HIPAA Compliant AI: Is ChatGPT HIPAA Compliant, and Which LLMs Sign a BAA?

HIPAA compliant AI is possible in 2026, but not with the free or personal versions of popular chatbots. Consumer ChatGPT, Claude and Gemini plans do not come with a business associate agreement (BAA), so putting protected health information (PHI) into them breaks HIPAA. Enterprise plans and APIs from OpenAI, Anthropic, Microsoft, Google and AWS can be used with PHI once a BAA is signed and the account is configured correctly.

So, is ChatGPT HIPAA compliant? The short answer: the consumer app is not; ChatGPT for Healthcare and ChatGPT Enterprise can support HIPAA compliance under a BAA; and the API can, for eligible accounts with the right data-retention settings. No AI tool is "HIPAA certified", because no such certification exists. Compliance comes from the contract plus how you configure and use the tool.

This guide gives a sourced vendor table, a safe architecture for LLM apps that handle PHI, and the equivalent rules in the UK, UAE and India.

What makes AI HIPAA compliant?

HIPAA compliant AI is an AI system used under three conditions: a signed BAA with every vendor that creates, receives, stores or transmits PHI on your behalf; technical safeguards that meet the HIPAA Security Rule, such as access control, encryption, audit logs and retention limits; and organisational practices, including a risk analysis, minimum-necessary use and staff training.

The BAA matters because it makes the AI vendor legally responsible for protecting PHI and reporting breaches. Without one, sharing PHI with the vendor is an impermissible disclosure, however secure the vendor's technology is.

Which AI vendors sign a BAA? (as of November 2026)

Vendor terms change often. The table below reflects vendor documentation and announcements we reviewed in November 2026; confirm current terms with each vendor before any PHI flows.

Sources include Anthropic's HIPAA-ready Enterprise help article, which states that Team and individual plans cannot enable HIPAA, and OpenAI's January 2026 OpenAI for Healthcare announcement, which introduced ChatGPT for Healthcare with a BAA option and named early users including AdventHealth, Cedars-Sinai, HCA Healthcare and UCSF.

Is ChatGPT HIPAA compliant? The detail

There are three ways clinicians and hospitals use ChatGPT, and each has a different answer.

• Personal ChatGPT accounts. Not HIPAA compliant. OpenAI does not sign a BAA for consumer plans. A clinician pasting a discharge summary into a personal account is a disclosure of PHI, even with names removed if other identifiers remain.
• ChatGPT for Healthcare and Enterprise. These workspaces can support HIPAA compliance once the organisation signs a BAA, with admin controls, audit logs and data retention settings. Your organisation still decides who gets access and what they may enter.
• The OpenAI API. Developers building apps can request a BAA. OpenAI ties eligibility to the account's data-retention configuration, and some endpoints and tools that keep application state sit outside the covered scope. Check the current list before you design the app.

The same pattern applies to Claude: consumer and Team plans are out; the HIPAA-ready Enterprise plan and the API under a BAA are in, with specific features excluded.

HIPAA compliant AI note takers and scribes

"HIPAA compliant AI note taker" is one of the most searched variants of this question, because clinicians want ambient documentation. General meeting note takers vary widely: some offer a BAA only on a paid HIPAA plan, and others offer none at all. Medical scribe vendors normally sign BAAs, but you should still ask where audio is stored, how long it is kept, whether it is used for training, and whether clinicians must sign every note. Our AI medical scribe guide covers these questions in detail.

A safe architecture for a HIPAA compliant LLM app

Signing a BAA is step one. The architecture determines how much PHI is exposed and how easy it is to prove control. This is the pattern we use for clinical documentation AI and other PHI workloads.

1. Host the application in your own cloud account under a BAA, in a HIPAA-eligible region.
2. Call the model through a BAA-covered route, such as Amazon Bedrock, Azure OpenAI, Vertex AI or a direct vendor API with a BAA.
3. Minimise PHI before the prompt: strip direct identifiers the task does not need, and use tokens that you map back afterwards.
4. Keep source documents in your own store and use retrieval (RAG) to pass only the relevant passages, rather than sending whole records.
5. Turn off or limit vendor-side retention where the vendor allows it.
6. Log every prompt, response, user and model version in your own audit store, with access controls.
7. Put a human review step before any output reaches a chart or a patient.
8. Run a documented HIPAA risk analysis and update it when the model, vendor or workflow changes.

Our RAG development and AWS and cloud teams build this pattern so PHI stays in an account you control.

A quick decision tree: can I put this patient data into this AI tool?

1. Is the data PHI (identifiable health information held by a covered entity or business associate)? If no, HIPAA does not apply, though other privacy laws may.
2. Is there a signed BAA with the AI vendor that covers this exact product and feature? If no, stop.
3. Is the tool configured as the BAA requires (retention, region, covered features only)? If no, stop.
4. Is this the minimum PHI needed for the task? If no, reduce it first.
5. Is the output reviewed by a person before it affects care or the record? If no, add a review step.
6. Is the use logged and covered in your risk analysis? If yes to all, proceed.

Common mistakes with AI and PHI

These are common errors when healthcare teams start using LLMs.

• Shadow AI. Staff use personal chatbot accounts because no approved tool exists. Banning them without offering an alternative rarely works.
• Assuming the BAA covers every feature. Vendors exclude specific features, beta tools, web search or cloud sessions from BAA scope. Check the implementation guide for each one you enable.
• Logging PHI in the wrong place. Application logs, error trackers and analytics tools often capture prompts and responses. Each of those vendors needs a BAA, or the data must be stripped first.
• Treating "anonymised" as de-identified. Dates, rare diagnoses, postcodes and free-text details can identify a patient even without a name.
• Forgetting state and local law. Some US states add rules for health data or AI use beyond HIPAA, and outside the US, local law applies to the same data.
• No exit plan. Ask how data is returned or deleted when the contract ends, and how long backups are kept.

Fixing these is mostly policy and configuration work, and it is far cheaper before go-live than after an incident.

Equivalent rules in the UK, UAE and India

HIPAA is US law, but the same question comes up everywhere.

In the UK, NHS England's guidance on AI-enabled ambient scribing expects DTAC and DSPT compliance, clinical safety assessment and a decision on whether the product is a medical device.

In the UAE, Article 13 of the ICT health law restricts storing or processing health data outside the country unless a health authority approves it; Al Tamimi's summary of permitted transfers explains the exceptions. In practice, that means using a UAE cloud region and checking that the model you want is available there. Our post on health data residency in India and the UAE covers cloud options.

In India, the DPDP Rules were notified in November 2025, with most obligations applying 18 months later. See DPDP Act for hospitals for a checklist.

How Algon approaches this

Algon is a Kochi-based team that builds LLM and AI systems for hospitals and health companies in India, the UAE and the USA. We design the data path first: which fields the model actually needs, which cloud account and region it runs in, which vendor agreements cover it, and who reviews outputs. Then we build in short cycles, so compliance and clinical staff can check working software rather than diagrams. Our LLM development services follow this order for every healthcare project.

Conclusion

HIPAA compliant AI is a matter of contracts, configuration and discipline, not a badge on a product. Use enterprise or API routes with a signed BAA, minimise PHI, keep records in your own environment, log everything and keep a human in the loop. Ban personal AI accounts for patient data, and give staff a sanctioned alternative so they are not tempted to use them.