FDA's seven-step credibility framework for AI, in plain terms

What is the FDA draft guidance on AI for drugs and biologics?

In January 2025 FDA published the draft guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. It gives sponsors a risk-based framework for establishing and assessing the credibility of an AI model for a particular context of use. The comment period closed in April 2025.

A draft guidance describes FDA's current thinking and is not binding. It is still the clearest statement of what the agency expects to see when AI output supports a decision it will review.

When does it apply?

It applies to AI models that produce information or data intended to support regulatory decisions about the safety, effectiveness or quality of drugs and biological products. That reaches across the product life cycle: nonclinical work, clinical trials, postmarketing surveillance and manufacturing.

It does not address AI used in drug discovery, or AI used for operational efficiency, such as drafting or resource planning, where the use does not affect patient safety, drug quality or the reliability of study results.

What are the seven steps?

• Step 1. Define the question of interest: the specific question or decision the model will address.
• Step 2. Define the context of use: what the model does, what it is used for, and whether other evidence is used alongside its output.
• Step 3. Assess the AI model risk, from model influence and decision consequence.
• Step 4. Develop a plan to establish the credibility of the model output within the context of use.
• Step 5. Execute the plan.
• Step 6. Document the results of the credibility assessment plan and discuss deviations from it.
• Step 7. Determine whether the model is adequate for the context of use.

How is AI model risk assessed?

Model risk is not the risk of the model in the abstract. It is the risk that its output leads to a wrong decision with an adverse outcome. The draft builds it from two factors.

• Model influence: how much the AI output contributes to the decision, compared with other evidence. A model whose output is confirmed by an independent test has less influence than one that decides alone.
• Decision consequence: how serious the outcome is if the decision turns out to be wrong.

What does this look like in manufacturing and quality?

Take a model that screens batch data and flags batches for closer review. If every batch still goes through the full release testing and a qualified person makes the release decision, the model's influence is low: it directs attention and decides nothing. If the same model were used to skip testing on batches it scores as normal, its influence would be high, and so would the evidence FDA expects.

The practical lesson is that design choices move the risk. Keeping a qualified reviewer and an independent check in the process lowers model influence, and that lowers the validation burden along with the risk to the patient.

What should you document?

• The question of interest and the context of use, written precisely enough that a reviewer can tell what is in and out of scope.
• The model and how it was developed: the data used for training and tuning, and how it was managed.
• How the model was evaluated: test data independent of development data, the performance metrics and the acceptance criteria.
• Life cycle maintenance: how performance is monitored after deployment and how changes to the model are controlled.