Short answer: yes, US healthcare providers can use Claude and ChatGPT under HIPAA — but only on BAA-covered plans for anything touching Protected Health Information, and only after your practice has a written data-classification policy, a tool allowlist, and trained staff behind it. Everything else in this guide is how to get there.
HIPAA Privacy Rule meets generative AI
The HIPAA Privacy Rule governs Protected Health Information (PHI): any individually identifiable health information a covered entity creates, receives, maintains, or transmits. That includes the obvious (name, date of birth, MRN, insurance ID, diagnosis, treatment notes) and the less obvious (ZIP code combined with age, appointment times, device identifiers, voice recordings).
When a clinician or admin pastes PHI into a chat window, HHS treats it as a disclosure. If the AI vendor is not a Business Associate under a signed agreement, that disclosure is impermissible — full stop. It does not matter that the data was typed rather than emailed, that the chat was private, or that the model "forgets" after the session. The disclosure happened when the characters left your network.
The practical result: every generative AI prompt your team writes is a mini risk assessment. Either the destination is covered by a BAA and the data can include PHI, or it is not and the data must be de-identified before it arrives.
Business Associate Agreements: who offers one in 2026
A Business Associate Agreement is the contract that lets a vendor touch PHI on your behalf. It imposes HIPAA Security Rule duties, breach notification timelines, and audit rights. Without one, the vendor is a stranger in HIPAA's eyes.
- Anthropic Claude. Anthropic offers a BAA on Claude for Work (Enterprise) and on Claude via Amazon Bedrock and Google Cloud Vertex AI when those cloud providers' BAAs are in place. The consumer claude.ai free and Pro plans are not HIPAA-covered.
- OpenAI ChatGPT. OpenAI offers a BAA on ChatGPT Enterprise, on the ChatGPT Team plan when a BAA is specifically executed, and on the OpenAI API. Free, Plus, and Pro consumer tiers are not HIPAA-covered.
- Microsoft Copilot. Microsoft 365 Copilot, when deployed under an existing Microsoft BAA for M365, inherits that coverage — one of the simplest routes for practices already on Microsoft 365.
- Google Gemini. Gemini for Google Workspace is covered under Google's existing Workspace BAA when the customer has one in place.
A BAA is required whenever the AI vendor will process PHI. If your workflow strictly uses de-identified data, a BAA is not strictly mandated — but most practices find it simpler to standardize on a BAA-covered tool than to police the line prompt by prompt.
De-identification: Safe Harbor vs Expert Determination
HIPAA recognizes two paths to lawfully de-identified data. Once data is de-identified correctly, HIPAA no longer applies to it — so it can safely flow to any AI tool.
Safe Harbor method
Remove all 18 categories of identifiers and confirm you have no actual knowledge the remaining information could re-identify someone. The 18 identifiers are:
- Names; geographic subdivisions smaller than a state (including ZIP, with narrow exceptions); all date elements (except year) tied to an individual; telephone, fax, and email; Social Security numbers; medical record numbers; health plan beneficiary numbers; account numbers; certificate/license numbers; vehicle identifiers; device identifiers and serial numbers; URLs; IP addresses; biometric identifiers; full-face photos and comparable images; and any other unique identifying number, characteristic, or code.
Expert Determination method
A qualified statistician certifies that the risk of re-identification is "very small." Useful when you need to keep more granular data (e.g. full dates for a clinical research dataset) and are prepared to pay for the analysis.
[Patient], DOB with [age X], address with region only, MRN with [Patient ID], and any dates with relative time ("2 weeks post-op"). Review the prompt out loud before sending. If a stranger could identify the person from what's left, keep redacting.OCR enforcement posture in 2024–2026
The HHS Office for Civil Rights (OCR) has not yet published an AI-specific breach settlement, but its 2024 and 2025 guidance is unambiguous: generative AI vendors that touch PHI are business associates and the usual rules apply. OCR's 2024 bulletin on tracking technologies — which has driven seven-figure settlements against health systems that let third-party scripts receive patient data — applies the same analytical framework. A consumer AI tool receiving PHI is, legally, the same shape of problem as a web analytics pixel receiving PHI.
Expect the first AI-specific OCR settlements to follow the tracking tech pattern: multi-million-dollar resolutions with a vendor-management corrective action plan attached.
State laws that layer on top
- California — CMIA. The Confidentiality of Medical Information Act extends to any provider, contractor, or corporation that handles medical information, with a private right of action and statutory damages. California's CPRA also treats health data as sensitive personal information with extra disclosure duties.
- Texas — HB 300. Expands the definition of covered entity well beyond HIPAA (anyone who assembles, collects, analyzes, uses, evaluates, stores, or transmits PHI), mandates biennial training within 90 days of hire, and gives the Texas AG enforcement authority.
- New York — SHIELD Act. Applies reasonable administrative, technical, and physical safeguards to any business holding New Yorkers' private information, with breach notification duties that run alongside HIPAA's.
If your patient base spans multiple states, the strictest overlay effectively sets your floor.