Short answer: no — not the ChatGPT most practices are actually using. The free, Plus, Team, and Business tiers cannot be made HIPAA compliant, because OpenAI will not sign a Business Associate Agreement for them. Putting patient information into those tiers is a HIPAA violation. But that is not the whole story, and the whole story matters if you are making a decision for your practice.
There are now several ChatGPT products, and they sit on opposite sides of the compliance line. Some of them can be used with protected health information under the right contract. Most of the ones people already have open in a browser tab cannot. Below is what actually applies as of August 2026, what a BAA does and does not buy you, and how to decide between cloud AI and keeping it in your building.
First, the part that trips everyone up: PHI and BAAs
Two definitions do most of the work here.
Protected health information (PHI) is health information tied to an identifiable person. A patient name, a date of birth, an account number, a scan, a chart note — and combinations that identify someone even without a name attached. If you paste a chart note into a chatbot to "clean up the wording," you have disclosed PHI.
A Business Associate Agreement (BAA) is the contract HIPAA requires before you hand PHI to an outside vendor who processes it on your behalf. No BAA means no PHI. It is that binary. A vendor being secure, or encrypted, or well-intentioned, does not substitute for the contract.
So the real question is never "is this AI tool safe." It is "will this vendor sign a BAA for this specific product, and did we actually sign it."
Which ChatGPT products can handle PHI
OpenAI has split its lineup considerably, and the tier you are on decides everything.
Cannot be used with PHI
- ChatGPT Free, Plus, Team, and Business — OpenAI does not offer a BAA for these. This is the important one, because it is what most small practices are actually using.
- ChatGPT Health — the consumer health product. OpenAI has been explicit that it will not sign a BAA for it under any circumstances, because it exists for personal health literacy, not regulated healthcare operations. It is for patients, not practices.
Can be used with PHI, under contract
- The OpenAI API platform — a BAA is available. You email
[email protected]with your company and use case, and OpenAI reviews case by case, typically responding within one to two business days. An enterprise agreement is not required. Most requests are approved. A few API services are excluded from BAA coverage, so the scope matters. - ChatGPT Enterprise and Edu — a BAA is available, but only for sales-managed accounts. You have to go through OpenAI sales; you cannot self-serve your way into it.
- ChatGPT for Healthcare — launched January 2026 for hospitals and health systems, built on the Enterprise compliance stack with a BAA included.
- ChatGPT for Clinicians — aimed at verified individual clinicians, with its own separate in-product BAA flow.
Notice how much of that changed inside a single year. Two of those products did not exist in 2025. That pace of change is itself a planning problem, and we will come back to it.
What a BAA actually buys you — and what it doesn't
This is where practices get a false sense of safety. A signed BAA does not make you compliant. It makes one vendor relationship permissible. Everything else is still your responsibility:
- Access controls — who in your office can use the tool, with individual accounts, not a shared login
- Audit logging — a record of who accessed what, retained and reviewable
- Workforce training — documented, and repeated when your policies change
- A security risk analysis that actually includes the AI tool
- Policies and procedures governing what staff may and may not put into it
- Breach notification procedures if something does go wrong
Put plainly: the BAA covers the vendor's handling of your data. It does not cover your handling of your data. A practice with a signed BAA, a shared login, no audit trail, and no staff training is not compliant. It just has better paperwork than the practice next door.
The mistake we see most often
It is almost never a dramatic breach. It is a front desk employee with a personal ChatGPT Plus subscription pasting in a patient's insurance denial letter to draft an appeal, because it saves twenty minutes and nobody ever told them not to.
That is an impermissible disclosure. The practice usually has no idea it happened, because it happened in a browser tab on a personal account that generates no logs the practice can see. There is no vendor to subpoena, no audit trail, and no way to scope what left the building.
If you take one thing from this article: find out what your staff are already doing before you decide what tool to buy. Ask directly, without blame — you will almost certainly find some AI use already happening. That conversation is free, and it is more urgent than any purchasing decision.
The option people forget: de-identification
De-identified health information is not PHI, and it is not subject to the BAA requirement. If information has been de-identified using a method the HIPAA Privacy Rule permits, you can use ordinary AI tools with it.
This is genuinely useful for things like drafting policy language, generating patient education material, or summarizing clinical literature. It is also easy to get wrong — proper de-identification is a defined standard, not just deleting the name at the top. Rare conditions, precise dates, and small-population details can re-identify someone. If you go this route, do it deliberately, train staff on the method, and document it.
When cloud AI is the right answer
We build private on-premise AI systems, so you would expect us to tell you cloud AI is a trap. It isn't, and we would rather be useful than persuasive. Cloud AI with a proper BAA is often the better call when:
- You are a large organization with real IT and compliance staff who can manage vendor risk properly
- You need the absolute frontier of model capability and are willing to pay for it
- Your use case is occasional rather than constant, so per-seat pricing stays cheap
- You want somebody else responsible for uptime, patching, and hardware failure
If that describes you, get the right tier, sign the BAA, configure it properly, and you will be fine. You do not need a server in your closet.
When keeping it in your building makes more sense
The argument for on-premise AI is not that cloud AI is illegal. It is a different set of tradeoffs:
- Data never leaves. A BAA reallocates liability; it does not change physics. If your vendor is breached, the PHI was still at your vendor, and you still own the notification obligation to your patients. When the model runs on hardware in your office, there is no third party to be breached.
- Cost stops scaling with headcount. Per-seat AI pricing is a permanent operating expense that grows as you hire. A server is a one-time capital cost the whole team shares.
- Product terms stop moving under you. The compliance landscape above changed twice in one year. Hardware you own does not get re-tiered, deprecated, or repriced.
- It works offline. Internet outage, and your AI still runs.
- The vendor risk conversation gets short. "Nothing leaves the building" is a simpler thing to document, explain to a compliance advisor, and defend in an audit than a matrix of vendor contracts.
The honest tradeoffs on this side: you own the hardware, someone has to maintain it, and open models — while genuinely capable now — are not always the sharpest option on the hardest reasoning tasks. For document Q&A, drafting, summarizing, and the daily grind of a practice, the gap is small and shrinking.
None of this removes your obligations. On-premise AI is a strong technical foundation for compliance. Your policies, training, access controls, and risk analysis still complete the picture, and you should still run the decision past your compliance advisor.
What to do this week
- Ask your staff what they are already using. No blame — you need the truth, not a clean answer.
- Check your tier. If anyone is on Free, Plus, Team, or Business and touching patient information, stop that today.
- Decide your path — a BAA-eligible cloud product, strict de-identification, or on-premise. Pick one deliberately instead of letting it happen by default.
- Write the policy down and train on it. Undocumented rules are not controls.
- Add AI tools to your security risk analysis. If they are not in it, it is out of date.
Want a straight answer for your practice?
We build private AI servers that run entirely inside your office — the model, your documents, all of it on hardware you own, with nothing leaving your network. We also tell people when that is the wrong fit, because a bad install helps nobody.
The consultation is free, and it starts with what your team actually does all day rather than what we would like to sell you. If cloud AI with a BAA is the better answer for your situation, we will say so.
See our private AI services Book a free consultation
Not legal advice. WebPCDesigns builds and supports technology; we are not attorneys and we are not a compliance consultancy. This article is general information current as of August 2026, and AI vendor terms have been changing quickly — verify current terms directly with the vendor. Decisions about HIPAA compliance in your practice should be made with your compliance officer or healthcare attorney.
Sources: OpenAI Help Center — BAA for API Services; The HIPAA Journal — Is ChatGPT HIPAA Compliant?; OpenAI for Healthcare.