How-to guides

Can ChatGPT write an ABA treatment plan? HIPAA, BAAs and alternatives

ChatGPT can write the words, but a personal ChatGPT plan has no BAA. What counts as PHI, Safe Harbor de-identification, the BACB code and a covered workflow.

Can ChatGPT write the plan at all?

In a narrow sense, yes. Give a general-purpose model a diagnosis, a list of assessment scores and a target format, and it will produce fluent goal statements, a medical-necessity paragraph and a parent-training section. What it can't do is know your client. It hasn't seen the baseline data unless you paste it. It won't notice that a goal contradicts the assessment unless the contradiction is sitting in the text you gave it. And it has no idea what your client's payer requires this quarter. The draft is a starting point that's only as good as the records behind it, and those records are exactly what you can't hand to a consumer chatbot.

That's why the HIPAA question comes first, not second. A plan written from de-identified fragments is thin. A plan written from the full record means the record has to go somewhere that's covered.

Why the BAA is the whole question

HIPAA's definitions in 45 CFR 160.103 make a business associate any person who, on behalf of a covered entity and outside its workforce, "creates, receives, maintains, or transmits protected health information" for a regulated function, including data analysis and practice management [2]. A vendor whose software drafts a treatment plan from a client's records receives and maintains PHI on your behalf. That makes it a business associate.

The Privacy Rule then sets the condition. A covered entity "may disclose protected health information to a business associate ... if the covered entity obtains satisfactory assurance that the business associate will appropriately safeguard the information", and that assurance "must be documented through a written contract or other written agreement" [3]. That written agreement is the BAA. It isn't paperwork bolted onto compliance. It's the thing that makes the disclosure allowed in the first place.

So "can I use ChatGPT for this?" really means "will OpenAI sign a BAA with me for the product I'm using?" On the pages we read, the answer depends on the product:

  • The API platform. OpenAI's developer documentation says data sent to the API hasn't been used to train its models since 1 March 2023 unless the customer opts in, that abuse-monitoring logs are kept for up to 30 days by default, and that approved customers can enable Zero Data Retention, which keeps their content out of those logs [1]. It says "BAA-eligible endpoints can be used for processing PHI" for customers who have executed OpenAI's Business Associate and Healthcare Addendum and whose organization has been provisioned with one of its retention controls, and that web search with live internet access "is not HIPAA eligible and is not covered by a BAA" [1]. OpenAI's list of HIPAA-eligible products puts the same two conditions plainly: "API HIPAA eligibility requires an executed OpenAI BAA. Your organization must also be provisioned with Modified Retention, unless OpenAI specifies otherwise" (HIPAA eligible products and functionality, read 1 November 2026). OpenAI's help-center article on the API BAA, read on 1 November 2026, adds that eligible organizations can accept the standard BAA in the API platform's settings, that self-serve enrollment needs an established history of API usage, that an enterprise agreement isn't required, and that accepting a BAA doesn't by itself make an application HIPAA compliant (Getting a Business Associate Agreement for the OpenAI API). This is the route software vendors take, not individual clinicians.
  • ChatGPT Enterprise, Edu and ChatGPT for Healthcare. The same article says a BAA for a sales-managed Enterprise or Edu account comes through OpenAI's sales team. OpenAI's separate list of HIPAA-eligible products names "ChatGPT for Enterprise with Regulated Workspace" rather than plain Enterprise, and doesn't list Edu, so ask which product a BAA covers. OpenAI's ChatGPT for Healthcare page describes an enterprise workspace with a BAA available, no training on customer data and admin-controlled retention, sold to healthcare organizations (ChatGPT for Healthcare, read 1 November 2026). It also notes that its "improved memory" feature isn't covered under the BAA and that PHI shouldn't be entered when using it.
  • ChatGPT for Clinicians. A free, individual-use version for verified clinicians in the United States that lets you sign a BAA in the product. Eligibility on 1 November 2026 named physicians (MD and DO), nurse practitioners, physician assistants and pharmacists, verified by NPI and license. The article says eligible clinicians "currently include" those groups and that OpenAI plans to expand access to more countries over time (ChatGPT for Clinicians). Behavior analysts aren't on that list today.
  • ChatGPT Business, Free, Plus and Pro. The BAA article states: "We do not offer a BAA for ChatGPT Business." It describes no BAA path for the Free, Plus or Pro plans. OpenAI's enterprise privacy page, read the same day, says business and API data isn't used for training by default and, separately, that OpenAI uses "data from versions of ChatGPT and other services for individuals" to train its models (Enterprise privacy at OpenAI, updated 8 January 2026). OpenAI's data-controls article adds that on personal Free, Plus and Pro workspaces "data sharing is enabled by default" and that you can turn off "Improve the model for everyone" (Data controls in ChatGPT, read 1 November 2026). Turning training off changes what OpenAI does with your content. It doesn't create a BAA.

