Is ChatGPT HIPAA compliant? What a medical or dental practice’s AI policy needs to say

By Dan Blomquist, Aegis AI Advisory · Published November 3, 2026 · Vendor terms checked November 3, 2026

Not by default. The free, Plus and Business versions of ChatGPT do not come with a HIPAA business associate agreement (BAA), so patient information should not go into them. OpenAI offers a BAA only for specific products, such as ChatGPT for Healthcare and the API with Modified Retention, and even then the practice is still responsible for how the tool is set up and used. HHS puts it plainly: the Office for Civil Rights “does not endorse, certify, or recommend specific technology or products.” A practice’s AI policy is where you decide which tools are approved for PHI and which are not.

The question that matters is not whether a tool is “HIPAA compliant” but whether your use of it is. That comes down to a signed BAA for the exact product and features your staff use, a risk review, and rules your front desk can follow. This page covers what HHS guidance requires, where the major AI tools stand today, the risks that are specific to a clinic or dental office, and sample policy wording. The broader basics of any staff AI policy are in our explainer on what an AI acceptable-use policy should contain.

Why an AI tool needs a BAA before it sees PHI

HHS’s guidance on HIPAA and cloud computing was written for cloud vendors, and a hosted AI tool is one. Three points from it decide most AI questions:

The vendor is a business associate. When a covered entity uses a service provider to create, receive, maintain or transmit electronic PHI on its behalf, the provider is a business associate, even if it only stores encrypted data it cannot read.
No BAA means a violation. HHS says a covered entity that uses a cloud provider to process or store ePHI without first signing a BAA is in violation of the HIPAA Rules (45 CFR 164.308(b)(1) and 164.502(e)).
De-identified data is the exception. A provider that receives only data de-identified under the Privacy Rule is not a business associate. But de-identification means removing 18 identifier types under the Safe Harbor method, including dates tied to the patient (other than the year), phone numbers, email addresses, record numbers and full-face photos, wherever they appear, including free-text notes. HHS says initials or the last four digits of a Social Security number do not qualify. Deleting the name from a patient message does not make it de-identified.

A BAA is a specific contract. HHS’s sample business associate provisions list what it must do, including limiting the vendor’s uses of PHI to what the contract allows, requiring Security Rule safeguards, reporting breaches, and returning or destroying PHI when the contract ends. A vendor’s “HIPAA-ready” marketing page is not a BAA.

Where common AI tools stand (checked November 3, 2026)

BAA coverage is granted product by product, and sometimes feature by feature. This is what each vendor’s own documentation says today. Plans and coverage lists change, so re-check before you approve a tool.

Tool or plan	BAA available?	What to watch
ChatGPT Free, Plus or Business	No. OpenAI says it does not offer a BAA for ChatGPT Business, and the consumer plans are not on its HIPAA-eligible list.	The tool most staff already use on their phones. Treat as no PHI, ever.
ChatGPT for Clinicians	Yes, for the individual verified clinician (physician, NP, PA or pharmacist) through an in-product BAA, per OpenAI.	Covers that one clinician’s account, not the office, and OpenAI says not to enter PHI unless you are authorized to sign the BAA. Dentists are not on the current eligible list. OpenAI points multi-user practices to ChatGPT for Healthcare.
ChatGPT for Healthcare, or Enterprise with Regulated Workspace	Yes, through OpenAI sales.	OpenAI lists features not covered, such as improved memory and some agent and browsing options. Keep them off for PHI users.
OpenAI API	Yes, with an accepted BAA and Modified Retention, on listed endpoints (OpenAI).	Relevant if a vendor builds on the API for you; ask which agreement covers your data.
Microsoft Copilot and Copilot Chat (Microsoft 365 commercial)	Yes. Microsoft lists both among the services in scope for its HIPAA BAA.	Microsoft says web search queries are not covered by the BAA. Consumer Copilot is a different product.
Gemini in Google Workspace	Yes, once an admin accepts Google’s BAA; Gemini features such as Help me write in Gmail are part of the HIPAA Included Functionality.	Google says third-party add-ons are not covered, and customers without a BAA must not use PHI in Workspace at all.

Summaries of vendor documentation, not an endorsement. HHS does not certify AI products, and a BAA does not by itself make a use compliant; both OpenAI and Microsoft say so in the pages linked above.

Risks that are specific to a practice

A general office AI policy misses the places PHI actually leaks in a clinic or dental office:

