Can therapists use ChatGPT for notes? The privacy rules from the clinician's side
The Morse teamPublished 5 min read
Therapists can use ChatGPT for notes only if no client can be identified from what they type, or if their practice has a Business Associate Agreement (BAA) with the AI company that covers the exact product they use. As of October 2026, HIPAA lets a covered practice hand client information to a vendor only with written assurances, and its de-identification standard requires stripping 18 kinds of identifiers, including those of a client's relatives, employers, and household members.
This isn't legal advice; it covers privacy while you draft. We read the HIPAA regulation text on the government's eCFR site (45 CFR 160.103, 164.501, 164.502, 164.504, 164.508, and 164.514), the American Counseling Association's recommendations for practicing counselors on AI, Anthropic's BAA page, and Google Workspace's HIPAA page on October 23, 2026. HHS's own guidance pages and the American Psychological Association's AI guidance didn't load for us that day, so we quote the regulation itself.
Is it a HIPAA violation to use ChatGPT for therapy notes?
HIPAA reaches a practice that bills electronically, and it treats an AI company that receives identifiable session content as handling it on the practice's behalf, which we explain from the patient's side in private AI for health questions. For a clinician, HIPAA's contract rule comes first. It requires "satisfactory assurance" that the vendor will safeguard client information (eCFR, 45 CFR 164.502), documented "through a written contract," and that contract must bind the vendor to "Not use or further disclose the information other than as permitted or required by the contract or as required by law" and to report breaches (eCFR, 45 CFR 164.504). A consumer chatbot's terms of use are not that contract. A cash-only practice may sit outside HIPAA, but state confidentiality law and your licensing board's ethics code still apply, so ask your board or a health-law attorney before you treat that as a green light.
Does the plan you use come with a BAA? Companies that offer one tie it to specific business products. Anthropic says its BAA covers "HIPAA-ready services, such as use of our first-party API, Claude in Microsoft Foundry, or Enterprise plans," that a standard Enterprise plan isn't covered until its Primary Owner turns on HIPAA compliance, and that "the BAA only covers the single organization that accepted it" (Anthropic). Claude Free, Pro, and Max are handled in a separate consumer article. Google says Workspace customers without a BAA "must not use PHI in Google Workspace or Cloud Identity services," and that a super administrator accepts the BAA in the Admin console (Google). Check the BAA's list of covered features for the AI tool by name, not the vendor's brand.
What counts as de-identified for a therapy note?
The bar is higher than deleting the name. HIPAA's Safe Harbor method requires removing 18 identifiers "of the individual or of relatives, employers, or household members of the individual," and the practice must not have "actual knowledge" that what's left could identify the client (eCFR, 45 CFR 164.514). The other route, Expert Determination, needs a qualified expert to find "the risk is very small" and document the analysis, which few solo practices will commission for a progress note.
Therapy notes are hard to de-identify because they're about relationships. A note names a partner, a child's school, a boss, a hometown, and the date things went wrong. Here is a typical note, element by element, against the Safe Harbor list as read on October 23, 2026.
| What's in the draft | Safe Harbor item it falls under | What to type instead |
|---|---|---|
| The client's name or initials | (A) Names | "The client" |
| A partner's, child's, or parent's name | (A) Names, which covers relatives and household members | "Their partner", "their child" |
| The employer or school | (A) Names, which covers employers; (B) for a school's location | "Their workplace", "school" |
| Town, neighborhood, or street | (B) Every geographic unit smaller than a state | Leave it out |
| Session date, birth date, admission or discharge date | (C) "All elements of dates (except year)" | The year only, or no date |
| An age of 90 or over | (C) Ages over 89 | "90 or older" |
| Phone, email, or a portal link in a letter | (D), (F), and (N) | A placeholder like [PHONE] |
| Insurance member ID or record number | (H) and (I) | Leave it out |
| An audio recording of the session | (P) Biometric identifiers, "including finger and voice prints" | A written summary you've de-identified |
| A detail only this client has, like a rare job or a news event they were in | (R) "Any other unique identifying number, characteristic, or code" | A general description, or leave it out |
Can AI draft letters and treatment ideas without client data?
Yes, if you split the task. A letter to a school, an employer, or an insurer exists to identify the client, so ask the AI for the template with placeholders like [CLIENT NAME] and [DATE], then fill the placeholders in your records system, not in the chat. For treatment ideas, describe the presentation in general terms ("an adult with panic symptoms after a job loss") rather than the person, and leave out everything in the first column of the table above.
What about psychotherapy notes?
Keep them out of AI tools unless your compliance advice says otherwise. HIPAA defines psychotherapy notes as notes "documenting or analyzing the contents of conversation during a private counseling session," kept separate from the rest of the record (eCFR, 45 CFR 164.501). They get the strictest rule in the Privacy Rule: a covered entity "must obtain an authorization for any use or disclosure of psychotherapy notes," with a short list of exceptions such as the note's author using it for treatment (eCFR, 45 CFR 164.508). Progress notes, with diagnosis, treatment plan, and symptoms, are a separate category.
Do I have to tell clients I use AI?
The ACA says yes. Its recommendations for practicing counselors say to "clearly inform clients about the use of AI tools in their counseling process" and to "obtain explicit informed consent from clients for the use of AI-assisted tools." The same page says AI platforms used for counseling "must adhere to the standards set by applicable local privacy laws and regulations (e.g., HIPAA in the United States)" (American Counseling Association). Your state board may set its own rules.
How Morse keeps your drafting off someone else's servers
A consumer chatbot keeps your draft on its servers, tied to your account, under a policy written for the public rather than for your clients. Whatever tool you use, the workflow starts the same way: swap each item in the de-identification table for what its last column suggests, then draft.
Morse changes where that de-identified draft goes. In an E2EE chat, your browser checks the secure enclave's proof that it's genuine and encrypts the draft so only that enclave can read it, and our servers relay text they can't read; the guarantee rests on the enclave hardware and attestation working as designed. The chat history stays in your browser, encrypted, and Morse doesn't train AI on it. A Temporary chat is kept in memory only, which fits a note you'll paste into your records system and close. Clearing your browser deletes your chats for good, so the finished note belongs in your records, not in the chat.
- A progress note from a de-identified summary: E2EE.
- A letter template with placeholders: Private, the default.
- A psychoeducation handout: any mode.
For how the encryption works, read end-to-end encrypted AI chat. Try a chat from the Morse home page without an account, or start free. End-to-end encrypted chats come with Pro, at $19 a month.
Common questions
Is replacing the client's name with initials enough?
No. Initials are part of the name, and HIPAA allows a re-identification code only if it "is not derived from or related to information about the individual." Every other identifier on the list still has to go too.
Does my EHR vendor's BAA cover the AI features inside it?
Not automatically. Google says third-party add-ons are not included in what its Workspace BAA covers, and Anthropic's BAA excludes beta features. Check your vendor's list of covered features for the AI tool by name.
Once a note is properly de-identified, do HIPAA's rules still apply to it?
The Privacy Rule says its requirements "do not apply to information that has been de-identified" under its standard. Re-identifying it, or disclosing the key that would re-identify it, brings the rules back.