Not the doctor's to sign
Free AI tools for doctors come with a HIPAA contract one doctor can accept alone. A hospital's records need the hospital's signature, and the companies should make the safe path the default.

Listen to this special topic
read by an AI voiceMonument Health decided to block OpenEvidence, STAT reported on Nov. 5. The popular tool answers clinicians' medical questions. By STAT's account, the decision drew a few objections at the health system, which is based in South Dakota and Wyoming. One came from a doctor who had been storing more than a hundred patient records in OpenEvidence. Patrick Woodard, Monument's chief information officer, told that doctor it would have to end, the report says.
The tool all but invites those records. OpenEvidence announced in April 2025 that it was “fully HIPAA compliant.” Covered entities, the law's term for the doctors, clinics and others it binds, could now enter patients' protected health information under the company's business associate agreement, or BAA. That is the written contract the federal privacy law requires with a company a provider engages to handle its records. It is where the trouble starts. Patient details should go only into AI tools covered by an agreement the clinician's own organization has signed. And the companies that give these tools away should make the safe path the default, rather than leave it to a doctor between patients.
Under OpenEvidence's terms, a covered entity that chooses to send patient information agrees, by sending it, to have it handled under the company's agreement. Its rivals offer much the same. Doximity's agreement takes effect on the day a provider registers as a member. OpenAI made ChatGPT for Clinicians free to verified physicians in April, with what it calls “optional support for HIPAA compliance” through an agreement for eligible accounts. OpenEvidence's own profile, printed in a Mount Sinai release, says it alone is trusted by hundreds of thousands of verified clinicians.
That agreement can mean less than it seems. HIPAA lets a covered entity give records to a contractor once it has written assurances that the contractor will safeguard them. The covered entity must get them. The federal health department's examples of such contractors include a vendor's chatbot on a provider's patient portal. A doctor who runs her own practice and bills insurers electronically is a covered entity, and for her that agreement may be enough. A doctor on a hospital's payroll is handling the hospital's records. A MaineHealth notice about the tool says that “such individual BAAs do not cover the sharing of MaineHealth data with external parties.” Federal guidance on cloud companies is blunter. A covered entity that stores patients' electronic records with one, without an agreement, “is in violation of the HIPAA Rules.”
When records go somewhere a hospital never approved, its patients carry the risk. In September a worker at Johns Hopkins Bayview Medical Center uploaded a spreadsheet of patient information to a public AI tool the hospital had not approved. It may have held names, medical record numbers and the medications patients were given, among other details. The worker deleted the chat. Hopkins asked the company that runs the tool to delete the data. Hopkins did not answer questions from Straight Arrow News about which company was involved or whether the deletion had been confirmed. About 3,200 patients have been sent letters and offered a year of free credit monitoring, for a spreadsheet somebody else uploaded.
OpenEvidence knows where the danger lies. Its own fine print says so. The terms recommend that users remove “any information that identifies an individual” before they submit anything, from names and insurance numbers to photographs of tattoos or scars. “You shall be solely responsible for any claims arising from your failure to de-identify,” they tell the doctor. And anyone who does enter a patient's details confirms having all necessary consents and authorizations from the patient, “your employer, or other relevant parties.” The safe path is already written down. Following it has been left to a physician with a full clinic.
Doctors who rely on these tools will say a block does more harm than any contract could. The tools help with real decisions. A block does not stop the questions. Wolters Kluwer Health, which sells clinical decision support of its own, paid for a survey of 518 hospital and health system staff. In its results, nearly one in five said they had used an unauthorized AI tool, often because it was faster. Block the tool, on this view, and the questions move to personal phones. They have a point. A ban with nothing behind it fails doctors. But that is a reason to buy the tool properly, not to settle for the click. Northwestern Medicine has signed an agreement with OpenEvidence. MaineHealth has tried, and “thus far has not come to acceptable terms with them.” And many questions need no patient's name at all, as OpenAI's own launch post points out.
Health systems should move first. Each one whose doctors use a free tool can sign with its maker or name an approved alternative, then tell clinicians in one plain sentence what may go in the box. MaineHealth already puts its warning on a page that users acknowledge by proceeding to OpenEvidence. Medical staffs should ask for the same. The companies have the easier fix. An account not tied to an agreement signed by the clinician's organization should catch names, record numbers and insurance numbers before a question leaves the screen, strip them out and say why. A company that already tells doctors to remove identifiers could do it for them, rather than leave it to a tired person at the end of a shift.
The doctor at Monument was told to stop after objecting to the block. The patients at Hopkins learned what had happened from a letter. In both places, someone noticed after the records had gone in. A box that takes patients' details should ask whose agreement covers them before it accepts the first name.
