The consumer versions of ChatGPT — Free, Plus, Pro, or self-serve Team or Business — are not HIPAA compliant, and you should never paste protected health information (PHI) into them. OpenAI signs a Business Associate Agreement (BAA) only for a few specific configurations: the API on Zero Data Retention endpoints, sales-managed ChatGPT Enterprise, and ChatGPT for Healthcare. A BAA is the starting line, not the finish line.
"Is ChatGPT HIPAA compliant?" is really three questions wearing one coat: which ChatGPT, doing what, under which agreement. The chatbot in your browser and an API call governed by a signed BAA with Zero Data Retention are not the same product, and treating them as the same is exactly how PHI ends up where it never should. This article separates the consumer app from the enterprise paths, explains what actually makes an AI feature HIPAA compliant, and shows how BeevR builds it.
A note before we begin: this is a general explanation, not legal or compliance advice. Your HIPAA obligations depend on your role, your data, and your contracts. Vendor terms change too — verify the current BAA terms directly with OpenAI, and have qualified counsel and a security professional review your specific configuration.
No — not the consumer version, and not at all without a signed BAA.
The ChatGPT that most people use — the Free, Plus, Pro, and self-serve Team and Business plans — is a consumer product. OpenAI does not offer a BAA for these plans, and under HIPAA you cannot send PHI to a vendor that has not signed one. Pasting a patient's name, diagnosis, dates, or record numbers into that chat box is an impermissible disclosure of PHI. It does not become acceptable because the tool is useful, because you were careful, or because you delete the conversation afterward.
There is a clean way to use a consumer chatbot for health-related work: strip out the PHI first. Under HIPAA's Safe Harbor method, once you remove the 18 specified identifiers — names, geographic detail smaller than a state, any date element more specific than a year, contact information, record and account numbers, device codes, full-face photos, and so on — and you have no actual knowledge that the remainder could identify a person, that data is no longer PHI. De-identified data falls outside the rules that require a BAA. It's worth noting that true de-identification is harder than deleting a few obvious fields, and free-text clinical notes are riddled with identifiers hidden inside ordinary sentences. If there is any doubt that the data is truly de-identified, treat it as PHI.
OpenAI now has HIPAA-eligible paths, but "eligible" does not mean "compliant out of the box."
As of 2026, OpenAI will sign a BAA for three things: the API, but only on Zero Data Retention (ZDR)-eligible endpoints; sales-managed ChatGPT Enterprise (the version negotiated through OpenAI's sales team, not the self-serve plans); and ChatGPT for Healthcare, an enterprise product launched in January 2026, built specifically for hospitals and clinicians, where the PHI you enter is not used to train the model. To start the process for the API, you email OpenAI's BAA team with your company information and use case; they typically respond within a few business days, while a sales-managed agreement can take longer to finalize.
There are two things to be careful about here. First, a HIPAA-eligible product is not automatically compliant — ChatGPT for Healthcare and the API still have to be configured correctly, with appropriate access controls and organizational safeguards, before any PHI touches them. The vendor gives you a compliant foundation; the compliant system is still something you have to build and operate yourself. Second, pay attention to the boundary of what the BAA covers. On the API, only ZDR-eligible endpoints are in scope; the standard endpoints that still retain data (historically about 30 days) are not appropriate for processing PHI. And features such as live web search are not HIPAA-eligible and not covered by the BAA, even when they remain available — route PHI through them and you have stepped outside the scope of the agreement. Also note: a separate consumer health experience called "ChatGPT Health" is governed by ordinary consumer terms, not HIPAA, and is not a clinical product backed by a BAA.
A signed BAA with OpenAI covers only one vendor in your chain. HIPAA compliance is a property of the whole system — everywhere PHI is stored, transmitted, logged, or processed — not of any single model. Here is what has to be right from end to end.
| Control | What it means in practice | Why it matters |
|---|---|---|
| A BAA with every subcontractor | A signed BAA with each vendor that touches PHI — the model provider, the cloud provider, logging and analytics tools, anything on the data's path | One unsigned link breaks the whole chain; PHI reaching a vendor without a BAA is an impermissible disclosure |
| Minimum necessary & de-identification | Send only the exact PHI the task truly needs; de-identify or tokenize anywhere the model does not need real identifiers | Narrows your exposure and audit scope; data that never left as PHI cannot leak as PHI |
| Encryption in transit and at rest | TLS for every call; strong encryption for stored prompts, outputs, and logs | A baseline expectation; PHI unencrypted in transit or at rest is a classic failure |
| Access controls | Role-based access, least privilege, individual identities, and MFA; no shared service account that can read raw PHI | Limits who and what can see PHI, and makes every access accountable |
| Audit logging | Tamper-resistant logs recording who accessed what, when, and which prompts and outputs went where | You cannot prove compliance — or investigate an incident — for activity you never recorded |
| Human in the loop | A qualified person reviews AI output before it affects a patient, a record, or a decision | The model can be wrong; for clinical or high-risk purposes, a human is the final check |
| Retention & deletion | Zero or very short retention at the model layer; a clear lifecycle for any prompts, outputs, and logs you do keep | PHI kept longer than necessary, in more places than necessary, is pure risk with no benefit |
The pattern running through the table above: HIPAA-compliant AI is mostly disciplined engineering, not a special "compliant model." The model is one component inside a system designed so that PHI is minimized, encrypted, access-controlled, logged, and never processed by a vendor that has not signed a BAA — with a human accountable for anything that matters.