OpenAI's help center and corporate pages block automated retrieval, so they're linked inline above rather than numbered in the Sources list. The developer documentation can be fetched, so it's cited as source [1]. Re-read them before you rely on this page; OpenAI's product names and eligibility lists change.

What counts as PHI in an ABA treatment plan?

Individually identifiable health information is information, including demographic information, created or received by a health care provider that relates to a person's physical or mental health, the care provided, or payment for it, and that identifies the person or gives "a reasonable basis to believe the information can be used to identify" them [2]. Protected health information is that information when it's transmitted or maintained in electronic media or "in any other form or medium", with narrow exclusions for education and employment records [2]. The Privacy Rule protects it in the hands of a covered entity or its business associate.

Almost everything in a treatment plan, and in the documents behind it, qualifies. The obvious items are the client's name, date of birth, address, Medicaid or plan member number and the diagnosis tied to them. The less obvious items are the ones clinicians strip last: caregiver names, the school or daycare, the referring physician, session dates and assessment dates, the client's ID in your practice-management system, a photo on the cover sheet, and the free-text history that mentions a sibling, a street or an employer. Health information plus any one of those is PHI.

Here's a made-up example, invented for this page and not drawn from any real client. A BCBA pastes a four-year-old's prior plan into a chatbot after swapping the name for "Client A". The plan still has the child's birth date, health plan member number, the dates of the last three assessments and the name of the preschool. Every one of those is on the Safe Harbor list below, so the record is still PHI. Using initials instead of "Client A" wouldn't have helped either. HHS's guidance says a record that contains patient initials doesn't meet Safe Harbor.

De-identifying a record properly

The Privacy Rule gives you two ways to turn PHI into information it no longer protects [4]. The first is Expert Determination: someone with the right statistical and scientific expertise determines, and documents, that the risk of re-identification is very small [4]. The second, and the one a small practice can actually use, is Safe Harbor. HHS's guidance on de-identification walks through both methods with examples (Guidance Regarding Methods for De-identification of Protected Health Information, read 1 November 2026); the regulation it explains is 45 CFR 164.514(b) [4].

Safe Harbor means removing the following identifiers of the individual and of the individual's relatives, employers and household members [4]:

  1. Names
  2. All geographic subdivisions smaller than a state, including street address, city, county, precinct and ZIP code, with a narrow exception for the first three digits of a ZIP code
  3. All elements of dates except the year for dates directly related to the individual, including birth date, admission date, discharge date and date of death, and all ages over 89
  4. Telephone numbers
  5. Fax numbers
  6. Email addresses
  7. Social Security numbers
  8. Medical record numbers
  9. Health plan beneficiary numbers
  10. Account numbers
  11. Certificate or license numbers
  12. Vehicle identifiers and serial numbers, including license plates
  13. Device identifiers and serial numbers
  14. Web URLs
  15. IP addresses
  16. Biometric identifiers, including finger and voice prints
  17. Full-face photographs and any comparable images
  18. Any other unique identifying number, characteristic or code

Then comes the second half of the test, which is the part most people miss: you must not have "actual knowledge that the information could be used alone or in combination with other information to identify" the person [4]. If you know the de-identified record is still recognizable, it isn't de-identified.

