Patient-directed AI: where HIPAA ends, a provider’s duties change shape
When consumer AI products invited patients to connect their medical records, HIPAA protection ended at the patient’s request. Providers could not block that flow, but they still owed patients a safe pathway, a clean record and accurate context.
Looking Back revisits one week at a time, from January 2020 to today, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
The week
Two launches in early January 2026 brought consumer AI to the medical record. On January 7, OpenAI introduced ChatGPT Health, a separate space where U.S. users could connect medical records, through the health data company b.well, and wellness apps. OpenAI said health conversations would not be used to train its foundation models and that the product was meant to support, not replace, medical care. Access began with a waitlist. On January 11, at the start of the J.P. Morgan Healthcare Conference, Anthropic announced Claude for Healthcare. It combined tools for providers and payers with an option for some U.S. subscribers to connect their own health records. Neither launch created a new legal pathway. Patients already had a right to their records, and federal rules already required providers to make records available to the apps patients choose.
Coverage focused on the AI. For health systems, the more important facts were older. Under HIPAA, patients have a right of access to their health information. Under the 21st Century Cures Act, providers and their health IT developers generally may not interfere with that access, including through standardized APIs to third-party apps. Once a patient sends records to an app that is neither a covered entity nor a business associate, HIPAA stops applying to that copy.
None of that was new in January. What changed was scale and salience. Large consumer brands were now asking patients to connect records, and patients were bringing the answers back to their clinicians. The provider’s position splits into three parts: what it could not do, what it still controlled, and what it owed patients.
What providers could not do
A provider cannot refuse or slow a patient’s request to send records to a consumer app because it doubts the app’s privacy practices or dislikes its business model. The information blocking rules contain exceptions, including for privacy and security, but they are narrow and conditional. A practice that adds friction outside those exceptions, such as extra verification steps only for certain apps or a warning screen designed to discourage, risks an information blocking complaint. Organizations should check with counsel before restricting any app.
Providers also cannot govern what the app does next. Retention, secondary use, sharing and security of the copy are governed by the company’s own terms, by state consumer health and privacy laws, and by the Federal Trade Commission. The FTC polices unfair or deceptive practices and enforces its Health Breach Notification Rule, which it updated in 2024 to make clear it reaches health apps. Those protections differ from HIPAA. Patients rarely know the difference, and many assume the hospital’s protections travel with the data.
What providers still controlled
Providers controlled the pathway. They decide how identity is proofed at the portal, how app authorization screens look, how long access tokens last, and how patients see and revoke connected apps. A screen that names the app, lists the data categories it will receive and shows how to disconnect is permitted. It is also the strongest protection the provider can offer. Security teams should treat third-party app connections as an access channel. That means logging which apps hold active authorizations, watching for unusual volumes, and making revocation as easy as consent.
Providers also controlled what they sent. A consumer AI tool will summarize whatever is in the record. A problem list that was never reconciled, a medication list with drugs stopped years ago, or a note full of shorthand becomes, in a chatbot’s answer, a confident statement about the patient. Record hygiene was always a safety issue for clinicians reading the chart. It is now also a patient experience issue, because patients read summaries generated from the same data. The practical response is to prioritize reconciliation of problem and medication lists for patients who use portals and APIs heavily, and to fix the templates that generate the most ambiguous text.
Finally, providers controlled their own AI use. Several of the same vendors sell enterprise tools to health systems under business associate agreements, where HIPAA does apply. A patient may not easily tell a tool chosen by the care team from one they chose themselves. Health systems that use AI in messaging, documentation or triage should say where and how, so the boundary is visible. That means a short public statement of which functions use AI, which vendors provide it under business associate agreements, and how a patient can reach a person instead. It also means making sure the patient-facing disclosure matches the contracts. If a vendor’s enterprise terms allow use of data for product improvement, the statement should not imply otherwise. Mismatches of that kind are what turn a patient question into a regulatory complaint.
What providers owed patients
The third part is explanation. Patients who ask a chatbot about a lab result and then call the clinic, worried, are now part of the workload. Front-desk staff, nurses and portal message teams need one consistent position. The provider supports the patient’s right to use these tools. The tool is not part of the care team, and its answers are not reviewed by the patient’s clinicians. Questions about results should come back to the care team, and here is how to ask them.
Patients also deserve a plain account of the privacy difference. Data in the portal is protected under HIPAA. A copy sent to a consumer app is governed by that company’s terms and other laws. ONC has said that educating patients about the privacy and security risks of apps is not information blocking, provided the information is accurate, unbiased and applied consistently. That gives providers room to inform without discouraging. The test is tone: the same factual notice for every app, not a warning reserved for products the organization dislikes.
These three parts need separate owners. Legal and compliance should own the limits on intervention. Portal, identity and security teams should own the pathway and record quality programs. Patient experience and nursing leaders should own the scripts. Left unassigned, each tends to fall between groups, and the first sign of the gap is a complaint.
With hindsight
The January launches accelerated a trend rather than starting one. Through 2026, patients arriving with an AI interpretation of their results became an ordinary enough experience that a standard staff response was worth having. We have no reliable figures on how many patients connected records, and we would be cautious about any claim of precision.
The legal picture did not change in substance. HIPAA still stopped at the point of patient-directed disclosure. The information blocking rules still limited how far providers could intervene. The FTC and state laws remained the main protections for the copy an app held. What health systems controlled was the preparation around that pathway: clear authorization and revocation screens, clean problem and medication lists, consistent staff scripts and a factual privacy notice. Organizations that discouraged connections informally took on information blocking risk. Those that said nothing left patients to assume the provider stood behind whatever the app said.