FDA's enforcement activity on AI medical devices is still relatively limited — the cleared device base is young, and the agency has been focused on building the clearance framework rather than enforcement. But the warning letters and enforcement actions that have been issued reveal where FDA is drawing its lines.

AI DEVICE WARNING LETTER VIOLATIONS BY CATEGORY Software validation deficiencies 43% Labeling inaccuracies / omissions 30% Unauthorized algorithm modifications 18% Quality system / design control gaps 9% Approximate distribution based on publicly available CDRH warning letters citing AI/ML device violations
Software validation documentation deficiencies account for the largest share of FDA warning letters to AI device manufacturers. Illustrative distribution based on public enforcement records.

The three violation categories FDA is finding

Software validation violations are the most common. Under 21 CFR Part 820.30, device manufacturers must validate device design including software. For AI devices, this means documenting validation methodology, maintaining records of training and test datasets, and verifying performance across the clinical conditions of intended use. FDA's inspectors are looking specifically for gaps between how the model was actually developed and what the validation documentation says.

Labeling violations fall into two patterns: overstated performance claims, and inadequate disclosure of the device's limitations and conditions under which performance may be degraded. Warning letters have cited manufacturers for failing to update labeling when post-market experience revealed performance gaps not reflected in the original labeling.

Unauthorized modification violations are the most AI-specific: modifications to the cleared algorithm that were not covered by the original clearance or an approved PCCP. FDA has taken action against manufacturers who updated their AI model without submitting a new 510(k). These cases reveal the enforcement risk of continuous deployment practices common in software development.

What's notably absent from enforcement

FDA has not yet brought enforcement actions specifically addressing AI algorithmic bias or AI device failures in clinical settings where adverse events occurred. This is likely to change as the bias guidance matures and the cleared AI device base grows. Companies operating cleared AI devices should treat current warning letters as signals, not as the ceiling of FDA's enforcement interest.

ⓘ This article is for informational purposes only and does not constitute regulatory or legal advice. AIFDA Intel is an independent platform and is not affiliated with the U.S. Food and Drug Administration. Consult a qualified regulatory affairs professional before making regulatory decisions.