Shadow AI just hit Johns Hopkins — and AI governance is now your job

Dr. Ashley Hussain-Okorafor, DBA
Former lecturer and hiring-side healthcare administrator

Here's the short version: a Baltimore hospital employee pasted a spreadsheet of 3,182 patients' records into an AI tool nobody had approved, and a few weeks later the federal rule that was supposed to make a hospital's AI transparent quietly got rolled back. Put those two things together and they say one thing: AI governance just became a healthcare administrator's job, not an IT department's.

I'll break it down, because this is the first AI story I've seen this year that's actually about you — the person who's going to run the operation — and not about a robot reading an X-ray.

What actually happened in Baltimore

On September 21, a worker at Johns Hopkins Bayview Medical Center was doing an internal review and uploaded a spreadsheet to a public AI tool the hospital had never approved. That spreadsheet held admission and treatment data on 3,182 patients treated between May and September — names, admission and treatment dates, medical record numbers, account numbers, the departments that treated them, medications, and provider names.

To the hospital's credit, it caught the upload the same day, had the employee delete the prompt and the document, asked the tool's operator to purge the data, and started mailing letters to patients on October 22 with 12 months of free credit monitoring. Social Security numbers and financial account data were not in the file. You can read the hospital's own account via Becker's Hospital Review.

Here's the part that should stop you cold: there was no attacker. No hack, no breach in the way you're picturing it. A staff member was just trying to work faster, and the easiest tool to reach was one IT never cleared.

That's the "shadow AI" problem, and it's everywhere

The industry has a name for this: shadow AI — staff using ChatGPT or some other public tool that IT, security, and compliance haven't approved, because it's faster than whatever the hospital officially gave them. In a hospital, that usually means pasting patient records, transcripts, or a spreadsheet into a chatbot to save an hour.

One employee, one paste, 3,182 people getting a breach letter. That's the whole problem in miniature — and most organizations can't even tell you how often it happens, because the tool is public and the upload leaves the building in a single click.

Now the safety rail quietly came off

Here's where it gets serious for your career. On October 28, the federal government published a rule called HTI-5, and buried in it is a change that directly affects every hospital's AI. The Office of the National Coordinator for Health IT finalized the removal of the "model card" requirements for predictive AI — the source-attribute rules that made the AI built into an electronic health record carry something like a nutrition label: who trained it, on what data, how it performs across different patient groups, and its known risks.

That transparency requirement is gone, effective December 28, 2026. ONC removed 14 certification criteria in all and revised seven, in what it calls a deregulatory move to "unleash prosperity." I'm not here to relitigate the politics. I'm here to tell you what changed for the job you want: the government is no longer guaranteeing you a label on the AI inside your hospital's software. If you want to know what a predictive tool is trained on, or whether it performs worse on certain patients, you now have to get that answer yourself — from your EHR vendor, through a contract, through your own oversight.

Why this lands on the administrator's desk

And here's the kicker: while ONC removed the transparency requirements, it did not remove the hospital's legal duty not to discriminate. Under Section 1557 of the Affordable Care Act, a provider still has to identify and reduce bias in the predictive tools it uses on patients. So the obligation to know what your AI is doing didn't go away — only the guarantee that someone will hand you the information to check it.

That gap? It's not a technology problem anymore. It's an operations and leadership problem. Somebody has to decide which tools are approved, what happens when a patient's data gets pasted into the wrong one, and how to tell — before the breach letter goes out — that it's happening. That somebody is a healthcare administrator.

What this means for the job you're training for

Let me translate this into your résumé, because that's the part most news sites skip.

Most people applying for healthcare admin jobs right now are polishing the same generic résumé and hoping. Meanwhile, systems are quietly discovering they need someone who can govern AI — and almost nobody has that vocabulary yet. That's an opening, and it's wide.

Here's what "AI governance" looks like when it's your job, in plain English:

• An approved-tool list people will actually use. If staff have no sanctioned way to summarize or draft, they'll find an unsanctioned one. Your job is to make the right tool the easy tool.
• Vendor terms, in writing. Does the AI company train its models on your data? Will it delete on request? Will it sign a business associate agreement? That's contract language, and it's an administrator's job to read it.
• Detection, not just policy. A policy nobody enforces is a hope, not a control. You need to know when patient data is headed toward a public AI site.
• An incident playbook. What happens the day someone pastes the wrong thing? Who calls the vendor, who runs the privacy assessment, who decides whether 3,182 letters go out?

None of that requires you to be a data scientist. It requires you to understand operations, compliance, and contracts — three things a healthcare administrator already owns.

What I'd do if I were starting today

Three moves, and you can start all of them this week.

First, learn what shadow AI is and be able to explain it in one sentence — because in your next interview, "how would you keep patient data out of unauthorized AI tools?" is exactly the kind of question that separates the people who've been paying attention from the people who haven't.

Second, learn the vocabulary: model cards, source attributes, Section 1557, business associate agreement. You don't need to be an expert. You need to be the person in the room who knows these are the questions to ask.

Third, reframe AI in your own head. It's not something coming for your job. It's a new responsibility being created — and the people who volunteer to carry it are the ones who get promoted while everyone else worries about the robots.

This week, AI governance stopped being a tech story. It became a healthcare administration story. That's the door. Walk through it.