Is ChatGPT HIPAA compliant? What to check first

Is ChatGPT HIPAA compliant? Not on the plans most people use. As of October 2026, OpenAI offers no business associate agreement (BAA) for ChatGPT Free, Plus, Pro or Business, so patient information should not go into them. OpenAI does sign BAAs for certain business offerings: ChatGPT Enterprise and Edu with the Regulated Workspace, ChatGPT for Healthcare, ChatGPT for Clinicians and, by request, the API. Even then, no software is HIPAA compliant on its own. Compliance depends on how your practice sets up and uses the tool.

Someone on your team has probably already pasted a patient detail into ChatGPT. Maybe it was a front desk note, a hard insurance letter, or a summary for a referral. This guide is for the practice manager or owner who now has to answer the question properly: is that allowed, and if not, what is?

Short answer: it depends on the plan and whether a BAA is in place

ChatGPT is not one product. It is a set of plans, and OpenAI treats them very differently. The plans most people use (Free, Plus, Pro and ChatGPT Business) do not come with a business associate agreement (BAA). OpenAI's HIPAA Implementation and Configuration Guide says that without a BAA with OpenAI, you may not use its services with protected health information (PHI). So for those plans the answer is simple: no patient information goes in.

OpenAI does sign BAAs for a short list of products, covered below. But a signed BAA is the starting line, not the finish. HIPAA does not certify software. Compliance comes from how your practice uses the tool: which settings are on, which staff can sign in, what they paste, and whether you can show later who saw what. A BAA with an unmanaged workspace and no staff rule still leaves you exposed.

Consumer ChatGPT vs Business, Enterprise and the API

Here is where each plan stands, taken from OpenAI's own pages. Plan names and terms change, so check the linked page before you rely on a row.

A few details matter for a small practice.

• Business and Teams. ChatGPT Team was renamed ChatGPT Business in August 2025. Paying for seats does not change the BAA answer: there is no BAA for it.
• Enterprise and Edu. OpenAI's BAA article says only Enterprise or Edu customers with a sales-managed account can get a BAA for ChatGPT. You arrange it with OpenAI's sales team; it is not a checkbox.
• ChatGPT for Clinicians. It is free for verified U.S. clinicians, and the eligible list currently names physicians (MD/DO), nurse practitioners, physician assistants and pharmacists. Dentists are not on that list. It is built for one person. For a BAA that covers several users with admin controls, OpenAI points to ChatGPT for Healthcare.
• The API. This is for software your practice or a vendor builds. You email OpenAI to request a BAA, and an enterprise agreement is not required. PHI can only go through the listed endpoints once your account has Modified Retention.
• Cost. OpenAI says ChatGPT for Healthcare pricing is based on ChatGPT Enterprise and depends on organization size, so the honest answer to "what does a HIPAA-eligible ChatGPT cost?" is a quote from OpenAI sales, not the price on the public plans page.
• Health in ChatGPT. The consumer Health feature lets people connect their own records. OpenAI's Health help page says it is not intended for covered-entity use and does not offer a BAA.

Even on an eligible plan, the BAA covers only listed features. OpenAI's HIPAA eligible products page says features not listed are not automatically covered, and PHI should not be entered when using them. Someone has to own those settings.

The common real-world risk is personal accounts. Your practice may hold a BAA for an approved workspace while a hygienist or biller uses a personal Plus account on their phone, because it is quicker. That chat sits outside your BAA, outside your admin controls and outside your logs. The plan you bought does not cover the account they actually used.

What counts as PHI, and de-identification

PHI is health information that can be tied to a person. In a practice that includes far more than a diagnosis: a name with an appointment time, a phone number on an insurance note, a photo of an X-ray with a date on it.

Data that is properly de-identified is no longer PHI. HHS's de-identification guidance describes two methods. One is a formal determination by a qualified expert. The other, Safe Harbor, means removing a list of identifiers (names, most geographic detail, all elements of dates except the year, phone numbers, account numbers, full-face photos and more) and having no actual knowledge that what remains could identify the person. HHS gives a useful warning: a record that still contains patient initials or the last four digits of a Social Security number does not meet Safe Harbor.

"I took the name out" is not de-identification. A rare condition, a visit date and a town can point to one patient. In a busy office, "mostly removed" is the realistic outcome, which is why a written rule works better than case-by-case judgment.

Using ChatGPT for clinical or progress notes: the risks

Drafting progress notes is the most common way clinicians want to use ChatGPT, and it is also the riskiest. A note is full of PHI by design. On a consumer or Business plan, pasting it in is a disclosure to a vendor with no BAA.

On an eligible plan, three more risks remain:

• Accuracy. The model can add details that were never said or drop ones that were. A clinician has to read and correct every draft before it goes in the chart.
• Too much information. HHS's minimum necessary standard asks covered entities to take reasonable steps to limit uses and disclosures of PHI to what the purpose needs. Pasting a whole chart to draft one paragraph rarely passes that test.
• Who can see it. OpenAI's HIPAA Guide tells customers to confirm that each new user is authorized to access PHI before inviting them, and to turn on workspace sharing only if everyone in the workspace may see the shared PHI. One shared login for the front desk undoes that.

If patient details were already pasted into ChatGPT

This is the situation many practices are actually in. Treat it as a privacy incident to look into, not something to quietly delete.

• Tell your privacy officer. Every covered entity has to name one. They own the next decision, not the person who pasted the note.
• Write down what happened. Which account and plan, roughly when, what kind of information, and how many patients. That record is what the assessment below is built on.
• Stop the habit, not just the chat. Deleting a conversation does not undo a disclosure. Find out why the person reached for ChatGPT and give them an approved way to do that task.

Under HHS's Breach Notification Rule, an impermissible use or disclosure of PHI is presumed to be a breach unless the covered entity shows, through a documented risk assessment, a low probability that the information was compromised. HHS lists the factors to weigh, including what information was involved, who received it and whether it was actually viewed. Whether notice is owed is a call for your privacy officer or counsel, made on the facts.

Then look for the rest of it. One pasted note usually means others, often on phones and personal accounts. Our guide to finding shadow AI covers where the evidence is.

A safer setup: policy, approved tools and a controlled layer

None of this means your team should stop using AI. It means deciding where it is allowed.

• Write the rule down. Name the approved tools and the plan for each, say plainly that patient details never go into personal accounts, and say who to ask. Our AI policy template gives you a first draft.
• Match the plan to the task. Marketing copy, job posts and general questions can live on an ordinary business plan with no patient data. Anything with PHI needs a plan covered by a signed BAA, or should not happen in a chatbot at all.
• Use individual accounts. Each person signs in with their own account through single sign-on, so access can be removed when someone leaves.
• Check the vendor, not just the plan. Our HIPAA and AI buyer checklist covers retention, training on inputs, audit logs and breach notice terms for any AI tool.

Many practices find that the work they want help with (insurance checks, intake, follow-up letters) is not a person typing into a chat box. It is a system doing a task. A controlled layer between the AI and your systems helps there: each task sees only the records it needs, a person approves anything that changes a record or reaches a patient, and every access is logged. That does not make any tool compliant, and it does not replace a BAA with each vendor. It makes your rules easier to keep. AI for healthcare teams shows how that work starts, and the healthcare and dental AI guides cover the rest of this series.

This guide is educational. It is not legal advice and does not replace your privacy officer or counsel.