Five healthcare AI news stories making headlines this week: Johns Hopkins notifies 3,182 patients after a staff member uploads patient data to an unapproved AI tool; two Australian nurses answer to their regulator for an AI documentation error; Ohio Medicare patients wait months under an AI prior authorization pilot; a study finds AI models split on borderline test decisions; and Microsoft publishes diagnostic AI results in Science. The common thread is accountability for what AI gets wrong.
On Sept. 21, a staff member at Johns Hopkins Bayview Medical Center uploaded a spreadsheet of patient data to a public AI tool the hospital had not approved. The hospital learned about it the same day. A month later, 3,182 patients are getting letters in the mail.
The same week, Australia's nursing union told a parliamentary committee about two nurses who spent months answering to their regulator for an error made by an AI system they were required to use. In both cases, the people who paid for the mistake were not the ones who chose the technology or set its limits.
That pattern runs through the rest of this week's news, from Medicare's AI prior authorization pilot to new research on how much AI models disagree with each other, and Microsoft's latest diagnostic results.
The Bayview employee uploaded the spreadsheet while doing an internal review, Becker's Hospital Review reported. It held admissions and treatment information from May through September 2026, including patient names, medical record numbers, medications given and provider names. It did not include Social Security numbers or financial account information.
Bayview had the employee delete the prompt, thread and file from the tool's history, then asked the company behind the tool to delete its copy. Letters went out Oct. 22 with 12 months of free credit monitoring. The hospital said it is reinforcing policy and training on unauthorized AI tools. It did not name the tool.
Why it matters: Hopkins caught this within a day, and the data had still left. Policy and training shape what staff intend to do. They don't act at the moment someone submits PHI to an AI tool, which is exactly where shadow AI risks turn into notification letters. Stopping that requires policy enforcement in real time, at the point of use.
Annie Butler, federal secretary of the Australian Nursing and Midwifery Federation, told a parliamentary committee on AI that an AI-assisted documentation system listed two registered nurses as nurse practitioners. That is a separate registration category in Australia, and holding yourself out as one is an offense. The nurses saw the error and tried to fix it, but they had no control over how the system saved their entries.
They were reported to the national regulator, which eventually accepted that the error was outside their control. That took months. The union now wants Australia's Fair Work Act amended so workers can refuse AI tools they consider unsafe without penalty. In a September submission, it warned that clinicians risk becoming "liability sinks" for systems they never chose.
Why it matters: The employer picked the tool and required its use. The nurses carried the license risk. US health systems rolling out AI documentation tools have the same setup: if the tool records something wrong and staff can't correct it, the error becomes part of the chart with a clinician's name on it. Governance needs a clear path to flag and fix bad output, plus continuous monitoring of what these tools write once they are live.
Some Ohio Medicare beneficiaries waited months for care under the Wasteful and Inappropriate Services Reduction (WISeR) model, The Columbus Dispatch reported, citing records the Electronic Frontier Foundation obtained through a public records lawsuit. WISeR launched in January in six states. It uses AI-driven prior authorization for 13 services that never needed it before, including certain nerve stimulators and skin substitutes for chronic wounds.
Ohio's vendor, Innovaccer, warned CMS on Dec. 12, 2025, that its system would not be fully operational at launch. The program went live anyway, and Innovaccer approved every request automatically to avoid a backlog. Some providers still waited more than a month and couldn't reach the company. According to Becker's ASC, two vendors denied nearly 6,000 claims in the first three months. CMS says the issues are resolved and is considering adding cardiac catheterization, pacemakers, MRIs and specialized cancer care to the model.
Why it matters: A vendor said its AI wasn't ready, and the launch date held. Health systems make that same call with their own AI vendors, usually under deadline pressure. Go-live readiness belongs in governance, with written criteria and someone who has the authority to delay. And the first months after launch need the closest watch, because oversight that stops once vetting is done misses the problems that only show up in use.
Researchers from Ben-Gurion University of the Negev and Clalit Health Services published a study in Annals of Internal Medicine that tested six large language models from Google, Anthropic and OpenAI on scenarios built from Choosing Wisely recommendations. Each scenario paired a case where a test or treatment was unnecessary with a close match where it was needed.
Overall, the models did well. They withheld needed care in 0.6% of cases and recommended unneeded care in 5.4%. Most of the errors came in areas where medical guidelines already conflict, such as routine EKGs for low-risk patients and ultrasounds for small thyroid nodules. On those same borderline cases, one model recommended an unnecessary test less than 10% of the time. Another did so 80% of the time. The authors also noted that models tend to go along with the person they're talking to, and plan to study what happens when patients push back.
Why it matters: Two tools built on different models can give different advice about the same patient. A vendor can also swap the model inside a product during a routine update, and the advice may shift with it. Health systems need visibility into which models sit behind the AI tools their staff use, and a way to know when that changes.
Microsoft published peer-reviewed results in Science for its Diagnostic Orchestrator, a system that coordinates several AI agents to work through a case the way a panel of physicians might. On New England Journal of Medicine case records, it reached 83% accuracy, compared with 36% for generalist physicians who could use reference materials. On emergency department cases, it scored 80% against a physician average of 47%.
The paper is candid about its limits. All 15 authors work for Microsoft. The cases were historical, and the comparison was against individual generalists, not specialists or care teams. Microsoft AI CEO Mustafa Suleyman told The Verge the system still needs clinical validation, and Microsoft plans a prospective study with Beth Israel Deaconess Medical Center.
Why it matters: Diagnostic AI is moving from benchmarks toward hospital pilots, and Microsoft is bringing its Dragon Copilot teams into the same group. A high score on published cases says little about how a tool will perform on one hospital's patients, or after its underlying model is updated. Validating locally before use and treating AI risk as patient safety risk after go-live will still fall to health systems.
Every story this week comes back to the moment after AI gets something wrong, and who is left holding the result. Five questions for health system leaders to put to their teams this week:
Vitea is an AI governance platform built specifically for healthcare. Command enforces policy in real time, inspecting prompts and responses and blocking PHI and unsafe output at the point of use. Lens shows where AI is running across the organization, and Pulse keeps watching tools after go-live, including when vendors change them.
Talk with our team about putting real-time enforcement in front of every AI tool your staff use.
Follow Vitea on LinkedIn for weekly healthcare AI governance news.
.png)
.png)
.png)
.png)