The failures are rarely exotic. A handful of patterns account for most of the risk.
BeevR builds AI for heavily regulated industries the way the regulations assume you will: deterministic rules first, the model kept on a short leash, and a human accountable for anything that matters. We start by asking whether a given step needs a model at all — a clear rule or lookup is easier to audit than a probabilistic call, so PHI only flows through the model where it genuinely adds value.
The Kite framework we developed treats the model as an untrusted component by default. Its output is verified, constrained, and logged rather than trusted unconditionally, and a qualified person reviews anything that affects a patient, a record, or a decision. Around that core, we build the controls HIPAA actually expects: minimum necessary and de-identification so real identifiers reach the model only when needed, encryption in transit and at rest, role-based access with least privilege, and tamper-resistant audit logs recording who accessed what and which prompts and outputs went where. We map every place PHI lives or moves and make sure there is a BAA with each vendor on that path — the model provider, the cloud provider, every subcontractor — and we keep PHI away from the endpoints and features the agreement does not cover.
The result is an AI feature that is auditable by design, not patched together under deadline pressure. And because of how BeevR works, it's yours: fixed price, fixed timeline, and you own 100% of the codebase — source code, infrastructure configuration, and IP — with GitHub owner rights from the first commit. When an auditor asks how PHI moves through the system, you can show them, because it was built to be shown.
Is ChatGPT HIPAA compliant? The consumer versions — Free, Plus, Pro, and self-serve Team and Business — are not, and OpenAI will not sign a BAA for them. There are HIPAA-eligible paths: the API on Zero Data Retention endpoints, sales-managed ChatGPT Enterprise, and ChatGPT for Healthcare, each of which requires a signed BAA and correct configuration. Even then, "eligible" is not "compliant out of the box."
Can I use ChatGPT with patient data? Not in the consumer app, and not anywhere without a signed BAA. Putting PHI into consumer ChatGPT is an impermissible disclosure under HIPAA. You can use the chatbot with properly de-identified data, because de-identified data is no longer PHI — but true de-identification is harder than deleting a few obvious fields.
Does OpenAI sign a BAA? Yes, for a few specific products as of 2026: the API on ZDR-eligible endpoints, sales-managed ChatGPT Enterprise, and ChatGPT for Healthcare. For the API, you request a BAA by emailing OpenAI's BAA team with your company information and use case. OpenAI will not sign a BAA for the Free, Plus, Pro, Team, or self-serve Business plans.
What is Zero Data Retention and why does it matter for HIPAA? Zero Data Retention (ZDR) means OpenAI processes a request and then does not retain the prompt or output afterward. On the API, the BAA covers only ZDR-eligible endpoints; the standard endpoints that still retain data (historically about 30 days) are not appropriate for PHI. Some features, such as live web search, are not HIPAA-eligible even when ZDR-style handling is available.
Beyond a BAA, what makes an AI feature HIPAA compliant? A BAA covers one vendor; compliance is a property of the whole system. You need a BAA with every subcontractor on the PHI path, minimum-necessary data and de-identification, encryption in transit and at rest, role-based access controls, tamper-resistant audit logs, a clear retention and deletion policy, and a human in the loop for decisions that affect a patient.
Is ChatGPT for Healthcare automatically HIPAA compliant? No. It is a HIPAA-eligible product, launched in January 2026, built for clinical environments and with a BAA available to sign, but it still has to be configured correctly — appropriate access controls, organizational safeguards, and an actual signed BAA — before any PHI touches it. The vendor gives you a compliant foundation; the compliant system is something you have to build and operate yourself.
If you're building an AI feature that will touch PHI, the question is not whether the model is smart enough — it's whether the whole system can stand in front of an auditor. That means a BAA with every vendor on the path, PHI minimized and encrypted, access controlled and logged, and a human accountable for the decisions that matter.
That's how BeevR builds: deterministic rules first, the model treated as untrusted, a human in the loop, and compliance designed in from the first commit — fixed price, fixed timeline, and you own 100% of the codebase. If you're building software or AI for healthcare or another heavily regulated industry, tell us what you're building and book a consultation. You can reach us anytime at connect@beevr.ai.
This article is general information, not legal or compliance advice. Your HIPAA obligations depend on your role, your data, and your contracts, and vendor terms change — verify the current BAA terms with OpenAI and consult qualified counsel and a security professional before processing PHI.
Go deeper: beyond chat — HIPAA-compliant AI agents: architecture & compliance checklist.