FDA’s 2025 AI-device draft: govern the whole lifecycle

The FDA announced its lifecycle draft on 6 January 2025; the document itself was issued on 7 January. Its title is Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. The document explicitly says draft guidance and distribution for comment purposes only. It contains non-binding recommendations and is not for implementation as final guidance. The jurisdiction is the United States. Reading it as a final rule for every AI application in healthcare would misstate both its legal status and its subject matter. 1

The draft concerns devices containing AI-enabled software functions and information supporting marketing submissions. A scheduling assistant, a general-purpose chatbot and software with a diagnostic intended use are not interchangeable categories. Document what the product is intended to do, who uses it, the patient population and the conditions in which its output informs decisions. These boundaries determine what evidence is relevant. They also prevent a demonstration on convenient data from being mistaken for validation in a different clinical workflow, where missing information and time pressure can change system behaviour.

The FDA announcement emphasized transparency, bias and lifecycle risk management. For a proposed evaluation, examine which sites, devices and patient groups are represented and which are missing. Separate training, tuning and independent evaluation data, and inspect whether records from the same underlying source leak across splits. Report subgroup findings with their limitations rather than hiding variation behind one headline score. A dataset description is part of understanding where the system may fail. It is not enough to say that a model was trained on a large amount of medical information.

An accurate isolated output does not prove a safe workflow. Examine how users see uncertainty, recognize unsupported inputs and recover when the system is unavailable. Define when an output requires independent review and how disagreement is handled. These are governance questions, not patient-specific medical recommendations. Connect them to AI literacy only as a broader regulatory context, and to AI security governance for access and incident controls. A clinician’s nominal presence does not establish effective oversight if the interface hides the information needed to challenge the result.

The total-product-lifecycle perspective means evidence cannot stop at the initial model evaluation. Data collection practices, equipment and patient populations can change. A proposed monitoring plan should identify signals, review responsibility, escalation thresholds and a rollback process. Keep the model version and relevant configuration connected to each evaluation record. Do not interpret the draft as permission for unlimited autonomous updates. The FDA announcement explicitly distinguished this lifecycle draft from separate guidance on predetermined change control plans. Different documents should not be combined into an invented blanket authorization.

A Swedish purchaser needs an assessment of applicable European and national requirements, not an assumption that an FDA document replaces them. Company knowledge can help organize versioned documentation, but clinical validation and regulatory classification need qualified specialist review. Published on 25 October 2026, this analysis describes the January 2025 draft and its historical status. It does not verify whether a later final document superseded it, nor whether any particular product is authorized. It provides neither medical nor legal advice and makes no claim about patient outcomes or deployment success.