FDA and AI-enabled device software: the 2025 draft guidance, the device list and the 2026 CDS guidance, explained
Three FDA documents shape how AI in clinical software is reviewed: a January 2025 lifecycle draft guidance that is still a draft, a public list of authorized AI-enabled devices, and final clinical decision support guidance issued in January 2026.
Three documents with three different jobs
A health system or developer looking at AI in clinical software meets three FDA documents. The first is a draft guidance on what a manufacturer should put in a marketing submission for an AI-enabled device [1]. The second is a public list of AI-enabled devices FDA has authorized [4]. The third is final guidance on which clinical decision support software is not a device at all [6]. They answer different questions: how a device is reviewed, which devices have been authorized, and whether FDA oversight applies in the first place.
The January 2025 draft guidance
FDA announced the draft guidance in the Federal Register on January 7, 2025 and asked for comments by April 7, 2025 to ensure they were considered before work on a final version [2]. The notice says the draft “is not final nor is it for implementation at this time” [2]. As checked on November 14, 2026, FDA’s guidance page still labels it a draft, “Not for implementation,” under docket FDA-2024-D-4488, with the page content current as of January 7, 2025 [1].
When finalized, it would recommend the contents of marketing submissions for devices that include AI-enabled device software functions, using a total product life cycle approach, and would offer design and development recommendations manufacturers may wish to consider [1]. The draft is organized by the parts of a submission [3]:
- device description;
- user interface and labeling;
- risk assessment;
- data management;
- model description and development;
- validation, including performance validation;
- device performance monitoring;
- cybersecurity;
- public submission summary, with appendices that include an example model card and an example 510(k) summary with a model card.
Several recommendations bear on data and operations more than on the algorithm itself:
- Data management. The draft asks sponsors to explain how data has been or will be collected, processed, annotated, stored, controlled and used, and to characterize the data used in development and validation [3].
- Test independence. The model description should compare the training dataset to the validation dataset and to expected inputs, and describe how independence of test data from training data was ensured [3].
- Drift. The draft notes that AI-enabled devices can be sensitive to differences between development data and data in actual deployments, called data drift, and addresses performance in the postmarket setting [3].
- Cybersecurity. For devices that are cyber devices under section 524B of the FD&C Act, the draft lists AI-specific risks such as data poisoning, through injected inauthentic or maliciously modified data, and model inversion or stealing [3].
- Model card. The draft says sponsors should consider using a model card to organize information about the model [3].
Change control plans are already final
One related piece is final. On December 4, 2024, FDA announced final guidance on predetermined change control plans (PCCPs) for AI-enabled device software functions [7]. A PCCP describes planned modifications, the methodology to develop, validate and implement them, and an assessment of their impact. FDA reviews it as part of the marketing submission, so modifications described in an authorized PCCP do not each need a new submission [7]. The 2025 draft points sponsors to that guidance [3].
The AI-Enabled Medical Device List
FDA publishes a list intended to identify AI-enabled medical devices authorized for marketing in the United States [4]. Key points from the page [4]:
- Listed devices have met FDA’s applicable premarket requirements, and each entry links to the device’s record in FDA’s databases.
- The list is not comprehensive. Devices were identified primarily by AI-related terms in the summary descriptions of their marketing authorization documents or in their classification, using the terms in the FDA Digital Health and Artificial Intelligence Glossary.
- FDA says it will explore ways to identify and tag devices that incorporate foundation models, including large language models, and encourages sponsors to include that information in their public summaries.
- The list is updated periodically. Devices whose decision summaries were not published within a data collection period appear in a later update.
When checked, the page showed content current as of September 4, 2026 [4]. For a buyer, the list is a starting point for confirming that a product has an FDA authorization. Because FDA says it is not comprehensive, a product’s absence does not by itself show that it lacks one [4].
The January 2026 clinical decision support guidance
FDA issued final guidance on clinical decision support (CDS) software on January 29, 2026. It supersedes a version issued January 6, 2026 [6]. It sits under docket FDA-2017-D-6569 [5]. The guidance explains which CDS functions intended for health care professionals are excluded from the device definition by section 520(o)(1)(E) of the FD&C Act, added by the 21st Century Cures Act [6]. A function is excluded only if it meets all four criteria [6]:
- it is not intended to acquire, process or analyze a medical image, a signal from an in vitro diagnostic device, or a pattern or signal from a signal acquisition system;
- it is intended to display, analyze or print medical information about a patient or other medical information, such as clinical studies and practice guidelines;
- it is intended to support or provide recommendations to a health care professional about prevention, diagnosis or treatment;
- it is intended to enable the professional to independently review the basis for the recommendations, so that the professional does not rely primarily on them.
Software intended to acquire, process or analyze a medical image or signal fails the first criterion, so it cannot use this exclusion [6]. Software that provides a specific preventive, diagnostic or treatment output or directive fails the third criterion [6]. The guidance adds an enforcement policy: if only one option is clinically appropriate and the function otherwise meets all the criteria, FDA intends to exercise enforcement discretion [6].
On the fourth criterion, FDA considers the level of software automation and the time-critical nature of the decision, because automation bias, the tendency to over-rely on an automated suggestion, increases when urgent action leaves no time to consider other information [6]. The guidance says the software or labeling should describe the algorithm’s development and validation in terms the intended user can follow, including the general approach, such as meta-analysis, expert panels, statistical modeling or AI and machine learning techniques [6]. Like other FDA guidance, it is nonbinding [6].
Reading the three together
For a health system, the documents suggest a short set of questions about any AI function it buys or builds. Does it analyze images or physiological signals? Does it issue a directive rather than a recommendation? Can a clinician see the basis for its output? Those questions follow the four CDS criteria [6]. If the function is a device, the lifecycle draft shows what FDA would expect a manufacturer to document, from data provenance to drift monitoring [3], and the device list shows whether an authorization is on record [4]. That reading is ours; the documents themselves address manufacturers and FDA staff.
This explainer describes FDA guidance; it is not legal or regulatory advice, and organizations should confirm with counsel and regulatory specialists how it applies to their products.
Moraph’s teams help organizations with work like this. See our solutions.
Spotted an error? Tell us and we will check it and correct the explainer where needed.