FDA Draft Guidance Sets a Credibility Framework for AI Used in Drug Submissions

The FDA issued the draft guidance on 6 January 2025, and a notice of availability appeared in the Federal Register on 7 January 2025 under docket FDA-2024-D-4689. It was the agency’s first guidance focused on the use of artificial intelligence in the development of drugs and biological products, and it invited public comment until 7 April 2025.

The scope is defined by purpose rather than by technology. The draft applies to AI models that produce information or data intended to support regulatory decision-making about the safety, effectiveness or quality of a drug or biological product. Commentaries on the draft note that this can include uses across nonclinical, clinical, postmarketing and manufacturing activities. It does not apply to AI used in drug discovery, or to AI used to streamline operations, such as drafting a regulatory submission, where the use does not affect patient safety, drug quality or the reliability of results.

Because the guidance is in draft form, its recommendations are not binding and may change before any final version is published. At the time of checking for this brief, the FDA’s own guidance pages could not be accessed directly. Secondary sources dated up to September 2026 reported that the document was still labelled as a draft, and no final version could be confirmed.

The central proposal is a risk-based framework for establishing and documenting that an AI model is credible for a specific context of use, meaning the precise role and scope of the model in answering a regulatory question. The draft sets out seven steps: define the question of interest; define the context of use of the AI model; assess the model risk; develop a plan to establish the credibility of the model output within that context of use; execute the plan; document the results and any deviations from the plan; and determine whether the model is adequate for the context of use.

Model risk is assessed by combining two factors. The first, model influence, concerns how much weight is placed on evidence derived from the AI model compared with other evidence. The second, decision consequence, concerns how significant an incorrect decision would be, for example for patient safety. According to law firm analyses of the draft, the FDA expects the level of detail disclosed about a model and its credibility activities to increase as model risk increases. One commentary observes that the draft does not itself define a scale for these two factors.

The draft also addresses life cycle maintenance. Sponsors are expected to manage changes to a model, and to monitor its performance, so that it remains fit for its context of use over the life of the product. This is particularly relevant to models that are retrained or that operate on data that change over time, such as models used in manufacturing or postmarketing surveillance.

On 14 January 2026 the FDA and the European Medicines Agency published a joint set of ten guiding principles for good AI practice in drug development. As summarised by NSF, the principles cover human-centric design, a risk-based approach, adherence to standards, a clear context of use, multidisciplinary expertise, data governance and documentation, model design and development practices, risk-based performance assessment, life cycle management and clear essential information. The principles are high level and non-binding, and they are intended to inform future guidance in both jurisdictions.

The joint principles use several of the same concepts as the January 2025 draft, notably context of use, risk-based assessment and life cycle management. Sponsors developing a global evidence strategy can therefore expect broadly consistent expectations from the two agencies, while recognising that neither document creates binding requirements and that a final FDA guidance could differ in detail from the draft.

Although the guidance is not final, it is the clearest statement of the FDA’s current thinking on the evidence it expects when AI-derived information is submitted. Sponsors using AI to support decisions on dosing, patient selection, safety signal assessment, manufacturing controls or similar questions should be able to state the question of interest and the context of use in plain terms, and to explain how much the regulatory decision depends on the model output.

The draft encourages early engagement with the FDA to set expectations on credibility activities for a proposed model and to address potential challenges. Sponsors planning a higher-risk use are likely to benefit from raising it in a formal meeting before the evidence is generated, rather than at the time of submission. Documentation of training data, performance metrics, deviations from the credibility plan and change management should be prepared with the expectation that it may be requested.