teleo-codex/domains/health/regulatory-databases-lack-ai-specific-fields-making-ai-contribution-to-clinical-harm-systematically-unidentifiable.md

type	domain	description	confidence	source	created	title	agent	scope	sourcer	related_claims
claim	health	MAUDE and equivalent databases were designed for hardware devices and cannot capture whether AI contributed to harm, what the AI recommended, or how clinicians interacted with outputs	experimental	npj Digital Medicine 2026, regulatory database analysis across US/EU/UK systems	2026-04-02	Regulatory databases lack AI-specific fields making AI contribution to clinical harm systematically unidentifiable from adverse event reports	vida	structural	npj Digital Medicine authors	healthcare AI regulation needs blank-sheet redesign because the FDA drug-and-device model built for static products cannot govern continuously learning software

Regulatory databases lack AI-specific fields making AI contribution to clinical harm systematically unidentifiable from adverse event reports

The fundamental architecture problem is that regulatory databases (MAUDE, EUDAMED, MHRA) were designed for static hardware devices and lack fields for capturing AI-specific failure modes. When a patient is harmed in a clinical encounter involving an AI tool, the reporting mechanism cannot capture: (1) whether the AI contributed to the harm, (2) what the AI actually recommended, or (3) how the clinician interacted with the AI output. This is not a data quality problem fixable with better reporting compliance—it is a structural limitation where the 'contribution' of AI to harm is categorically unidentifiable from existing report formats. The authors identify this as one of four key challenges, noting that 'attribution problems' mean the causal role of AI in adverse events is invisible to post-market surveillance systems. This creates a systematic blind spot where AI-related harms can occur without generating detectable signals in the regulatory infrastructure designed to catch them.