The Confusion Around ChatGPT in Healthcare
Ask five healthcare professionals if you can use ChatGPT in practice, and you will get five different answers. Some say it is a HIPAA violation. Others use it every day to write patient emails.
Whether it is safe depends on *what* you put into it and *which tier* of the software you are using.
The Tier Distinction
First, you must understand the difference between the versions of ChatGPT.
Consumer ChatGPT (Free or Plus): This is the version most people use. By default, OpenAI may use the data you type into these versions to train future models. You can turn that off under Data Controls, and you should. But do not mistake that toggle for protection. Turning off training does not turn off storage. Your conversations are still retained on OpenAI's systems, and the zero data retention option OpenAI publishes is available on the API platform only, not on consumer ChatGPT. Opting out of training is hygiene. It is not a HIPAA control. Do not put Protected Health Information (PHI) into this tier. If you are a covered entity or a business associate, disclosing PHI to a vendor with no BAA in place is an impermissible disclosure, and you will be doing breach analysis you did not need to do.
One frame before the rest of this post. I am a pharmacist, not your attorney, and this is general education, not legal advice. HIPAA applies to covered entities and business associates, so whether a specific rule binds you depends on your role. Your state privacy law, your employer's policy, your malpractice carrier, and 42 CFR Part 2 for substance use disorder records can all be stricter than HIPAA, and vendor terms change without telling you. Verify the current terms with the vendor and run your AI plan past your compliance officer before you put it into practice.
ChatGPT for Healthcare, ChatGPT Enterprise with a Regulated Workspace, and the API with Modified Retention: These are the OpenAI products that can be covered by a Business Associate Agreement (BAA). Two things matter here. A BAA is not automatic with an Enterprise seat, you have to be provisioned into the eligible configuration and execute the agreement. And ChatGPT for Healthcare is sold by contacting OpenAI's team, not by clicking upgrade. In these configurations, content is not used to train models, and OpenAI will sign a BAA to support your HIPAA compliance. Read that carefully. A BAA supports your compliance, it does not deliver it, and there is no such thing as HIPAA certified software. OpenAI's own documentation is blunt that a feature being available does not mean that feature is covered by your BAA, and several features are excluded from coverage even inside the eligible products. You still owe the risk analysis, the workforce training, the access controls, and the audit trail.
If you are a solo practitioner on the $20/month Plus version, you are in the consumer tier. Act accordingly.
And here is the trap that catches small practices. ChatGPT Business, the small team plan, is not on OpenAI's HIPAA eligible product list. OpenAI does say it never trains on Business data by default, and clinicians read that sentence and assume PHI is now fine. It is not. "Not used for training" and "covered by a BAA" are two different promises, and only the second one lets PHI in the door.
Safe Use Cases (The Administrative Lane)
If you are using consumer-tier ChatGPT, you can still get a lot out of it by keeping it strictly in the administrative lane.
1. Patient Education Materials
Prompt: *"Explain the mechanism of action of SSRIs at an 8th-grade reading level. Use an analogy about a recycling plant."*
Why it's safe from a privacy standpoint: There is no patient data here. You are using the AI to translate a complex clinical concept into accessible language.
Now the part most people skip. Privacy safe is not the same as clinically safe. Anything generated here goes out under your license, to a patient, as your professional communication. Models fabricate mechanisms, invent onset timelines, and quietly drop the warning that mattered most. Every piece of AI drafted patient education gets verified against a primary source before it reaches a patient. Use the AI for the reading level and the analogy. Do the clinical accuracy check yourself.
2. Standard Operating Procedures (SOPs)
Prompt: *"Draft an SOP for how our front desk should handle a patient who arrives 15 minutes late to a 30-minute appointment."*
Why it's safe: Purely operational.
3. Marketing and Outreach
Prompt: *"Write a 3-part email sequence for our clinic's new weight management program. Focus on the holistic approach."*
Why it's safe: General business marketing.
Unsafe Use Cases (The Clinical Lane)
Never use consumer-tier ChatGPT for the following:
1. Clinical Summaries
Unsafe: *"Summarize these notes for John Doe: [Patient Notes]."*
Why: You just uploaded PHI to a consumer AI model.
2. Diagnostic Assistance
Unsafe: *"A 45-year-old female presents with..."*
Why: Deleting the name is not de-identification. Under the Safe Harbor method at 45 CFR 164.514(b)(2) you have to strip 18 categories of identifiers, not one. That list includes every geographic subdivision smaller than a state, every element of a date except the year, medical record and account numbers, device identifiers, URLs and IP addresses, biometric identifiers, full face photographs, and a catch all for any other unique identifying number, characteristic or code. On top of that, the rule requires that you have no actual knowledge the remaining information could identify the person, alone or combined with anything else. A detailed presentation in a small practice or a rare condition can fail that test with the name already gone, and relying on a general LLM for diagnostic reasoning carries real liability risk.
3. Specific Treatment Protocols
Unsafe: *"Write a supplement protocol for my patient with Hashimoto's and these specific lab values: [Labs]."*
Why: Two separate problems, and it helps to keep them separate. First, privacy. Labs plus a diagnosis plus "my patient" is almost never de-identified in the Safe Harbor sense, so you are most likely disclosing PHI to a vendor with no BAA. Second, and this one survives even if you did strip every identifier: this is individualized treatment planning. The model has no access to the chart, no medication list, no renal or hepatic function, and no ability to verify itself. It will generate plausible interactions and contraindications that are simply wrong. De-identifying the input does not make the output safe to act on.
The Rule
If the task involves a specific patient, it belongs in an environment where a BAA is actually executed and actually covers the feature you are using. Your EHR usually qualifies. Google Workspace can qualify, but only if your administrator has reviewed and accepted Google's BAA, and only for the services on Google's HIPAA Included Functionality list. Google states plainly that the BAA does not extend to Additional Google Services and does not cover third party add ons. A personal Gmail account is not a BAA covered environment, no matter how carefully you write the email. If the task involves general education, business operations, or marketing, consumer ChatGPT is fine to use.