- Monument Health and OpenEvidenceKatie Palmer reported in STAT on Nov. 5, 2026 that Monument Health, a health system based in South Dakota and Wyoming, had decided to block access to OpenEvidence, “the popular free artificial intelligence tool,” and that Patrick Woodard, Monument's chief information officer, had gotten a few objections. One came from a doctor, described as a super user, who was storing more than a hundred patient records in the tool; the chief information officer told him that would have to end. STAT reports that free clinician chatbots, including OpenEvidence, Doximity Ask and ChatGPT for Clinicians, claim to be compliant with HIPAA, and that hundreds of thousands of doctors use them. Beyond its opening, the article is available only to STAT+ subscribers.Source: STAT, Nov. 5, 2026
- OpenEvidence's HIPAA announcement and termsIn an announcement datelined Cambridge, Mass., April 25, 2025, OpenEvidence said it was “fully HIPAA compliant” and that U.S. covered entities could securely input protected health information; covered entities that choose to do so agree that it will be handled under the company's Business Associate Agreement. Its terms of use (last updated May 6, 2025) incorporate that agreement by reference. They recommend that, before submitting content, users remove information that identifies or could reasonably enable identification of an individual, such as a name, e-mail address, social security number, insurance number, biometric identifiers, facial photographs or photographs of identifying marks such as tattoos or scars. They make the user solely responsible for claims arising from a failure to de-identify or to obtain appropriate authorizations and consent, and a user who provides such information confirms having all necessary consents and authorizations from the patient, the user's employer or other relevant parties. The announcement says user conversations are private by default.
- Doximity's and OpenAI's agreementsDoximity's business associate agreement is between Doximity, Inc. and the healthcare provider that has agreed to its Terms of Service, and is effective as of the date the provider agrees to those terms by electronically registering as a Doximity member. It covers the tools Doximity makes available to the extent Doximity identifies them as appropriately secure for use with protected health information, and it refers separately to enterprise agreements. OpenAI announced ChatGPT for Clinicians on April 22, 2026, free for any verified physician, nurse practitioner, physician assistant or pharmacist, starting in the United States. Under the heading of optional support for HIPAA compliance, its post says many clinical tasks don't require protected health information but that, if needed, HIPAA support is available through a business associate agreement for eligible accounts. It says conversations are not used to train models, and that ChatGPT for Healthcare allows organizations to deploy ChatGPT to clinicians, administrators and researchers.
- What the HIPAA rules and guidance sayThe Department of Health and Human Services treats doctors and clinics as covered providers only if they transmit information electronically in connection with a transaction for which it has adopted a standard; a covered entity that engages a business associate must have a written contract or other arrangement with it (page last reviewed Aug. 21, 2024). The business associate guidance (last reviewed Oct. 21, 2026) permits disclosure if the covered entity obtains satisfactory assurances, in a contract or other written arrangement, that the information will be appropriately safeguarded; its examples of business associates include a third-party vendor's chatbot on a provider's patient portal that handles patients' information for symptom assessment, medical reminders and appointment scheduling. Cloud computing guidance from the Office for Civil Rights (last reviewed Dec. 23, 2022) says a cloud provider holding such information for a covered entity is a business associate even if it stores only encrypted data and lacks the key, and that a covered entity using one without an agreement is in violation of the HIPAA Rules.
- Three health systems and OpenEvidenceMaineHealth's undated page on the tool is headed “Do NOT enter PHI or other sensitive data into OpenEvidence.” It says MaineHealth does not have a business associate agreement with OpenEvidence; that although OpenEvidence advertises itself as HIPAA-compliant when a user agrees to its agreement, individual agreements do not cover the sharing of MaineHealth data with external parties; that MaineHealth has tried to negotiate one but has not come to acceptable terms; and that proceeding to OpenEvidence from the page acknowledges the notice. An undated Northwestern Medicine physician forum page says OpenEvidence has signed a business associate agreement with Northwestern Medicine, advises clinicians to use discretion and provide PHI only when it is necessary to finding relevant information, and says NM single sign-on will be added in the coming months. Mount Sinai announced on March 31, 2026 that OpenEvidence would be accessible from within its Epic record for physicians, registered nurses and pharmacists.
- The Johns Hopkins Bayview uploadWMAR reported from Baltimore on Oct. 23, 2026 that Johns Hopkins Bayview Medical Center had begun notifying about 3,182 patients after a worker uploaded a spreadsheet of patient information to a public AI tool the hospital had not approved. The details may have included patient names, dates of admission and treatment, medical record numbers, account numbers, the department that treated them, medications given and the names of their providers. The hospital said it discovered the problem on Sept. 21, 2026. The worker deleted the prompt, the thread and the document from the tool's history, and the hospital asked the company that runs the tool to delete the data. Straight Arrow News (Devin Pavlou, Oct. 26, 2026; the outlet closed on Nov. 6) reported that the upload was found the same day it happened, that Hopkins is offering a year of free credit monitoring, and that Hopkins did not answer its questions about which company was involved or whether it confirmed the deletion.
- Unapproved AI tools in hospitalsWolters Kluwer Health released on Jan. 22, 2026 a survey conducted online by CITE Research for Wolters Kluwer Health in December 2025, of 518 healthcare professionals evenly divided between providers and administrators. Forty percent had encountered unauthorized AI tools in the workplace and nearly 20 percent admitted using them. Half pointed to a need for faster workflows. Healthcare Dive, describing more than 500 respondents at hospitals and health systems, reported that more than 50% of administrators and 45% of care providers cited a faster workflow, and that about a quarter of providers and administrators ranked patient safety as their top concern about AI in health care.