For ABA records, item 3 is the one that bites. Session dates, assessment dates and the date of birth all have to go, leaving the year at most. An age in years and months isn't itself a date, and the rule allows ages under 90, but next to any dated event it lets a reader work out the birth month, so treat it the same way. Item 18 covers the client ID your software assigns. Item 2 covers the preschool and the town. Initials are derived from the name, and HHS's guidance answers that question directly: a data set that contains patient initials "would not meet the requirement of the Safe Harbor method". Even with the initials gone, a birth year, a diagnosis and a small town together can give you the "actual knowledge" that fails the second half of the test. Done honestly, de-identifying a VB-MAPP summary and a prior plan takes real time, and what survives is a sketch of a client. That's why the draft that comes back is generic.

Two practical rules follow. De-identify before the record leaves your systems, not after; a redaction done inside the chatbot has already made the disclosure. And keep any re-identification key in your own records only. The Privacy Rule allows a code for re-identification as long as it isn't derived from the client's information and isn't disclosed [4].

What the BACB Ethics Code says

The Ethics Code for Behavior Analysts has applied to every BCBA and BCaBA since 1 January 2022 [6], and the current text was updated in August 2024 [5]. It has no standard that names artificial intelligence, chatbots or large language models [5]. It does have five standards that settle the question anyway.

1.02, Conforming with Legal and Professional Requirements. "Behavior analysts follow the law and the requirements of their professional community" [5]. HIPAA is the law, and the BAA requirement above is part of it.

2.03, Protecting Confidential Information. Behavior analysts "take appropriate steps to protect the confidentiality of clients", "prevent the accidental or inadvertent sharing of confidential information", and comply with applicable confidentiality requirements. The Code says "the scope of confidentiality includes service delivery ... documentation and data; and verbal, written, or electronic communication" [5]. A paste into a chat window is electronic communication of documentation and data.

2.04, Disclosing Confidential Information. Confidential information is shared only with informed consent, to protect the client or others from harm, to resolve contractual issues, to prevent a crime, or when the law compels it, and even then only "information critical to the purpose of the communication" [5]. Drafting convenience isn't on that list.

2.05, Documentation Protection and Retention. Behavior analysts "comply with all applicable requirements (e.g., BACB rules, laws, regulations, contracts, funder and organization requirements) for storing, transporting, retaining, and destroying physical and electronic documentation" [5]. Your payer contracts and your organization's approved-technology policy are "funder and organization requirements".

2.11, Obtaining Informed Consent. Behavior analysts must know and follow the conditions under which informed consent is required, including "when exchanging or releasing confidential information or records", and must explain, obtain and document it [5].

Put those together: if the record is identifiable and no BAA and no consent cover the disclosure, the problem isn't that the tool is new. It's that standards 1.02, 2.03, 2.04, 2.05 and 2.11 already describe the situation.

What a compliant AI workflow looks like

The pieces are the same whichever vendor you pick, and you can check each one before a single document is uploaded.

  • A BAA between you and the vendor, signed before any PHI moves. This is the written assurance 45 CFR 164.502(e) requires [3]. Read it. A good one says what the vendor's subcontractors may do.
  • The vendor's AI provider under its own BAA. A plan-writing tool that calls OpenAI's API is itself a business associate passing PHI to a subcontractor, and the Privacy Rule requires the business associate to get the same satisfactory assurances from that subcontractor [3]. On OpenAI's platform that means an executed Business Associate and Healthcare Addendum and the use of BAA-eligible endpoints only [1].
  • Zero data retention and no training. OpenAI's documentation describes Zero Data Retention as an approval-gated control that keeps customer content out of its abuse-monitoring logs, and says API data isn't used for training unless the customer opts in [1]. Ask whether both apply to the vendor's account.
  • Access controls on the vendor's side. Encrypted storage, access limited to the account that uploaded the documents, and a way to delete them.
  • You in the loop and on the signature line. The vendor drafts. You review every goal against the assessment, correct the medical-necessity narrative, decide the hours, and sign. The Ethics Code's responsibility for the plan doesn't transfer to the software [5].

