What is HIPAA, and how does it apply to AI?
HIPAA, the Health Insurance Portability and Accountability Act of 1996, is the US federal law that governs how protected health information (PHI) is used, shared, and secured. PHI is health, treatment, or payment information that can be tied to a specific person. The law applies to health plans, healthcare clearinghouses, and most healthcare providers, known as covered entities, and to their business associates, the vendors that handle PHI on their behalf.
HIPAA does not mention AI. It regulates the data, so any model, agent, or transcription tool that touches PHI falls under the same rules as the health record. The Privacy Rule limits who may use PHI and why. The Security Rule requires safeguards for electronic PHI. The Breach Notification Rule requires notice to patients and regulators when unsecured PHI is exposed.
How HIPAA works for AI systems
Every vendor that handles PHI must sign a business associate agreement (BAA) before patient data reaches it. In an AI system, that can include the model vendor, the cloud host, the transcription service, and a hosted vector database. A clinician who pastes a patient note into a personal chatbot account has disclosed PHI without a BAA. The practice must treat that as a presumed breach unless a documented risk assessment shows a low probability of compromise.
Prompts, outputs, and logs that contain patient data are electronic PHI, and embeddings built from patient records should be protected the same way, with access controls, audit logging, and encryption. Under the minimum necessary standard, an assistant should retrieve only the records its task requires. Data de-identified under the Privacy Rule, by removing 18 listed identifiers or through a qualified expert’s determination, is no longer PHI, which makes it the usual route for model training.
HIPAA in practice
A home health agency wants an assistant that turns nurses’ visit dictation into draft documentation. To avoid adding a model vendor as a new business associate, the agency runs an open-weight model on cloud servers already covered by its cloud provider’s BAA. Drafts and logs stay in that environment under the clinical record’s access rules. The agency adds the system to its HIPAA risk analysis, and each nurse reviews and signs the final note.
What to watch for
- Ask each AI vendor whether it signs a BAA, which products the BAA covers, and whether it may keep prompts or train on them.
- Test PII redaction tools on real clinical notes, since general tools often miss medical record numbers and names in free text.
- HIPAA does not address whether a model’s output is accurate or fair, so clinical validation and bias testing need their own review.