The Direct Answer
ChatGPT itself is not HIPAA compliant. Two specific OpenAI tiers — ChatGPT Enterprise and ChatGPT Edu — can be configured to support HIPAA-regulated workloads under a Business Associate Agreement (BAA) signed with OpenAI. The free, Plus, and Team tiers cannot.
But that single answer is misleading on its own, and it is the answer most blog posts stop at.
The harder question — the one the compliance team is going to ask after legal signs the BAA — is where does Protected Health Information actually live, and who has technical control over it when an AI model processes it.
For most healthcare organizations evaluating AI in 2026, that question has a different answer than "we signed a BAA."
What HIPAA Actually Requires of an AI Vendor
The HIPAA Security Rule has three categories of safeguards: administrative, physical, and technical. Every category applies to the systems that touch Protected Health Information (PHI), and an AI inference endpoint absolutely touches PHI if PHI ever appears in a prompt.
The technical safeguards include access control, audit controls, integrity controls, transmission security, and authentication. The administrative safeguards include workforce training, incident response, and risk analysis. The physical safeguards include facility access controls and workstation security.
A vendor signing a BAA is acknowledging that it will meet these safeguards — but the BAA is the contractual layer. The technical implementation is what auditors test. And the technical implementation depends entirely on where the system runs.
When PHI flows from a hospital into a third-party AI inference endpoint, the auditor's question is no longer "did the vendor sign a BAA." It is "what is the chain of custody for that PHI, end to end, and can you produce evidence."
What OpenAI's BAA Actually Covers
OpenAI's BAA, as documented on their compliance page, applies to ChatGPT Enterprise, ChatGPT Edu, and certain API usage tiers. It covers PHI processed through the covered services and commits OpenAI to standard business-associate obligations under HIPAA.
It does not cover:
- ChatGPT Free or ChatGPT Plus.
- ChatGPT Team (the small-business tier).
- Any usage where the customer routes prompts through unsupported endpoints.
- PHI that ends up in the model's training data (it shouldn't, under the enterprise terms, but the BAA does not change what training data was used historically).
- Anything that happens after the response leaves OpenAI — what your downstream system does with it.
The BAA also doesn't change the underlying architecture: prompts and PHI still travel to OpenAI's infrastructure for inference. The covered tiers add contractual and technical protections. They do not bring the model to your data center.
The Question That Actually Matters
For most healthcare CISOs and compliance officers, the strict question of "is the vendor signed up for a BAA" is necessary but not sufficient. The deeper concern, in practice, is:
- Data residency. Where does PHI physically sit during inference, and is that location aligned with our data-residency commitments to patients and to state regulators?
- Audit trail. Can we produce a defensible audit log showing every prompt that touched PHI, every model that processed it, and every response it generated?
- Continuity. If OpenAI changes its terms, deprecates a model, or has an outage during a HIPAA audit, what is our fallback?
- Scope creep. When a clinician copies a patient note into a chat window on a personal account, the BAA does nothing. How do we govern actual usage at the workforce level?
These are infrastructure questions, not contract questions. And the answer most healthcare organizations end up at — once they have lived with a SaaS AI deployment for a few months — is some version of "we wish we owned this."
The Sovereign-AI Answer
The cleanest HIPAA posture for AI is the one where PHI never leaves the covered entity's perimeter.
That requires three things, none of which a SaaS chatbot can give you:
- An AI platform deployed inside your infrastructure. Not a managed SaaS that holds your data on the vendor's cloud — a platform that runs in your VPC, your data center, or your air-gapped environment, under your network and identity controls.
- A choice of model. Frontier models when the task allows it; open-weights models (Llama, Mistral, Qwen, Phi) running locally when the data is too sensitive to leave. The same governance layer routes both.
- Audit logging owned by you. Every prompt, every response, every model invocation logged into your SIEM, retained on your schedule, defensible under your HIPAA documentation.
This is the model the ibl.ai platform implements. The complete codebase ships to the covered entity. It runs on infrastructure the customer controls. PHI never leaves the customer's perimeter unless the customer explicitly routes a non-PHI workload to a frontier API — and even then, the platform sits between the user and the API, so the customer keeps the audit trail.
It also gives you the build-and-buy posture compliance teams ask for: you get a production-grade platform on day one (no green-field build), and you own the source code (no vendor lock-in if priorities change).
A Practical Decision Framework
For a healthcare organization that has already deployed ChatGPT Enterprise — or that is considering it — the right path is rarely "rip it out." It is usually:
- Keep ChatGPT Enterprise for clearly non-PHI workloads (general drafting, marketing, internal research) where the BAA is sufficient and convenience is high.
- Move PHI-touching workloads to a self-hosted platform (ours or someone else's) so the data never leaves your perimeter.
- Layer governance on top of both so the audit trail is unified and the workforce policy is enforceable regardless of which model handled a given prompt.
This is the posture most large healthcare systems we work with end up in. It is also the posture that survives the next compliance audit, the next vendor pricing change, and the next time OpenAI deprecates a model your clinical workflow depended on.
What to Take Away
- ChatGPT is not, by default, HIPAA compliant.
- ChatGPT Enterprise and ChatGPT Edu can be configured for HIPAA-regulated workloads under a BAA with OpenAI.
- A BAA is a necessary contractual layer, not a complete technical answer.
- For real HIPAA defensibility, PHI should not leave the covered entity's perimeter — which means a self-hosted AI platform, with your choice of model, audit logs you control, and the source code in your possession.
If you want to see how this looks in practice, see how ibl.ai handles self-hosted and private LLM deployments, and how the Healthcare solution stack is structured around HIPAA, data residency, and the BAA chain. The ChatGPT Enterprise alternative page is the side-by-side for procurement.