The chart leaves the building
When a patient links a medical record to a chatbot, HIPAA stays behind. The protections should go with the record; until Congress acts, physicians should tell patients plainly what they give up.
Listen to this special topic
read by an AI voiceThe number of Advocate Health patients who link their MyChart accounts to ChatGPT rose more than 7,000% from July to August, the health system told Becker's Hospital Review. The report gave no count of patients, and OpenAI had opened the feature to American adults generally only on July 23. The patients make the links themselves, on OpenAI's site, and Advocate, which runs 69 hospitals, did nothing to enable them. Each link draws on Advocate's Epic record system. Bobbie Byrne, its chief information officer, says the system has no real visibility or influence over the links day to day and can absorb the cost of serving them for now. "Like many things in our healthcare system," she said, "there is a disconnect between who benefits and who pays."
The larger disconnect is in the law. Inside the hospital, the record is protected by the Health Insurance Portability and Accountability Act, or HIPAA. Once it reaches an app of the patient's choosing that HIPAA does not cover, guidance from the Department of Health and Human Services says, it is "no longer subject to the protections of the HIPAA Rules." From then on, what guards it is mainly the company's own policy. Patients have every right to their records and to send them where they please. But the protections should go with the record, and until Congress makes sure they do, physicians should tell patients plainly what they give up when they link.
The department's guidance also says a hospital "would not be liable under the HIPAA Rules" for what the app later does with the record. And the Privacy Rule, it says, "generally prohibits" a provider from refusing to send records to an app the patient designates, even out of concern about how the app will use them. Separate federal rules against information blocking forbid providers to interfere knowingly and unreasonably with patients' access to their electronic records. A provider found to have broken them can face disincentives under certain programs of the Centers for Medicare and Medicaid Services, and the department announced a crackdown on violations in September 2025. In general, then, a hospital may not stop the transfer, and under HIPAA it does not answer for what follows.
The companies make generous promises. In ChatGPT the records arrive through b.well, a company that says its network reaches more than two million American providers. OpenAI says that connected records, and the conversations that use them, "are not used to train our foundation models or target ads," and that they are encrypted and kept apart from other chats. Anthropic, which lets subscribers connect records to its Claude chatbot, says it does not use health data to train models. A law firm's analysis notes that HIPAA does not typically apply to either consumer product, and OpenAI's help page says its health product "does not offer a Business Associate Agreement," the contract under which a vendor takes on HIPAA's duties.
The promises carry some legal weight. The Federal Trade Commission can act against a company that breaks them. In 2023 it said the drug-discount company GoodRx had broken a promise never to share users' health information, and GoodRx agreed to a permanent ban on sharing such information for advertising. It also agreed to pay a $1.5 million civil penalty under the commission's Health Breach Notification Rule, for not telling users about the sharing. That rule, written for health apps outside HIPAA, requires notice after a breach, not limits on use. Washington state generally requires a consumer's consent to collect or share health data and a signed authorization to sell it, and it lets consumers sue. OpenAI's privacy notice for health data acknowledges that law and Nevada's, and says it may be updated "from time to time." The commission's staff have warned that quietly loosening such terms for data already collected may be unfair or deceptive.
What a patient gives up, then, is a law that binds whatever a company decides. Under HIPAA a hospital needs the patient's authorization before it sells a record or uses it for marketing, whatever its own policy says. Most Americans do not know where HIPAA's reach ends. In a national survey in 2022, more than four in five adults did not know that health apps can sell users' data to marketers without breaking HIPAA. A patient who links a record believing that HIPAA still applies is consenting on a false premise.
Hospitals, which carry HIPAA's duties, want the gap closed. In a statement to a Senate subcommittee on Sept. 30, the American Hospital Association said it is essential that holders of health information outside HIPAA, including "certain AI vendors," be "subject to similarly rigorous privacy and security standards." A bill to that end is already before the Senate. The Health Information Privacy Reform Act, led by Senators Bill Cassidy, a Louisiana Republican, and Maggie Hassan, a New Hampshire Democrat, would have the department write privacy, security and breach-notice rules within 18 months for companies outside HIPAA that hold identifiable health information, such as health apps. Those rules would have to require a patient's written authorization for any sale, licensing, transfer or marketing of the data, and give patients a right to have it deleted. The Senate health committee approved the bill 22 to 0 in July; it has been on the Senate calendar since August and has no House companion.
