Two things happened this week. A Baltimore hospital began notifying 3,182 patients after a staff member uploaded their information to a public AI tool. And the federal rule that made certified health software publish a plain sheet about its AI was deleted on October 26, 2026. For a solo practitioner the AI governance lesson is the same in both. The checking nobody else is doing is yours, and this week there is less paper to help you do it.
Signal 1
A Staff Member Put Patient Data Into a Public AI Tool
Johns Hopkins Bayview Medical Center is notifying 3,182 patients about an AI mistake. It happened on September 21, 2026, during an internal review. A staff member uploaded a spreadsheet of patient information to a public AI tool. The hospital had not approved that tool for any use. The spreadsheet covered patients treated between May and September 2026. It held names, medical record numbers, account numbers, admission and treatment dates, the treating department, medications given, and provider names. The hospital says Social Security numbers and financial details were not in it. The worker deleted the prompt, the conversation and the file from the tool's history. The hospital also asked the company that runs the tool to delete the data. Letters went out to patients on October 22, 2026, offering twelve months of credit monitoring.
Sources: WMAR-2 News, October 23, 2026 | Baltimore Sun, October 24, 2026
What this means for you
This is the one to sit with. A large hospital has an approved software list, a privacy office and a compliance team. One person still pasted patient data into a tool that was not on the list. You do not have any of that apparatus. You have yourself, a laptop and a full day. Nothing about a smaller practice makes this mistake less likely, and the notification afterward would be yours to make and yours to pay for. Notice also what did not cause this. The AI tool did not fail. No tool with better training data would have stopped it. The whole exposure came from one ordinary decision, made quickly, by somebody doing their job.
Signal 2
The AI Model Card Is No Longer Required Anywhere
On October 26, 2026, the federal office that certifies health record software issued a rule called HTI-5. It removes 14 of the 60 requirements software must meet to be certified, and changes seven more. One of those changes deletes the AI model card. A model card is a short standard sheet. It states what an AI feature was trained on, what it was tested on, and how its performance is checked. Certified software had to publish one. Now it does not. The office had proposed cutting 34 requirements. After comments came in it kept the privacy and security ones and said it would look at those later. It did not keep the model card. The HHS announcement led with 1.5 billion dollars in saved costs and did not mention the model card at all. Thirteen of the cuts take effect now and seven more on January 1, 2027.
Sources: HHS press release, October 26, 2026 | American Hospital Association, October 26, 2026 | STAT, October 26, 2026 | Federal Register, HTI-5 proposed rule, December 29, 2025
What this means for you
Here is the plain version. The one document that told you how a health AI tool was built is now required of nobody. A vendor who will not say what their tool was trained on is fully within the rules. That lands on you, because you are still the person who answers for what the tool does in your practice. If a client asks whether your intake form or your note-taker is safe, there is no official sheet to point at. There is also no standard to measure the answer against. What you get is whatever the company says, in whatever form they decide to say it.
The Pattern
Put the two together and the direction is easy to read. On one side, the paperwork a vendor owes you is getting shorter. The certification office proposed cutting 34 requirements on December 29, 2025. Comments closed on February 27, 2026. The final rule landed on October 26, 2026 and cut 14 of them, the AI model card included. On the other side, the thing that actually goes wrong has not changed at all. A spreadsheet went into a public AI tool and 3,182 people received letters. So there is less written down about how these tools work, while the failure that costs real money stays exactly what it always was. For a practitioner nobody is governing, that means two jobs are now yours and neither one is optional. You decide whether a tool is good enough to use. You decide what never gets typed into it. No agency is doing either of those for you, and after this rule there is less public information to do the first one with.
What To Watch
North Carolina Senate Bill 963 would require a license to run a health chatbot. Section 114B-3(a) says no person shall operate or distribute a chatbot that deals substantially with health information without first obtaining a license. Operate is the word that reaches you. A licensee would carry professional liability insurance and run security audits every six months. It would report a breach to the state within 24 hours and to affected consumers within 48. It would also have to show the tool works through peer-reviewed controlled trials. There is no exemption for a small practice, so an intake chatbot on a one-person website sits inside that text as written. The licensing part takes effect January 1, 2027. The bill was filed April 30, 2026 and has not moved since May 4. Watch which version of the duty survives. The licensing section reaches whoever operates the chatbot. The enforcement and inspection sections are written around manufacturers and importers, and around places where chatbots are made or held. Both readings are in the same bill. Which one survives amendment decides whether this lands on the companies or on the practices using them. Source: https://www.ncleg.gov/Sessions/2025/Bills/Senate/PDF/S963v1.pdf
New Jersey Assembly Bill 4733 would bar advertising a generative AI system as able to practice a regulated profession. The professions are the ones licensed under Title 45 of the Revised Statutes. A violation is an unlawful practice under the consumer fraud act. The duty lands on the deployer as well as the developer. Your own description of your own tool is what would create the exposure. The bill has sat in Assembly Appropriations since May 28, 2026, and its fiscal estimate was issued on August 6. Watch it because it now has two vehicles. An identical Senate bill, S 4088, was introduced on May 4 and sits in Senate Commerce. It can move from either chamber. Sources: https://www.njleg.state.nj.us/bill-search/2026/A4733 and https://www.njleg.state.nj.us/bill-search/2026/S4088
One Thing You Can Do This Week
Write your own list of which AI tools are allowed to see client information and which are not. Two columns, one page. In the first column put the tools you actually use for client work. In the second put everything else, including the general chat assistant you keep open in another tab. Add one line at the bottom: client information never gets pasted into anything in the second column. Keep the page with your other practice paperwork. The hospital in this issue had a list like this and somebody went around it, so the page is not a guarantee. It helps because once the list exists, pasting stops being a reflex and becomes a decision you have to make on purpose.