Charm ABA's published BAA, as one example of the shape, is signed electronically before your first plan. It says PHI sent to the AI provider is processed under terms that prohibit training on it and limit retention to what's needed to return the result, with subcontractors bound by written agreements at least as restrictive as the BAA itself [7]. Other vendors publish similar documents; our comparison ABA treatment plan software compared (2026) lists what each one states.

A practical checklist

Before a client's record goes into any AI tool, including a general chatbot:

  1. Is there a signed BAA between your organization and this vendor for this product? If not, stop, or de-identify to Safe Harbor first.
  2. Does the vendor's AI provider operate under a BAA, on HIPAA-eligible endpoints, with zero data retention and no training on your data? Ask in writing.
  3. Does your organization's technology policy allow the tool? Standard 2.05 points you to that policy [5].
  4. Does your client agreement or consent form cover disclosure to this vendor? Standard 2.11 asks [5].
  5. If you're de-identifying, have you removed, not just shortened, every Safe Harbor identifier, including dates and your own client ID [4]?
  6. Will a BCBA read every line of the draft against the assessment data before signing it?

If the answer to any of the first five is no, the convenient path is the one the Privacy Rule and the Code both close.

A note on sources

The regulatory text on this page is cited from the Code of Federal Regulations as mirrored by Cornell's Legal Information Institute, which this site's automated source check can verify [2] [3] [4]. HHS's own de-identification guidance, OpenAI's help-center articles and OpenAI's enterprise privacy page block automated retrieval, so a person read each of them on 1 November 2026 and they're linked inline where they're used. OpenAI's developer documentation is cited as source [1] because it can be fetched and re-verified. This page summarizes public regulatory text and vendor pages. It isn't legal advice, and a compliance officer or counsel should confirm how it applies to your organization.

Frequently asked questions

Is ChatGPT HIPAA compliant?

No product is 'HIPAA compliant' on its own. Compliance describes how a covered entity or business associate handles data, and there's no HIPAA certification. What we can say is narrower. OpenAI's BAA article says it signs Business Associate Agreements for its API platform, for sales-managed ChatGPT Enterprise and Edu accounts and for ChatGPT for Healthcare, and offers an in-product BAA to eligible clinicians on ChatGPT for Clinicians. Its list of HIPAA-eligible products names 'ChatGPT for Enterprise with Regulated Workspace' and doesn't list Edu, so ask which product a BAA covers. It says it doesn't offer a BAA for ChatGPT Business, and its BAA article lists no path for the Free, Plus or Pro plans. Without a BAA, sending PHI to the service means you haven't obtained the written assurance the Privacy Rule requires.

Can I paste a client's VB-MAPP or prior plan into ChatGPT if I remove the name?

Removing the name isn't enough. HHS's guidance says a record that still has the client's initials doesn't meet Safe Harbor either, because initials come from the name. Safe Harbor, in 45 CFR 164.514(b)(2), means removing eighteen categories of identifiers, including every element of a date except the year (birth dates, session dates, assessment dates), any geographic unit smaller than a state, medical record and health plan numbers, and any other unique identifying number, characteristic or code. You also need to have no actual knowledge that what's left could identify the person. A dated assessment with initials, a date of birth and a Medicaid ID is still PHI.

Does OpenAI sign a BAA for ChatGPT Plus, Pro or Business?

