Johns Hopkins Bayview incident shows what one unapproved AI upload costs a practice

One spreadsheet, one consumer AI tool and 3,182 patient notification letters. That is the tally from a Sept. 21 incident at Johns Hopkins Bayview Medical Center in Baltimore, and it is a concrete look at what shadow AI costs a health care organization once the data leaves the building.

A staff member uploaded a spreadsheet of patient information to a public AI tool the hospital had not approved, according to WMAR-2 News. The file covered patients treated from May through September and included names, admission and treatment dates, medical record numbers, account numbers, the treating department, medications and provider names. Social Security numbers and financial account information were not included, the hospital said.

An internal review caught the upload the same day. The employee deleted the prompt, the thread and the document from the tool, and Hopkins asked the AI company to delete the data and add safeguards. Letters began going out Oct. 22 with an offer of 12 months of free credit monitoring.

Hopkins has not named the tool or confirmed that the vendor actually erased the file, Straight Arrow News reported. Deleting a chat clears it from the user's history, but that does not necessarily remove it from the company's servers.

In a statement, Johns Hopkins said it is "committed to maintaining the privacy and security of its patients' information" and is strengthening policy and staff training on which AI tools are approved, The Baltimore Banner reported.

Smaller practice, same obligations

A five-physician practice would face a smaller mailing, not a smaller rulebook. Under the HIPAA Breach Notification Rule, an impermissible disclosure is presumed to be a breach unless a documented risk assessment shows a low probability the information was compromised. That assessment weighs four factors: what data was involved, who received it, whether it was actually acquired or viewed and how far the risk was mitigated. An upload to a consumer AI vendor with unknown retention practices is hard to argue down on the second and third factors.

If it is a breach, the practice has no more than 60 days from discovery to notify each affected patient in writing. A breach affecting 500 or more people also requires notice to HHS within that window and, when more than 500 residents of a state are involved, to prominent media outlets serving the area. Breaches affecting fewer than 500 people go into a log reported to HHS within 60 days after the end of the calendar year. State breach laws can add their own deadlines and attorney general notices on top of that.

Notification is only half the exposure. HHS guidance on cloud services says a service that receives or maintains electronic protected health information is a business associate even if it cannot view the data, and a covered entity that uses one without a business associate agreement is in violation of HIPAA. An employee's personal chatbot account comes with no such agreement.

What an acceptable-use policy needs to say

The Hopkins employee is not an outlier. In a Wolters Kluwer Health survey of 518 hospital and health system providers and administrators fielded in December 2025, 17 percent admitted using unauthorized AI tools at work. Among providers who had, 45 percent said faster workflow was the main reason.

Governance has not kept pace. Only 27 percent of physicians and nurses said they knew of formal AI governance policies at their organizations, up from 21 percent a year earlier, according to Wolters Kluwer's 2026 Future Ready Healthcare survey, which Ipsos fielded in March among 203 physicians and 152 nurses. Medical Economics recently reported that staff often turn to outside tools when a practice's approved AI hits its usage limit.

A written policy that holds up after an incident like this one should cover a few basics:

• A named list of approved tools. Staff should not have to guess. Every tool on the list should have a signed business associate agreement.
• A flat ban on patient data in anything else. That includes personal accounts, browser extensions and free tiers of tools the practice pays for elsewhere.
• Technical blocks, not just rules. Restrict consumer AI sites on practice devices and networks, and give staff an approved alternative so the work still gets done.
• An immediate reporting duty. Same-day discovery let Hopkins move fast on containment, which also feeds the mitigation factor in the risk assessment.
• Documented training. Sign-off records show regulators the policy existed before the mistake, not after.

For practices that already have a policy on paper, Physicians Practice has a checklist on proving security and AI controls actually work.

Hopkins did not say whether it has reported the incident to HHS, according to Straight Arrow News. With letters already in the mail, the 60-day federal clock for a breach of this size runs out in late November.