FDA's approach to AI models used in support of a regulatory decision is built around a question that sounds simple and isn't: how much should you trust this model's output, for this specific purpose? The agency's answer is a risk-based credibility framework — evidence scales with how much the model's output matters to the decision and how bad it would be for the model to be wrong.
Current as of September 2026 — FDA guidance in this area continues to evolve; verify current expectations against the agency's published guidance and, where relevant, through pre-submission engagement.
Context of use comes first
Credibility isn't a property of a model in isolation — it's a property of a model applied to a specific, precisely stated context of use: what question it answers, on what population or data, feeding into what decision. A model can be highly credible for a narrow, well-characterised context of use and inadequately supported for a broader one, even with no change to the model itself. This is the single most common place submissions run into trouble: the context of use as actually deployed drifts from the context of use the credibility evidence was built for.
Risk scales the evidence bar
Two dimensions matter: how much the model's output influences the ultimate decision, and how consequential that decision is if the model is wrong. A model contributing one supporting data point among several, all reviewed by a qualified expert before the decision is made, sits at lower risk than a model whose output is treated as largely determinative. Higher risk demands a heavier credibility package — broader validation, more rigorous characterisation of failure modes, more independent verification — not because the agency distrusts AI specifically, but because that's how evidentiary burden works for any tool used to support a consequential decision.
What a credibility package tends to include
A clearly stated context of use; a description of the model, its training and validation data, and known limitations; performance evidence appropriate to that context of use; and an honest account of where the model is expected to underperform or where its use has not been characterised. The honesty matters — a credibility argument that only presents favourable evidence tends to invite exactly the scrutiny it was trying to avoid.
Engaging early beats engaging late
Because credibility is assessed against a specific context of use, teams that lock in the intended context of use early and engage with the agency about it tend to avoid the expensive failure mode: building a substantial evidence package, then discovering the actual deployment's context of use has drifted from what that package supports.
Where Eclypse sits: deployments document a fixed context of use, performance evidence against that context, and known limitations as standard deliverables — the shape a credibility argument needs, built in rather than assembled retroactively for a submission.
This discipline pairs closely with GxP validation for AI systems and with hallucination mitigation in regulated settings, both of which are ultimately about the same thing: knowing precisely what a model is good for, and proving it.
Common questions about FDA's AI credibility framework.
What is a credibility assessment for an AI model?
A structured, evidence-backed argument for why a model's output can be trusted for a specific, defined context of use — risk-based, so evidence scales with the decision's consequence.
What is context of use in FDA's credibility framework?
A precise statement of what question the model answers, on what data, feeding into what decision. Credibility is assessed against that specific statement, not the model in the abstract.
Does using AI in a submission require a separate approval?
Not a separate approval, but credibility evidence becomes part of the submission's evidentiary package and is reviewed accordingly. Early agency engagement is generally advisable.
How is AI model risk assessed?
By the model's influence on the decision and the consequence of it being wrong — a supporting data point reviewed by an expert carries lower risk than an output accepted at face value.
Does the credibility framework apply outside FDA submissions?
The underlying logic — precise context of use, risk-scaled evidence, honest documentation of limitations — is a reasonable default for AI in any consequential regulatory or clinical decision.
Tell us the context of use — we'll show you what evidence the deployment already produces.
Book a working session
ECLYPSE
AI
Our proprietary AI orchestration platform for healthcare: one engine, a registry of reusable task modules and domain agents, and a governed knowledge base — deployed inside your walls and run by your team.