Bottom line up front: ChatGPT is not HIPAA-compliant by default. Consumer and Plus versions offer no Business Associate Agreement (BAA), so entering protected health information (PHI) into them discloses PHI to a vendor with no HIPAA obligations — a potential reportable breach. OpenAI supports BAAs only for its API and enterprise offerings, and a BAA alone does not make workflows compliant. Your practical choices in 2026: block PHI from reaching AI tools, or route AI through a BAA-covered, safeguarded path.
The scale of the problem is measured, not hypothetical
Netskope's May 2025 healthcare analysis found that 81% of data policy violations in healthcare organizations involved regulated data — overwhelmingly the kind of information staff paste into generative AI tools to summarize a chart, draft a letter, or translate discharge instructions. The intent is almost always good. The disclosure is still a disclosure.
What HIPAA actually requires here
- A BAA before PHI touches a vendor. If a third party creates, receives, maintains, or transmits PHI on your behalf, you need a Business Associate Agreement. Consumer ChatGPT has none, and OpenAI's terms for the consumer product do not contemplate PHI.
- The minimum necessary standard. Even under a BAA, sending a full chart to draft a two-line letter fails minimum-necessary scrutiny.
- Access controls and audit trails (Security Rule). You must be able to show who disclosed what, when — browser-pasted prompts leave you nothing.
- Breach analysis when it goes wrong. PHI sent to a non-BAA vendor triggers a four-factor breach risk assessment, and potentially notification to patients and HHS.
Why "just ban it" fails
Policy-only bans do not survive contact with a busy clinic. Staff use personal devices, personal accounts, and whatever tool answers fastest. A written AI policy with no technical control is an unenforced policy — and unenforced policies read badly in an OCR investigation.
The compliant architecture: allow AI, block PHI
The workable posture for most physician groups and clinics is a local AI firewall:
- Route staff AI traffic through a proxy running on your own infrastructure (HoundShield's HIPAA deployment runs as self-hosted Docker — prompt content never leaves your network to be scanned).
- Detection engines scan each prompt locally in under 10ms for PHI markers — names, MRNs, DOBs, diagnosis codes, insurance identifiers.
- Prompts containing PHI are blocked or require justification; clean prompts pass through, so staff keep the productivity win.
- Every event lands in a tamper-evident audit log — the artifact your next risk analysis and any OCR inquiry will ask for.
Note the contrast with cloud-based DLP products: tools that scan prompts in their own cloud receive your PHI in order to protect it, which puts you right back in BAA territory with the DLP vendor. Local scanning avoids the recursion entirely.
A Privacy Officer's 30-day checklist
- Week 1 — discover: survey which AI tools staff actually use (assume more than you think), and check whether any have BAAs in place.
- Week 2 — policy: publish a one-page AI acceptable-use policy: no PHI in any AI tool without an approved, BAA-covered pathway.
- Week 3 — enforce: deploy local scanning so the policy is a control, not a memo.
- Week 4 — evidence: add AI usage to your HIPAA risk analysis, and keep the audit log as standing documentation.
Common questions — including the AI-scribe vendors and BAA-covered API workflows — are answered at Is ChatGPT HIPAA compliant?. To see what your staff's AI traffic actually contains, the $499 AI risk assessment runs 14 days in your environment and reports every PHI event it caught.