Some will call warning patients paternalism under a privacy label. Patients have a legal right to their records, federal rules now largely bar providers from standing in the way, and if a chatbot helps a patient understand a lab result, that is what the access was for. Dr. Byrne herself said: "If using ChatGPT helps patients further understand their health, I want to support it." Hospital warnings, on this view, would scare patients off a useful tool and protect the hospitals' hold on the record. The answer is that nothing here argues for slowing access. The federal office that writes the information blocking rules says providers may tell patients about an app's privacy risks, so long as the information is "factually accurate, unbiased, objective, and not unfair or deceptive" and is given the same way for every app. A patient told that HIPAA stops at the door can still walk through it, and a law that sent the protections along would cost patients no access at all.
Congress should pass the Senate bill, with one change. As reported, the bill requires a patient's written permission for sale, licensing, transfer and marketing, and leaves the department to decide which other uses need it; it mentions training AI models only in rules on how much data the training may use. Congress should not leave training to rulemaking, and should require the patient's written permission for it too. OpenAI's notice says that by default it does not use these records to improve the models that power ChatGPT, and Anthropic says it does not use health data to train models, so the change would cost them little. Acting soon matters because the government is encouraging the traffic: the Centers for Medicare and Medicaid Services lists OpenAI, Anthropic, Google and b.well among early adopters of a pledge to build AI assistants that draw on patients' records with their consent.
Until then, the warning falls to hospitals and physicians. Health systems should make sure that the screen on which a patient authorizes an app says, in one plain sentence, that unless the app's maker is bound by HIPAA, the law will not protect the copy and the company's own policy will govern it instead. Federal guidance on information blocking says such a notice must be used the same way for every app. Physicians whose patients ask about linking should say the same, and pass on what OpenAI's help page says about leaving: disconnecting deletes the synced records within 30 days, but anything already in the chat history stays until the patient deletes it. None of this argues against linking. For many patients the trade will be worth making, and they should make it knowing what it is.
Dr. Byrne saw a disconnect between who benefits from these links and who pays for them. The larger one is between who holds the record and who answers for it. Until Congress closes that gap, patients should hear about it from their doctors before they link.
- Advocate Health's patients and ChatGPTBecker's Hospital Review (Giles Bruce, Oct. 6, 2026) reported that patients at Charlotte, N.C.-based Advocate Health, a 69-hospital system, linking their MyChart accounts to ChatGPT "rose more than 7,000% from July to August," according to Bobbie Byrne, MD, executive vice president and chief information officer. The article gives no number of patients and no dollar cost. Patients "link their accounts on OpenAI's site," and "Advocate did nothing to enable the patient connections"; each linked account and request "generates API calls against Advocate's Epic EHR." Dr. Byrne said Advocate does "not have any real visibility or influence on this day to day"; "There is compute power required for this, so that will be cost to someone"; and "Right now we can absorb the amounts, but I would be concerned if the growth rate escalates." The article says OpenAI launched ChatGPT Health in January and made it broadly available in July.Source: Becker's Hospital Review
- Federal guidance on records sent to appsThe HHS Office for Civil Rights page "The access right, health apps, & APIs" (content last reviewed May 30, 2025) says: "Once health information is received from a covered entity, at the individual's direction, by an app that is neither a covered entity nor a business associate under HIPAA, the information is no longer subject to the protections of the HIPAA Rules." The covered entity then "would not be liable under the HIPAA Rules for any subsequent use or disclosure," though it could be if the app was provided by or for the covered entity. The Privacy Rule "generally prohibits a covered entity from refusing to disclose ePHI to a third-party app designated by the individual if the ePHI is readily producible in the form and format used by the app," whatever its concerns about how the app will use the data. The guidance applies only to the extent consistent with the court's order in Ciox Health, LLC v. Azar.Source: HHS Office for Civil Rights
- Information blocking, and warnings to patientsASTP/ONC defines information blocking as "a practice by an “actor” that is likely to interfere with the access, exchange, or use of electronic health information (EHI), except as required by law or specified in an information blocking exception"; health care providers are among the actors. A final rule set disincentives for providers found by the HHS Office of Inspector General to have committed information blocking. On Sept. 3, 2025, HHS announced a crackdown: providers in certain CMS programs "could be subject to disincentives under those programs," and certified developers and networks could face "a substantial civil monetary penalty of up to $1M per violation." An ONC FAQ (last updated Sept. 12, 2025) says educating patients about the privacy and security risks of apps they choose is not interference if it focuses on current risks, is "factually accurate, unbiased, objective, and not unfair or deceptive," and is provided in a non-discriminatory manner, used the same way for all apps.
