The honest answer to “Is OpenAI HIPAA compliant in 2026?” is no, not by default. OpenAI does not sign a Business Associate Agreement on its public API or standard ChatGPT. If you need to process PHI with OpenAI models, the realistic route is Azure OpenAI Service, where Microsoft signs the BAA and you still run OpenAI models.
Key takeaways
- OpenAI’s public API and standard ChatGPT are not covered by an OpenAI BAA.
- Azure OpenAI Service is the practical BAA route for OpenAI models; Microsoft is the business associate.
- Zero data retention is a setting, not a compliance shield, and it does not replace a BAA.
- PHI requires a signed BAA plus technical controls: no payload logging, region residency, and an audit trail.
- If you need per-request PHI routing across models, you likely need a gateway layer in front of the API.
Is OpenAI HIPAA compliant out of the box?
No. OpenAI’s public API and standard ChatGPT products are not HIPAA-compliant by default, and OpenAI does not offer a BAA on those routes. HIPAA compliance for AI tools requires a signed BAA plus technical safeguards. Without a BAA, sending PHI creates compliance exposure, even if the model appears to forget your inputs after the chat ends.
There is an important distinction here: HIPAA is not a product certification you can buy. There is no such thing as “HIPAA-certified” software. A vendor is a business associate only when it signs a BAA and agrees to the safeguards, breach notification, and permitted-use rules in the contract. If OpenAI has not signed one for the route you are using, you should not send PHI.
As of October 2026, a 2026 roundup of HIPAA-compliant AI tools lists nine vendors that will sign a BAA, which shows the market has options, but it does not make OpenAI’s direct API one of them by default. Pabau’s 2026 guide to HIPAA-compliant AI tools
If you are in Europe, remember that GDPR and the Swiss nFADP apply to personal data broadly, not just health data. Data residency in Europe helps with those regimes, but it is not a substitute for a HIPAA BAA when US PHI is involved.
Does ChatGPT Enterprise include a BAA?
Not by default, and not unless you have it in writing. ChatGPT Enterprise includes administrative controls, SSO, and domain verification, but those are security features, not a HIPAA BAA. OpenAI’s standard contracts do not include a BAA. Some enterprise negotiations may include one, but you cannot assume it exists.
This matters because clinicians and product teams often assume “enterprise” means “HIPAA-safe.” It does not. Before any clinician pastes a patient note into ChatGPT Enterprise, someone needs to read the contract and confirm the BAA exists, covers the specific workspace, and names OpenAI as the business associate. If the answer is “we’re still waiting on legal,” the answer is no.
Does the OpenAI API support a BAA for healthcare?
No. The OpenAI API does not come with a standard HIPAA BAA, and OpenAI’s developer terms are not a BAA. If you need OpenAI models under a BAA, the practical path is Azure OpenAI Service. Microsoft signs the BAA as the business associate, and you deploy OpenAI models inside your Azure subscription.
This is the route most healthcare teams actually use. You get the OpenAI models, but the contractual HIPAA layer sits between you and Microsoft, not between you and OpenAI directly. You still need to configure the service correctly—private endpoints, managed identity, and no logging to storage accounts you do not control—but the BAA exists.
Can clinicians use ChatGPT with PHI?
No. Not on standard ChatGPT Plus, Team, or Enterprise unless you have confirmed a signed BAA in writing. Clinicians should not paste patient names, MRNs, diagnosis codes, or clinical notes into the web UI. The risk is not just that OpenAI might train on the data; it is that there is no BAA, no guaranteed access log, and no business associate accountable for a breach.
If your organization wants clinicians to use AI with patient data, the setup should look like this: a vetted application, a BAA-covered backend, and a UI that never sends raw PHI to a model without a contract. Until that exists, the rule is simple: no PHI in the chat box.
What is zero data retention with OpenAI?
Zero data retention means OpenAI does not store your prompts or outputs beyond the life of the request. It is useful, but it is not HIPAA compliance. It removes long-term storage risk; it does not provide a BAA, access controls, breach notification obligations, or an audit trail.
You can have zero data retention and still violate HIPAA if you send PHI to a vendor that has not signed a BAA. Think of retention as one control among many. It is necessary for reducing data gravity, but it is not sufficient.
What does secure LLM inference look like for healthcare?
Secure inference means stateless processing, no payload logging, no training on your data, region-resident endpoints, and an audit trail on every call. For PHI, you need a gateway that inspects and routes each request before it reaches the model.
OpenAI offers data residency in Europe, which helps if your data must stay in the EU under GDPR or the Swiss nFADP. OpenAI’s data residency documentation explains how to keep processing in European regions. But residency is a geographic control, not a HIPAA contract.
Here is what I look for in a secure inference path:
| Capability | What it means for PHI |
|---|---|
| Stateless processing | No prompt or output is stored after the response returns |
| No payload logging | The gateway or provider does not write request bodies to disk |
| No training | Your data is excluded from model training explicitly |
| Region residency | Inference happens in the jurisdiction your law requires |
| Audit trail | Every call is logged with user, timestamp, and endpoint |
| BAA coverage | A signed contract names the vendor as a business associate |
What is actually covered by an OpenAI BAA?
A BAA covers the vendor’s handling of PHI, not your entire application. It defines permitted uses, breach notification, and safeguards. It does not cover your employees’ misuse, your frontend security, or your downstream database.
Scope is narrow and contractual. If you sign a BAA with Microsoft for Azure OpenAI, Microsoft is responsible for the platform layer. You remain responsible for your application code, who has access to the API keys, and what your users paste into the prompt. A BAA is not a force field.
Build a gateway or use Azure OpenAI directly?
Build a gateway when you need per-request PHI routing, custom redaction, and a single audit trail across multiple models. Use Azure OpenAI directly when you want the BAA with minimal integration and your workload fits its regions. Do not build a HIPAA wrapper just to avoid signing a BAA; the wrapper does not replace the contract.
If you are already on Azure and need OpenAI models, start with Azure OpenAI Service. If you need to route PHI to several models, enforce redaction rules per request, and keep an immutable audit trail, a gateway layer is worth the engineering effort. Learn how CloudSitus handles per-request PHI routing and audit trails.
FAQ
Does OpenAI sign a BAA for its API?
No. OpenAI’s public API terms do not include a standard HIPAA BAA. Use Azure OpenAI Service if you need Microsoft to sign the BAA while running OpenAI models.
Is ChatGPT Enterprise HIPAA compliant?
Not automatically. ChatGPT Enterprise has security features, but HIPAA requires a signed BAA. Confirm with OpenAI in writing before sending PHI.
Can I send PHI to OpenAI if I enable zero data retention?
No. Zero data retention reduces storage risk but does not replace a BAA. Without a BAA, sending PHI is a compliance issue regardless of retention settings.
Is OpenAI HIPAA compliant in Europe?
OpenAI offers data residency in Europe, which helps with GDPR and the Swiss nFADP, but residency is not a HIPAA BAA. For US PHI, you still need the BAA route, typically through Azure.
What is the difference between HIPAA compliance and a BAA?
HIPAA is the law; a BAA is the contract that lets a vendor handle PHI on your behalf. You need both the contract and the technical safeguards. Ask your counsel before relying on any vendor’s compliance claims.