Is ChatGPT HIPAA compliant?
Not by default, and not on the plans most people use. As of September 2026 OpenAI will not sign a business associate agreement for ChatGPT Free, Plus, Team or Enterprise, or for the consumer ChatGPT Health product, so patient information entered into any of them is an impermissible disclosure. OpenAI will sign a BAA for ChatGPT for Healthcare, the enterprise product launched in January 2026, and for the API when it is embedded in a healthcare application by an organization that meets its requirements. With the agreement signed and the workspace configured, those two can be used with PHI. Nothing else can.
Every major AI vendor, BAA by BAA.
OpenAI, Anthropic, Google, Microsoft and Amazon, with what each will and will not sign.
Every major AI vendor, BAA by BAASend us what you have. We’ll tell you if it’s right.
Your current policy paperwork, a renewal notice, or nothing at all. A licensed agent reads it and replies the same business day.
The detail
OpenAI will not enter into a business associate agreement for the Free, Plus, Team or Enterprise versions of ChatGPT.
ChatGPT for Healthcare, launched in January 2026 for hospitals and clinicians, is offered under a business associate agreement with qualifying healthcare organizations and is not HIPAA compliant out of the box until configured.
The OpenAI API can be covered by a BAA when embedded in a healthcare application, subject to the organization meeting HIPAA requirements on its side.
Properly de-identified data, using a method the HIPAA Privacy Rule permits, falls outside HIPAA and may be used with the standard ChatGPT versions.
Azure OpenAI Service offers the same OpenAI models under Microsoft’s BAA, included in the Online Services Data Protection Addendum, for production text workloads.
What this means for you
- The plan name on the invoice decides it.ChatGPT Enterprise sounds like the compliant tier and is not one. A practice that bought Enterprise seats for HIPAA reasons has a policy gap, not a technology gap, and the fix is a different product under a different agreement.
- The API route runs through you.A vendor building a clinical tool on the OpenAI API can be covered, but the BAA and the controls are the vendor’s to hold. A practice using that vendor needs a BAA with the vendor, and the vendor needs one with OpenAI. Ask to see both.
- De-identification is a method, not an intention.Removing the name is not de-identification. The Privacy Rule permits two methods, expert determination and safe harbor, and safe harbor strips eighteen identifiers. Notes that still carry a date of service and a zip code are PHI.
- Your insurance inherits the answer.A breach through an uncovered ChatGPT account is a violation you caused. Cyber pays the response on most forms; a penalty for a knowing violation may not be covered, and the 2026 renewal may carry AI-specific wording either way. Put the approved tools in writing and hand the list to your carrier.
Related questions
- Is ChatGPT Enterprise HIPAA compliant?
No. OpenAI states it will not sign a business associate agreement for ChatGPT Enterprise. The product built for regulated healthcare use is ChatGPT for Healthcare, launched in January 2026, which is offered under a BAA and still needs configuring before it is compliant. Enterprise is a data-privacy tier, not a HIPAA tier, and the distinction is the one that produces breaches in practices that thought they had paid for compliance.
- Can a therapist or a small practice use ChatGPT at all?
For anything without patient information, yes. For notes, letters, summaries or messages that contain PHI, only through a product under a BAA: ChatGPT for Healthcare, a clinical tool built on the API by a vendor who will sign a BAA with you, or a model reached through Azure, Vertex AI or Bedrock under those providers’ agreements. A one-person practice is a covered entity like any other and the penalties do not scale down.
- Does using ChatGPT with de-identified data need a BAA?
No, because de-identified data is outside HIPAA. The catch is the method: the Privacy Rule recognizes expert determination or the safe-harbor removal of eighteen identifiers, including dates more specific than the year and any geographic unit smaller than a state. A clinical note with the name deleted usually still fails safe harbor. Do the de-identification properly and document it, or treat the data as PHI.
- What should we do if staff have already been using ChatGPT with patient data?
Treat it as a potential breach, because it is one. Stop the practice, identify what was entered and when, and run the four-factor risk assessment the Breach Notification Rule requires; if you cannot show a low probability the information was compromised, notification follows. Tell your cyber carrier early, since most policies cover the assessment and the notification and require prompt notice. Then write the tool policy that should have existed.
- Cyber, and what it pays after a disclosure
- Is Claude, Gemini or Copilot any different?
The pattern is the same for all of them: consumer plans are excluded, specific enterprise or cloud tiers are covered under a BAA. Anthropic signs for the Claude API, Bedrock and HIPAA-ready Claude Enterprise; Google for Gemini in Workspace and through Vertex AI; Microsoft through the DPA that covers Azure OpenAI and eligible Microsoft 365 services. The product names change by the quarter, so check the vendor’s own BAA page on the day you sign.
- The vendor-by-vendor list
Need this handled?
Reading about it and having it handled are different jobs, and the second one is ours. We place the accounts other brokers decline — claims history, a cancellation, an unusual operation, a business nobody wants to underwrite twice — and we will tell you when the policy you already have is the right one. Tell us the situation and a licensed human replies the same business day.
Written by the licensed brokers at Boring Insurance. Last updated 2026-09-24. See all guides.