- OpenAI's terms for ChatGPT HealthOpenAI announced ChatGPT Health on Jan. 7, 2026, and began launching it to U.S. users 18 and older on July 23, 2026. Its help page (read Oct. 11, 2026) says connected records and conversations that use them "are not used to train our foundation models or target ads"; that Health "does not offer a Business Associate Agreement"; and that after disconnection synced data "is deleted from OpenAI's systems within 30 days," while conversation history "remains until you delete those conversations." Its Health Privacy Notice (updated June 29, 2026) says that "By default" OpenAI does not use connected medical records or Health conversations "to improve our foundational models that power ChatGPT"; that "A limited number of authorized OpenAI personnel and trusted service providers might access data received via Health Features to improve model safety, unless you have opted out in ChatGPT"; and "We may update this notice from time to time."
- b.well, Anthropic and a federal pledgeb.well, through which ChatGPT users connect records, says its network covers "more than 2.2 million providers and 320 health plans, labs, and other sources" and tells those users "Your data isn’t used to train AI models." Anthropic announced on Jan. 11, 2026, that "In the US, Claude Pro and Max plan subscribers can choose to give Claude secure access to their lab results and health records," and said "We do not use users' health data to train models"; a note dated Feb. 7, 2026, says "HIPAA-ready products are specific to healthcare providers and payers." Lawyers at Epstein Becker Green (Jan. 28, 2026) write that HIPAA "does not typically apply to direct-to-consumer applications such as ChatGPT Health and Claude for Healthcare." A CMS page (last modified Aug. 6, 2026) lists Anthropic, b.well Connected Health, Google and OpenAI among early adopters of a pledge to build conversational AI assistants that, "with patient consent, securely access relevant health information."
- Rules for health apps outside HIPAAThe FTC's Health Breach Notification Rule, updated in 2024, "applies to health apps and similar technologies not covered by HIPAA," counts unauthorized disclosures as breaches and requires notice to individuals within 60 calendar days of discovery, but does not limit how companies use data. In its first case under the rule (Feb. 1, 2023), the FTC said GoodRx had "deceptively promised its users that it would never share personal health information" and broke the rule "by failing to notify consumers, the FTC, and the media" about the sharing; GoodRx agreed to a ban on sharing health data for advertising and to "a $1.5 million civil penalty for violating the rule." FTC staff warned in February 2024 that quietly loosening terms for data already collected may be unfair or deceptive. Washington's My Health My Data Act, in force since March 31, 2024, requires consent to collect or share consumer health data, with exceptions, and "valid authorization" to sell it, and allows private suits.
- The Senate billS. 3097, the Health Information Privacy Reform Act, was introduced Nov. 4, 2025, by Sen. Bill Cassidy, R-La., who chairs the HELP Committee; Sen. Maggie Hassan, D-N.H., became a cosponsor on Aug. 5, 2026, and the committee calls the bill "led by" both. The committee voted on July 30, 2026 (22-0, according to Alston & Bird), and the bill has been on the Senate calendar since Aug. 4 (Calendar No. 538), with no House companion and no later action as of Oct. 11. As reported, HHS would have 18 months to set "privacy, security, and breach notification standards" for entities outside HIPAA that process identifiable health information, including "requiring the individual's written authorization for any sale, licensing, transfer, or marketing" and deletion within 30 days of a request; other uses that need authorization are left to the rules. Section 4 covers how the minimum necessary standard applies to data used to "develop, train, validate, or fine-tune" AI models.
- HIPAA's limits and what the public knowsUnder the HIPAA Privacy Rule (45 CFR 164.508), a covered entity "must obtain an authorization for any use or disclosure of protected health information for marketing," with narrow exceptions, and "must obtain an authorization for any disclosure of protected health information which is a sale of protected health information." In a nationally representative survey of 2,014 U.S. adults, conducted in fall 2022 for the University of Pennsylvania's Annenberg School by the University of Chicago's National Opinion Research Center (report dated Feb. 1, 2023), 82% did not know that HIPAA does not stop apps that provide health information, "such as exercise and fertility apps," from "selling data collected about the app users to marketers"; 45% said they did not know.