Who does HIPAA actually apply to?
HIPAA binds covered entities — health plans, most healthcare providers, and healthcare clearinghouses — and, through contracts, their business associates: any vendor that touches PHI on their behalf. That includes software vendors, transcription services, analytics providers and offshore development teams.
- Covered entity — bound by HIPAA directly (a hospital, a clinic, an insurer).
- Business associate — bound through a BAA: must safeguard PHI, use it only as agreed, report incidents, and flow the same duties down to subcontractors.
- Neither — a consumer wellness app with no covered-entity relationship is typically outside HIPAA, though other privacy laws still apply.
Why is sending PHI to an AI tool a problem?
Because giving PHI to an external vendor is a disclosure under the Privacy Rule, and disclosures to service providers generally require a BAA. A clinician or vendor employee pasting a discharge summary into a consumer chatbot discloses PHI to a company with no BAA, no agreed safeguards and no HIPAA duties toward the patient.
Vendor policies vary and change; treat them cautiously and generically. Unless your organization has a signed BAA covering a specific AI service and has approved it for PHI, the safe assumption is simple: no identifiable patient data goes in — the same discipline as knowing what never to paste into ChatGPT.
How does de-identification make AI use lawful?
HIPAA's restrictions apply to identifiable information. Data that has been properly de-identified is no longer PHI, and the Privacy Rule's limits no longer attach — so a de-identified document can go into an AI tool without a BAA.
| Method | What you do | When it fits |
|---|---|---|
| Safe Harbor | Remove 18 categories of identifiers — names, granular locations, exact dates, phone numbers, emails, record and account numbers, photos and more — with no actual knowledge the person could be re-identified | The simple, common route; fits day-to-day document work |
| Expert Determination | A qualified expert applies statistical methods and documents that re-identification risk is very small | When you must keep fields Safe Harbor removes, e.g. for research |
What does a practical de-identification workflow look like?
The workflow is the same one used for any document anonymization, tightened to the Safe Harbor list. Four steps cover most documents:
- Detect every identifier — direct (John Miller, MRN, phone) and contextual (exact dates, rare conditions plus location).
- Replace consistently — a stable token per identity (PERSON_001 everywhere) keeps the clinical narrative readable instead of full of holes.
- Clean metadata — author fields, comments and tracked changes in Office files often carry patient or clinician names.
- Keep it local — if the de-identification tool uploads the original file to its own server, you have disclosed PHI to yet another vendor. Processing in the browser avoids creating a new disclosure.
What about AI vendors that do sign BAAs?
Some enterprise AI offerings will sign BAAs, and for integrated clinical products that can be the right path. Even then, data minimization remains good practice — feeding the model only what the task needs. For everything outside that approved channel — drafts, summaries, ad-hoc questions — de-identification first is the rule that keeps AI use in healthcare inside the lines.