The front desk and billing. Staff paste a patient portal message, an insurance denial or a collections note into a chatbot to draft a reply. This is the most common gap, and it usually happens on a personal account.
AI scribes and ambient listening. The vendor records or transcribes visits, so it is a business associate. The policy needs a BAA before go-live, a rule on whether the vendor may train on your data, clinician sign-off on every note, and a decision on how patients are told.
Clinical content. AI-drafted patient instructions, referral letters or chart summaries can be wrong. Someone licensed has to review anything clinical before it reaches a chart or a patient.
Decision-support tools. Practices covered by Section 1557 of the Affordable Care Act (generally, those that receive federal financial assistance from HHS) must not discriminate through patient care decision support tools. Under 45 CFR 92.210, they must make reasonable efforts to identify tools whose inputs measure race, color, national origin, sex, age or disability, and to mitigate the risk of discrimination from those tools. Note that Section 1557 “covered entity” is not the same definition as HIPAA’s.
Website chatbots and AI receptionists. A patient who types symptoms into a chat widget or AI phone line is handing PHI to that vendor. Our look at whether an AI receptionist is worth it covers the operational side; for a practice, the vendor also needs a BAA.
More PHI than the task needs. The Privacy Rule’s minimum necessary standard requires covered entities to limit uses and disclosures of PHI to what the purpose requires (with exceptions, such as disclosures to a provider for treatment). Even in an approved tool, staff should paste the paragraph that matters, not the whole chart.

Before any patient information goes into an AI tool

Build this check into the policy. If any answer is no, the PHI stays out.

Is the tool on the practice’s approved list for PHI? Personal accounts never are.
Is there a signed BAA for this exact product and plan? “Microsoft” or “OpenAI” is not enough; the agreement has to cover the plan your staff sign in with.
Is the feature you are about to use covered? Web search, memory, browser agents and third-party add-ons are often carved out.
Is this the minimum PHI needed for the task?
Will a licensed person review the output before it is used for care, added to a chart or sent to a patient?

What goes in a medical or dental practice AI policy

Scope and owner: who the policy covers (clinicians, front desk, billing, contractors) and who approves new tools, usually the HIPAA privacy or security officer.
Approved tools list with the plan, the BAA status, and the covered features for each.
PHI rule: no PHI in any tool without a BAA; minimum necessary in approved tools; no personal accounts.
Clinical review: licensed review of all AI-assisted clinical content, and AI-drafted notes signed only after the clinician has read them.
AI scribes and patient-facing tools: how patients are informed, and how to stop recording on request.
Decision-support tools: an inventory and the Section 1557 review, if it applies to you.
Risk analysis link: every new AI tool goes through the Security Rule risk analysis before use.
Training and sign-off for all workforce members, repeated when the tools list changes.
Incidents: what to do if PHI goes into an unapproved tool, and who decides whether it is a reportable breach.

The general sections (purpose, prohibited uses, annual review) follow the same pattern as any professional firm’s policy; see the free sample policy. If you share patients’ billing with an outside accountant, the CPA firm version shows what their side looks like.

Sample clauses for a practice

Excerpt for illustration only. Not legal advice. [Practice] and [Privacy Officer] are placeholders.

4. Protected health information

Workforce members must not enter protected health information into any AI tool unless the tool, plan and feature are listed on [Practice]’s Approved AI Tools list as covered by a signed business associate agreement. Removing a patient’s name does not make information safe to enter. Personal or free AI accounts may never be used for practice work involving patients.

5. Minimum necessary

When an approved tool is used with PHI, enter only the information needed for the task. Do not upload full charts, images or schedules when a single note or paragraph will do.

6. Clinical review

AI output is a draft. A licensed clinician must review any AI-assisted clinical content, including visit notes, patient instructions, referral letters and summaries, before it is added to the record or shared with a patient.

9. New tools

No AI tool may be used with practice data until [Privacy Officer] has confirmed the business associate agreement, reviewed the vendor’s data use and retention terms, and added the tool to the Security Rule risk analysis.

Notes for Michigan practices

Michigan’s breach notification law, MCL 445.72, generally requires notice to affected residents when unencrypted personal information is accessed and acquired without authorization. Subsection (10) says a person that is subject to and complies with HIPAA and its regulations is considered in compliance with that section, so for a HIPAA-covered practice the HIPAA breach notification process is the one to build around, and your AI incident rule should feed into it. Many small practices, from metro Detroit to the Upper Peninsula, run on Microsoft 365 or Google Workspace with an outside IT provider; check in your admin console that the HIPAA BAA has actually been accepted before anyone turns on Copilot or Gemini for patient work.

Start this week

Ask every clinician and staff member which AI tools they use for work, with no penalty for honest answers. Our guide to shadow AI explains why this comes first.
For each tool, find the plan and check for a signed BAA. If there isn’t one, the tool is no-PHI until there is.
Add AI tools to your Security Rule risk analysis. HHS’s free Security Risk Assessment Tool is meant to help small and medium-sized practices, and HHS’s risk analysis guidance says new technology should be assessed when it is being planned.
Adopt a short written policy and get signed acknowledgments. Our 90-day AI governance plan shows one way to sequence the rest.

What isn’t settled

Vendors change which plans and features their BAAs cover, sometimes month to month, so the table above is a snapshot, not a standing answer. In January 2025 HHS proposed changes to the HIPAA Security Rule that would, among other things, require a written technology asset inventory and a map of how ePHI moves through your systems; check whether a final rule has been issued before you rely on the current text. Parts of the 2024 Section 1557 rule have been challenged in court, though § 92.210 remains in the current eCFR text. State privacy and recording-consent laws add their own requirements. And whether an AI-drafted note meets your payer’s and licensing board’s documentation standards is a question HIPAA does not answer. A policy makes these decisions explicit; it does not replace your privacy officer or a health care attorney.

