Author ORCID iD

Document Type


Publication Date



Artificial Intelligence (“AI”), particularly its subset Machine Learning (“ML”), is quickly entering medical practice. The U.S. Food and Drug Administration (“FDA”) has already cleared or approved more than 520 AI/ ML-based medical devices, and many more devices are in the research and development pipeline. AI/ML-based medical devices are not only used in clinics by health care providers but are also increasingly offered directly to consumers for use, such as apps and wearables. Despite their tremendous potential for improving health care, AI/ML-based medical devices also raise many regulatory issues.

This Article focuses on one issue that has not received sustained attention in the legal or policy debate: labeling for AI/ML-based medical devices. Labeling is crucial to prevent harm to patients and consumers (e.g., by reducing the risk of bias) and ensure that users know how to properly use the device and assess its benefits, potential risks, and limitations. It can also support transparency to users and thus promote public trust in new digital health technologies. This Article is the first to identify and thoroughly analyze the unique challenges of labeling for AI/ML-based medical devices and provide solutions to address them. It establishes that there are currently no standards of labeling for AI/ML-based medical devices. This is of particular concern as some of these devices are prone to biases, are opaque (“black boxes”), and have the ability to continuously learn. This Article argues that labeling standards for AI/ML-based medical devices are urgently needed, as the current labeling requirements for medical devices and the FDA’s case-by-case approach for a few AI/ML-based medical devices are insufficient. In particular, it proposes what such standards could look like, including eleven key types of information that should be included on the label, ranging from indications for use and details on the data sets to model limitations, warnings and precautions, and privacy and security. In addition, this Article argues that “nutrition facts labels,” known from food products, are a promising label design for AI/MLbased medical devices. Such labels should also be “dynamic” (rather than static) for adaptive algorithms that can continuously learn. Although this Article focuses on AI/ML-based medical devices, it also has implications for AI/ ML-based products that are not subject to FDA regulation.

Publication Title

George Washington Law Review