' devices — Medical Device Network , January 7, 2026
The ACP has issued ethical guideposts for AI at the bedside. There are three.
Relationality, self-governance, competence. The position paper's starting premise is that consensus on privacy, disclosure and fairness has not been reached, and clinicians need guidance anyway.
Sepsis AI reaches an AUROC of 0.88 and a positive predictive value of 34.2%
A network meta-analysis of 53 studies and more than 7 million admissions finds machine learning out-discriminates traditional sepsis scores — and would raise roughly two false alarms for every real one.
UAE hospitals are using AI to flag sepsis six hours before doctors would catch it
The tool reads vital signs, lab results, and history the way clinicians already do — just faster and continuously. Its accuracy tops out around 90%, which means it's also still wrong a meaningful share of the time.
FDA opens comment on how to regulate AI medical devices that generate their own answers
A new discussion paper proposes a two-axis risk framework and a physician-training analogy for evaluating generative AI devices. It is not a rule, and the agency is asking what one should look like.