Not on the pages we read on 1 November 2026. OpenAI says it doesn't offer a BAA for ChatGPT Business, and its BAA article describes self-serve acceptance only for the API platform, sales-managed BAAs for ChatGPT Enterprise and Edu, and a separate in-product flow for eligible clinicians on ChatGPT for Clinicians. There's no BAA path listed for Free, Plus or Pro. OpenAI's enterprise privacy page also says content from the individual versions of ChatGPT is used to train its models, while business and API data isn't used for training by default. Its data-controls article says sharing is on by default for personal Free, Plus and Pro accounts and can be turned off, but that changes training, not the missing BAA.

Is ChatGPT for Clinicians an option for a BCBA?

Not today. The eligibility list OpenAI published as of 1 November 2026 names physicians (MD and DO), nurse practitioners, physician assistants and pharmacists, verified by NPI and license, and it's designed for individual use with an in-product BAA. The article says eligible clinicians 'currently include' those four groups and that OpenAI plans to expand access to more countries, so check the list again before you rely on it. A behavior analyst who isn't on it is in the same spot as any other ChatGPT user: no BAA.

Is it a BACB ethics violation to use AI to write a treatment plan?

The Ethics Code for Behavior Analysts doesn't mention artificial intelligence. It does require you to follow the law (1.02), protect confidential information across documentation, data and electronic communication (2.03), share confidential information only with informed consent or under the other listed exceptions (2.04), meet every applicable requirement for storing and transmitting electronic documentation, including funder and organization rules (2.05), and get informed consent when exchanging or releasing confidential information (2.11). Sending identifiable client records to a service with no BAA and no consent lands inside those standards, whatever the tool is called.

What is zero data retention, and why does it matter?

On OpenAI's API platform, inputs and outputs sit in abuse-monitoring logs for up to 30 days by default. Approved customers can get Zero Data Retention, which keeps their content out of those logs. PHI may be processed only on BAA-eligible endpoints after the organization has executed OpenAI's Business Associate and Healthcare Addendum and been set up with one of OpenAI's retention controls (Modified Retention, Zero Data Retention, or Private Retention with Private Safety Processing). For you, it means the record you upload isn't sitting in a third party's logs for a month after the draft comes back. Ask any vendor whether its AI provider is under a BAA and whether retention is zero.

Sources

  1. OpenAI, developer documentation. Your data: how OpenAI uses and retains API data (data retention, Zero Data Retention, HIPAA-eligible endpoints). Retrieved November 1, 2026.
  2. Code of Federal Regulations, via Cornell Legal Information Institute. 45 CFR 160.103, Definitions (protected health information, individually identifiable health information, business associate). Retrieved November 1, 2026.
  3. Code of Federal Regulations, via Cornell Legal Information Institute. 45 CFR 164.502, Uses and disclosures of protected health information: general rules (paragraph (e), disclosures to business associates). Retrieved November 1, 2026.
  4. Code of Federal Regulations, via Cornell Legal Information Institute. 45 CFR 164.514, Other requirements relating to uses and disclosures of protected health information (paragraph (b), de-identification). Retrieved November 1, 2026.
  5. Behavior Analyst Certification Board. Ethics Code for Behavior Analysts (effective January 1, 2022; updated August 2024). Effective January 1, 2022. Retrieved November 1, 2026.
  6. Behavior Analyst Certification Board. Ethics codes (current codes, effective dates and revision history). Retrieved November 1, 2026.
  7. Charm ABA (Indy Toy Lab LLC). Business Associate Agreement, version 2026-10-01. Retrieved November 1, 2026.

Change log

  • : Rewritten in a plainer, conversational voice at the owner's request; every fact, figure, source and citation is unchanged.
  • : Corrections after fact-check: initials fail Safe Harbor (HHS guidance); age-in-months reworded; OpenAI API PHI conditions, Clinicians eligibility wording, individual-ChatGPT training default and Enterprise/Edu hedge; example renamed Client A; not-legal-advice note.
  • : First published. OpenAI, HHS and BACB pages read on 1 November 2026; the regulatory text is cited from the CFR.