Source: JD Supra

Before deploying or testing AI tools, organizations must determine whether they are subject to HIPAA (see 42 USC § 1320d-6; 45 C.F.R. Parts 160, 162, and 164) because processing protected health information (PHI) in AI systems can constitute a HIPAA-regulated “use” and trigger Privacy, Security, and Breach Notification Rule obligations. PHI is described as information in an electronic health record or designated record set that (1) identifies an individual and (2) conveys information about that person’s health, treatment, or payment; if both elements are present, the information is likely PHI. The guidance emphasizes that PHI cannot be analyzed through publicly available AI models and that entities must assess whether processing occurs in a closed-loop environment, whether data are segregated across ingestion and output stages, and whether third parties can access the data.

Where PHI is involved, organizations should determine whether a HIPAA authorization is required or whether the TPO (treatment, payment, operations) exceptions apply, noting that a HIPAA authorization has specific regulatory requirements (45 CFR 164.508). Training general-purpose models on PHI is restricted: it requires a proprietary closed-loop system, appropriate permissions, and adequate destruction protocols, or use of masked/obfuscated or de-identified data. AI deployments frequently involve vendors, so business associate agreements (BAAs) should be evaluated and updated to expressly address AI processing, aggregation, de-identification, and data transmission within the scope of 164.504. The piece also warns that state privacy and AI laws (examples cited: the California Consumer Privacy Act, Washington’s My Health My Data Act, and certain Colorado restrictions) may impose additional or heightened requirements, that HIPAA is a floor rather than a preemption of more protective state laws, and that FDA or other federal regulations may be implicated for clinical decision-support tools. Finally, ownership and control of training data and model outputs should be contractually allocated, with organizations often using standardized AI riders to define data use and post-termination handling.