Is ChatGPT HIPAA Compliant? Health and Business

"ChatGPT and HIPAA" is really two questions wearing one headline. The first is personal: OpenAI's ChatGPT Health lets you connect your medical records for personalized answers, and HIPAA generally does not follow that copy once it leaves your provider's system. The second is professional: can a therapist, clinic, or hospital use ChatGPT, Claude, or Gemini for work touching patient data and still stay HIPAA compliant. The answer depends entirely on the product tier and whether a Business Associate Agreement is signed, and for anyone typing into the free or Plus version of any of these tools, the honest answer is no.

This piece covers both. It explains what ChatGPT Health actually does and what HIPAA does and does not cover for that consumer feature, then walks through the business side: what a Business Associate Agreement requires, which AI vendors currently offer a HIPAA-eligible tier, and what happens to a practice that puts patient data into a tool that is not covered by one.

Is ChatGPT HIPAA compliant? The short answer

No, not by default, and not on any plan a person signs up for on their own. Generic ChatGPT, meaning the Free, Plus, and Team tiers, is not HIPAA compliant and cannot be made compliant through careful prompting or a strong password. ChatGPT becomes usable with protected health information only through a signed Business Associate Agreement, which OpenAI currently offers for ChatGPT for Healthcare and ChatGPT Enterprise on a sales-managed account, plus select API access with modified data retention. The same rule holds for Claude and Gemini: a consumer login does not carry HIPAA protection no matter how the conversation is phrased.

• Free, Plus, Team, and a personal Gemini or Copilot login: no BAA is available, so patient data should not go in.
• ChatGPT Health, the consumer feature covered below: a personal feature on the consumer product, not a HIPAA-covered tool for a provider's patients.
• ChatGPT for Healthcare, ChatGPT Enterprise, Microsoft 365 Copilot and Azure OpenAI, Gemini for Workspace and Vertex AI, and Claude Enterprise or the API: HIPAA-eligible only after a signed BAA and the right configuration.

What ChatGPT Health actually does

ChatGPT Health is a feature that lets you connect health information so ChatGPT can answer questions about your own body, not just general medicine. It rolled out on July 23, 2026 to US users who are 18 or older, on web and iOS, across the Free, Go, Plus and Pro plans.

Once connected, it can draw on your medical records, things like lab results, medications and visit notes, and on Apple Health data such as fitness and activity. According to the launch reporting, it can compare lab test data over time, summarize what changed since a prior appointment, track medications, and explore how sleep, activity and workouts relate to your health.

OpenAI says ChatGPT will ask for permission before using connected medical records and Apple Health information to personalize a response. That consent step is useful, but it is a control over how the data is used inside the product, not a promise about who can reach the data later.

Also on MemX

AI & Health

Therapists, AI Notes Can Break HIPAA

11 min read→

AI & Health

Nurses Are Pasting Patient Data Into ChatGPT

11 min read→

AI & Health

AI Medical Scribes Remember the Visit, Not You

11 min read→

What HIPAA covers, and what it does not

HIPAA does not cover ChatGPT by default. The law applies to covered entities and their business associates: hospitals, health plans, physician practices, insurers, and the vendors they hire under a signed contract to handle protected health information. A consumer AI app you sign up for yourself, or a provider using one without that contract, falls outside the law either way.

Here is the part that trips people up. While your records sit in your doctor's or hospital's system, HIPAA governs how they are shared and disclosed. The same law gives you the right to get a copy of those records. But once you exercise that right and load the copy into a consumer app, HIPAA's protection does not travel with it. Privacy advocates have flagged this directly: Sara Geoghegan of the Electronic Privacy Information Center warned that moving records into ChatGPT "would remove the HIPAA protection from those records, which is dangerous."

What OpenAI says about training and ads

OpenAI has made specific commitments about health data, and they are worth reading precisely. Connected medical records and Apple Health information, and the conversations that use that health data, are not used to train OpenAI's models or target ads, regardless of the model-training setting chosen elsewhere in the account.

Read the boundary carefully. There is a difference between a chat where you have connected records and one where you simply type a health question. A health question typed into an ordinary chat, without invoking connected records, follows your standard model-training setting, the one you control under Settings, then Data Controls. So the strong no-training promise attaches to the connected health feature, not automatically to everything health-related you ever type.

These are policy commitments, and they are meaningful. They are also not the same thing as a legal guarantee. A policy is a promise a company makes and can revise. HIPAA is a law with enforcement and penalties behind it. When protection rests on a policy rather than a statute, the questions that matter are: can the policy change, and what happens to the data if it does.

The risks that remain even with good policies

Good privacy policies reduce risk. They do not remove it, for a patient connecting their own records or for a practice handling someone else's. Once a copy of health history lives in a consumer account, or in a business account without the right controls, a handful of real exposures come with it regardless of what the marketing says.

• Data breach: any company holding data can be breached. A copy of labs and medications sitting in an account is a target that would not exist if it had never been uploaded.
• Legal process: consumer data can be subject to subpoenas or law-enforcement requests, reached through a different and often broader door than HIPAA's own carve-outs.
• Account access: a weak password, reused credentials or a shared device can expose whatever health context is connected to that account.
• Policy and scope creep: today's no-training promise is a current policy that can be revised, and the more that gets connected, the larger the single place holding a map of someone's health.

