No. The free and personal versions of ChatGPT, Claude and the other big AI apps are not HIPAA compliant, and you shouldn't put patient information in them. A business version can be used with patient data, but only when the vendor has signed a business associate agreement with you and only for the features that agreement names. Law offices and tax preparers face a parallel set of rules, written by the bar and the IRS instead of HHS. None of this is legal advice. It's a map of where the rules are published, so you know what to ask your own lawyer or compliance adviser.
Is ChatGPT HIPAA compliant?
A business associate agreement, or BAA, is a contract in which a vendor that handles protected health information on your behalf promises to safeguard it the way HIPAA requires. Protected health information is anything that ties a person to their health or care, from a chart note to an appointment reminder with a name on it. The Department of Health and Human Services says in its guidance on HIPAA and cloud computing that a cloud provider storing or processing patient data for a practice is a business associate, even when the provider holds only encrypted data and has no key to read it. The HHS sample BAA provisions lay out what the contract has to contain. If there's no signed BAA, the patient data shouldn't go in.
The AI companies offer BAAs, but only on specific products. OpenAI's help article on business associate agreements says a BAA for ChatGPT is available only to sales-managed Enterprise and Edu accounts, not the Business plan a small office would usually buy. Its HIPAA guide, which forms part of the agreement, adds that early-access features don't support HIPAA use. API customers, meaning the programmers who build OpenAI's models into their own software, request coverage separately.
Anthropic's page on what its BAA covers is just as particular. Claude Enterprise chat and projects are covered once the account owner turns on the HIPAA setting and accepts the agreement. Connectors, the mobile app, several beta products and some API features are not. That catches people. A connector is the piece that lets the model read another system, and Anthropic says plainly that its BAA covers none of them, even after an owner switches them on. So a dental office can be on a covered plan and still send patient data outside the agreement the day someone links the scheduling software.
Those enterprise plans are sold to organizations larger than most single-location practices. If the vendor won't sign with an office your size, the remaining routes are a different vendor that will, or a model that runs on your own hardware. Whether that second route makes sense for you is the subject of whether you need private AI at all.
Does a firm with health-care clients count as a business associate?
It can, and this is the part offices outside health care tend to miss. HIPAA follows the data, not the type of business. A law firm defending a clinic in a malpractice claim, an accounting firm doing a medical practice's billing review, or an IT shop maintaining a pharmacy's server may each handle patient records on that client's behalf. When they do, HHS treats them as business associates, and the client should be asking them to sign a BAA.
Data passes from the patient to your office to the AI vendor and its subprocessors, and a signed business associate agreement covers each hop
Patient or client
Your office
a covered entity, or a firm with a duty of confidence
The AI vendor
a business associate, with a signed BAA before any patient data
The vendor's subprocessors
bound by the same terms
Source: HHS, 45 CFR 164.502(e) and 164.504(e)
That has a knock-on effect for AI. If your firm is a business associate and you paste a patient's records into an AI tool, the AI vendor becomes your subcontractor under HIPAA, and you need its BAA in turn. A firm that has never thought of itself as a HIPAA shop can end up needing an Enterprise contract for one client's file.
What do lawyers need to know about AI and client confidentiality?
For most client files, lawyers answer to confidentiality rules rather than HIPAA. Two opinions set the terms for a North Carolina firm.
The American Bar Association's Formal Opinion 512, issued July 29, 2024, is the national reference. Its central worry is "self-learning" tools, meaning AI that can keep what you type and use it later, possibly in an answer to someone else. For those tools, the ABA concluded that a lawyer needs the client's informed consent before putting information about the representation into them, and that a line of boilerplate in an engagement letter doesn't count as informed. ABA opinions guide the states without binding them.
The North Carolina State Bar went further toward practical rules in 2024 Formal Ethics Opinion 1, adopted November 1, 2024. It allows lawyers to use AI with client information when the lawyer is satisfied the program is secure enough to meet the confidentiality rule, and it advises against putting client-specific information into publicly available tools that may train on what users type. It also asks lawyers to vet AI vendors the way they would any technology vendor, looking at the terms of service, who keeps the data, and whether the firm can get it back or destroy it.
The part of the NC opinion I'd read twice says that running AI on the firm's own servers doesn't reduce the lawyer's duty, because an in-house system can be just as vulnerable as a vendor's. It recommends bringing in an IT or security expert before storing client data locally. A two-lawyer firm in Boone that sets up a model on an unpatched desktop with a shared password has arguably done worse than a firm that signed a careful vendor agreement.
| Field | The rule | Where it's published |
|---|---|---|
| Health | Business associate agreement before sharing patient data | HHS, 45 CFR 164.502(e) and 164.504(e) |
| Law, national | ABA Formal Opinion 512 | July 29, 2024 |
| Law, North Carolina | NC State Bar 2024 Formal Ethics Opinion 1 | Adopted Nov. 1, 2024 |
What about accountants and tax preparers?
Tax preparers have a federal rule of their own. Section 7216 of the Internal Revenue Code makes it a crime for a preparer to knowingly or recklessly disclose or use a client's return information for anything other than preparing the return. The penalty is a fine of up to $1,000, a year in prison, or both. As The Tax Adviser's overview of Sec. 7216 explains, disclosure covers making the information known to anyone in any manner, consent has to come before the disclosure, and individual filers must opt in affirmatively.
The law predates AI, and as of October 2026 I could find no IRS ruling that says whether pasting a return into an AI tool counts as preparation or as disclosure to a third party. Practitioners reading it cautiously get written consent from the client first, or strip the names, Social Security numbers and figures that identify anyone before the data goes in. A bookkeeping office in Watauga County doing a few hundred returns a season would want that consent language settled well before February.
The same reasoning carries to other professional services with duties of confidence. Each field publishes its rule somewhere, and the vendor publishes which of its products the rule can live on. The gap between those two documents is where an office gets into trouble, and it moves every time a vendor ships a new feature. Where AI fits in an office's work more broadly is in the guide to how to use AI in your business.