Patient Privacy and PHI
Understand when using an AI tool exposes protected health information, why a Business Associate Agreement matters, how de-identification works, and safe habits for everyday use.
The Fastest Way to Turn a Helpful Tool Into a Breach
A nurse pastes a patient’s discharge summary into a free chatbot to produce a plain-language version. A physician photographs a rash and uploads it to a consumer app for a “second opinion.” Both intentions are good, and both may be impermissible disclosures of protected health information.
This lesson uses US HIPAA as the reference point. Other jurisdictions have comparable or stricter rules (for example, GDPR in the EU and UK, PIPEDA and provincial laws in Canada). The principles transfer; the specifics belong to your privacy office.
Key Concepts
PHI
Protected health information is individually identifiable health information held by a covered entity or its business associate. Identifiability is broader than names: dates of service, MRNs, addresses, phone numbers, full-face photos, and rare-condition details can all identify a patient.
Business Associate Agreement (BAA)
When a vendor creates, receives, maintains, or transmits PHI on your organization’s behalf, HIPAA generally requires a BAA: a contract obligating the vendor to safeguard the data and limiting how it can be used.
- Enterprise AI tools your organization has contracted for (ambient scribes, EHR-integrated assistants, enterprise LLM deployments) typically come with a BAA and configured data protections.
- Consumer AI apps (free or personal paid accounts) typically do not. Their terms may allow data retention, human review, or use for training.
Whether a tool is covered is a contract question, not a features question. A consumer product and an enterprise product from the same company can have entirely different terms.
De-identification
HIPAA provides two pathways:
- Safe Harbor: remove 18 specified identifier types (names, geographic units smaller than a state, all date elements except year, contact numbers, MRNs, device identifiers, full-face photos, and others) and have no actual knowledge that the remainder could identify the individual.
- Expert Determination: a qualified expert certifies that re-identification risk is very small.
Removing the name alone is not de-identification. “68-year-old retired ferry captain from Port Townsend admitted March 4 with…” may identify someone in a small community.
A Decision Path for Every Use
- 1 Approved tool?
Organization-sanctioned for this use
- 2 BAA in place?
Confirmed by IT, privacy, or compliance
- 3 Minimum necessary?
Only what the task requires
- 4 Proceed
Within policy, with review
Safe habits
- Abstract the question. “What’s the renal dosing adjustment approach for [drug] at an eGFR of 25?” needs no patient identifiers at all.
- Apply minimum necessary even in approved tools. Don’t paste the whole chart when the task needs one lab trend.
- Never use personal accounts for patient work, even “just this once.”
- Images are high risk. Photos can contain faces, tattoos, wristbands, room numbers, and metadata.
- Know your organization’s policy and where to ask. Most health systems now have a published AI use policy and an approved-tools list.
Common mistake: believing that disabling model training in a consumer app makes it acceptable for PHI.
Turning off training may reduce one risk, but it doesn’t create a BAA, guarantee retention limits, or restrict human review under HIPAA terms. Compliance depends on the contractual and technical arrangement your organization has with the vendor.
Beyond Legal Compliance: Trust
Even where a use is technically permitted, consider how the patient would feel if they knew. Many health systems are adopting transparency practices, such as telling patients when ambient listening is used and offering an opt-out. Privacy is ultimately about preserving the trust that makes patients willing to tell you the truth.
- Locate your organization’s AI acceptable-use policy and approved-tools list (try the intranet, compliance, or IT pages).
- For each tool on your Module 1 inventory, note whether it’s approved and whether it’s cleared for PHI.
- Take a real question you might ask AI about a patient and rewrite it so it contains no identifiers and still gets a useful answer.
- If you can’t find a policy, write down who you’d ask (privacy officer, CMIO, compliance) and ask them this week.
1. A physician wants to paste a patient's full progress note into a free consumer chatbot to generate a summary. What's the main problem?
Without a BAA and approval, entering PHI into a consumer tool generally isn't permitted, and removing the name alone isn't de-identification.
2. Which statement about HIPAA de-identification is correct?
Safe Harbor covers 18 identifier categories, including most date elements and full-face photos.
3. What's the best way to ask a dosing question without exposing PHI?
Many clinical questions need no identifiers. Abstracting them is the simplest privacy safeguard.
Up Next
Module 3 turns to Description: how to specify clinical tasks so AI output is useful, appropriately cautious, and easy to verify.