Can a healthcare practice or therapist use ChatGPT and stay HIPAA compliant?

Under HIPAA, a covered entity such as a clinic, hospital, therapist, or health plan cannot hand protected health information to an outside vendor, including an AI company, without a signed Business Associate Agreement. HHS's own guidance sets out what that contract must include: the permitted and required uses of the data, a promise the vendor will not use or disclose it beyond what the contract or law allows, safeguards including the technical protections in the HIPAA Security Rule, a duty to report unauthorized use back to the covered entity, the same terms flowing down to any subcontractor, and a plan to return or destroy the data when the contract ends.

Signing a BAA is necessary, but not sufficient on its own. The agreement covers what the vendor does with data once received; it does not stop staff from pasting a chart note into a tool never configured for protected health information, or restrict what gets typed into a consumer app on a personal phone. A practice using AI on patient data needs both the signed contract and internal controls: the right product tier, the right settings, and staff trained on which login is actually approved.

Which AI vendors currently offer a HIPAA-eligible tier

The BAA landscape changes as vendors add healthcare-specific products, so a practice should confirm current terms directly with the vendor before relying on this for a live decision. As of September 2026, four vendors cover most of clinical AI use, and each draws the line between consumer and business access in a different place.

A signed BAA is an agreement about how the vendor handles data it receives, not a filter on what gets sent to it. Every one of these companies says a version of the same thing: the contract does not stop protected health information from reaching the model if an employee pastes it into an unapproved feature or a screenshot carries it into a chat. The BAA has to be paired with configuration: the healthcare-specific tier turned on, consumer features turned off, and staff told exactly which login is approved for patient data.

What happens if a provider uses a non-BAA tool with patient data

Pasting a clinical note, a lab result, or a patient's name into a version of ChatGPT, Claude, or Gemini that has no BAA in place is an impermissible disclosure under HIPAA, regardless of how secure the vendor's infrastructure actually is. The covered entity, not the AI company, carries the liability, because HIPAA obligations sit with the provider that disclosed the data, not with the tool that received it.

HHS's Office for Civil Rights enforces HIPAA with civil penalties adjusted for inflation each year. As of the January 2026 update, the four culpability tiers run from $145 to $73,011 per violation for a lack-of-knowledge case, up to $73,011 to $2,190,294 per violation for willful neglect that goes uncorrected, with an annual cap of $2,190,294 per violation category. An unauthorized disclosure can also trigger the HIPAA Breach Notification Rule, meaning the practice has to notify affected patients and, above a certain size, HHS and the media, on top of whatever penalty follows.

How to use ChatGPT Health more safely

For an individual connecting their own records, there is real value in personalized health answers while keeping exposure small. The habits below cost nothing and cut most of the avoidable risk.

• Share the specific, skip the sensitive: connect the lab trend or medication list actually needed, rather than an entire chart by default.
• Keep a personal copy: download and store records separately so the app is never the only place the health history exists.
• Lock the account: use a strong, unique password and turn on two-factor authentication, since account access is one of the most common ways private data leaks.
• Delete when done: if records were connected for a one-time question, remove them afterward and confirm what deletion actually clears.
• Do not treat it as medical advice: use it to understand and organize, then confirm anything that affects treatment with a clinician.

Where your personal context lives is the real question

Connect health data to any AI, personal or professional, and a different question becomes urgent: where does that context live, and who can reach it. That question is not unique to health. It applies to documents, receipts, photos, voice notes and messages, the everyday record of a life that a memory app holds.

This is the problem MemX is built around. MemX is a consumer memory app: snap or save something now, ask a plain question later, and get the answer with its source. It is private by architecture, meaning per-user isolation so data is not pooled with other people's, customer-managed encryption keys, encryption at rest, and on-device processing where possible, with no training on your data. The design goal is data minimization and isolation, so personal context stays reachable by the person who owns it and is not repurposed.

Being precise about what that does and does not mean: MemX is not HIPAA-covered, and private by architecture is an engineering and data-handling stance, not a legal shield or medical protection. If records need HIPAA coverage, the safest place for that copy is inside the healthcare system that HIPAA governs, or inside a vendor's product that sits under a signed BAA. What a private-by-design store offers is a smaller, isolated home for everyday personal context, so fewer parties can reach it in the first place.

The bottom line

ChatGPT Health can genuinely help make sense of labs, medications and activity trends, and OpenAI's stated policy of not training on connected health data or using it for ads is a real commitment, though it is a policy rather than HIPAA. On the business side, the picture is stricter: a clinic, therapist, or hospital cannot use ChatGPT, Claude, or Gemini on patient data without a signed BAA and the right configuration, and every major vendor now offers some path to that, gated behind an enterprise or healthcare-specific tier.

Use these tools deliberately either way. As a patient, connect narrowly, keep a personal copy, lock the account, and delete what is no longer needed. As a provider, confirm the BAA is signed and the approved tier is what staff are actually using before any patient detail goes in, because the fine schedule and the breach notification duty land on the practice, not on the AI company.