Get a policy drafted for your practice

Answer 10 questions, choose Medical / dental, and the generator emails you an AI acceptable-use policy that addresses the HIPAA Privacy and Security Rules, no PHI without a signed BAA, the minimum necessary standard and clinician review of AI clinical content, using the tools and state you enter.

$49 Policy: AI acceptable-use policy (Word + web version) and a one-page staff checklist.
$149 Pro Pack: adds an AI vendor risk register and a staff training sign-off sheet.

Build my practice’s AI policy →

Not ready? Read the free sample policy first, or fill in our $9 DIY Starter Kit.

Disclosure: Aegis AI Advisory sells this generator. It produces a template, not legal advice, and it does not sign or review business associate agreements. It is not the right fit if your privacy officer or health care counsel has already written an AI policy, or if you are a hospital or health system that needs a full AI governance program. For a practice that wants an AI usage inventory, vendor review and staff training, see the AI governance track or book a free AI Operations Briefing.

Frequently asked questions

Is ChatGPT HIPAA compliant?

Not by default, and no tool is HIPAA compliant on its own. HHS says a vendor that creates, receives, maintains or transmits electronic PHI for a practice is a business associate, so the practice needs a business associate agreement (BAA) with it before any PHI goes in. OpenAI offers a BAA only for specific products: ChatGPT for Healthcare, ChatGPT Enterprise with Regulated Workspace, ChatGPT for Clinicians and the API with Modified Retention, among others. It says it does not offer a BAA for ChatGPT Business, and the free and Plus plans are not on its HIPAA-eligible list. Even with a BAA, OpenAI says accepting it does not by itself make your use HIPAA compliant; you still have to configure the tool and run your own safeguards.

Can my front desk paste a patient message into ChatGPT if they remove the name?

Usually not safely. Removing the name does not make information de-identified under HIPAA. The Safe Harbor method requires removing 18 kinds of identifiers, including dates (other than year) tied to the patient, phone numbers, email addresses, medical record numbers and full-face photos, and HHS says initials or the last four digits of a Social Security number are not allowed either. Identifiers also have to come out of free-text notes. A practical policy rule is simpler: no patient information in any AI tool that is not on the practice's approved list with a signed BAA.

Does a free ChatGPT for Clinicians account cover my whole office?

No. OpenAI describes ChatGPT for Clinicians as a free product for verified U.S. clinicians (physicians, nurse practitioners, physician assistants and pharmacists) with an in-product BAA for that individual's account. OpenAI says a BAA covering multiple users, central admin controls or an organization-wide deployment requires ChatGPT for Healthcare. OpenAI also says not to enter PHI unless you are authorized to sign the BAA for your account, which an employed clinician may not be. Dentists are not among the eligible clinician types OpenAI currently lists, and your schedulers, billers and hygienists are not covered by a physician's individual BAA.

Are Microsoft Copilot and Google Gemini HIPAA compliant?

They can be used with PHI under the right business plan and agreement. Microsoft lists Microsoft Copilot and Copilot Chat among the commercial Microsoft 365 services in scope for its HIPAA BAA, but its Copilot documentation says web search queries are not covered by the BAA. Google says Workspace customers must accept its HIPAA BAA before using PHI, that Gemini features such as Help me write in Gmail are part of the HIPAA Included Functionality, and that third-party add-ons are not covered. Consumer Copilot and personal Gmail accounts are a different matter; check the specific plan your staff signs in with.

Is an AI scribe or ambient listening tool a business associate?

If the vendor records, transcribes or stores visit audio or notes for your practice, it is creating, receiving or maintaining PHI on your behalf, which is the HHS definition of a business associate. You need a BAA before go-live, and the BAA should say whether the vendor may use your data to train or improve its models and what happens to recordings when the contract ends. Your policy should also require the clinician to review and sign every AI-drafted note, and decide how patients are told the tool is in use; recording-consent rules vary by state.

Do we need a separate AI policy if we already have HIPAA policies?

You need at least an AI section. The Security Rule already requires a risk analysis, and HHS guidance says new technology should be analyzed for risk when it is being planned, not after. Existing HIPAA policies rarely name which AI tools are approved, which features are outside the BAA, or who checks AI output before it reaches a chart or a patient. A short AI policy that plugs into your existing HIPAA program fills that gap.

What does the Aegis AI policy generator produce for a medical or dental practice?

Choose Medical / dental as your industry and answer 10 questions. The $49 Policy is an AI acceptable-use policy in an editable Word file and a web version, plus a one-page staff checklist. For medical and dental practices it is written to address the HIPAA Privacy and Security Rules, a rule that PHI never goes into a tool without a signed BAA, the minimum necessary standard, and clinician review of any AI clinical content, using the tools and state you enter. The $149 Pro Pack adds an AI vendor risk register and a staff training sign-off sheet. It is a template, not legal advice.

Sources

Also free: check whether ChatGPT and Google’s AI can find your practice → · More guides on the resources page.