The First National Rulebook for Hospital AI Just Dropped. Here’s What It Still Doesn’t Say About Us.
By Amber K. McClendon · Founder & CEO, Melanin Bliss Media
Visualize the exam room before you picture the algorithm. A Black woman sits on the table, describing pain that's been dismissed twice already. Out of frame, an AI tool listens to the visit and drafts her note. Another flags her "risk." A third drafts the message she'll read at home. She will never see any of them — but each one is now shaping what her doctor writes, believes, and does next.
For years, those tools ran in your hospital with almost no rules. As of late 2025, that changed. The Joint Commission — the body that accredits more than 23,000 U.S. healthcare organizations partnered with the Coalition for Health AI (CHAI), a coalition of roughly 3,000 health systems and tech companies, to publish the first national guidance on the responsible use of AI in healthcare. A voluntary certification will follow, rolling out through 2026 to those 22,000-plus organizations.
This is a genuinely good and overdue thing. It is also where Health Equity Journalism does its job: read the rulebook the way the people most at risk would have to live with it.
The rules now say AI must be validated but mostly "locally." And "local" is exactly where equity quietly falls through.
What the guidance actually says
Strip away the press language and the first installment rests on a handful of plain expectations. Health systems should set policies for how AI is used. They should do local validation confirm a tool actually works in their own setting, on their own patients, before trusting it. They should monitor performance over time rather than assume it holds. And all of it is meant to be flexible "interpreted and integrated" to fit each organization, at any stage of its AI journey.
CHAI's own leaders are explicit that the goal is to make responsible AI usable in hospitals of every size and resource level not just the wealthy academic centers. That intent is right. The risk lives in the gap between the intent and the word "local."
The part you only catch from inside
Here is what 15 years inside the system teaches you to notice. "Local validation" sounds like rigor and it is. But validation is only as honest as the data and the people doing it. A tool "validated locally" at a hospital whose records underrepresent Black patients, non-English speakers, and Medicaid populations will pass its local check and still fail the very people who already get failed. The bias doesn't announce itself. It hides inside a green checkmark.
"Flexible" carries the same double edge. Flexibility lets a small clinic adapt the rules to its reality good. It also lets a system that doesn't prioritize equity simply… not. Nothing in high-level, voluntary guidance forces anyone to test whether an AI tool performs differently across race, language, or income. It is permitted. It is not required.
And the people most often left holding this work? Researchers who study AI governance at academic medical centers have found equity treated inconsistently, with oversight frequently handled by clinicians on a voluntary basis, squeezed in around full-time jobs. The safety-net hospitals and community clinics serving the most vulnerable patients have largely been left out of the rooms where these standards get shaped at all.
A tool validated only on a hospital's own skewed records will pass its local check and still fail the patients the system was already failing.
Why this is a health story, not a tech story
It is tempting to file AI governance under "IT." Don't. The note an AI drafts becomes the record that follows a patient for years. The "risk score" an AI assigns can decide who gets a closer look and who gets sent home. The patient instructions an AI writes too often at a reading level no one checks decide whether a family actually understands what to do next. When those outputs carry bias and no rule requires anyone to look for it, the harm is not theoretical. It is measured in delayed diagnoses, dismissed pain, and preventable deaths the same disparities, now running at the speed of software.
That is the line this new standard walks right up to and does not quite cross. It tells hospitals to make sure AI works. It does not yet tell them to make sure it works for everyone.
What to do with this if you run, fund, or use one of these systems
For practices and health systems adopting AI:
Add one question to every validation: "Did we test this tool's output across race, primary language, and insurance type not just overall?" If the answer is no, the validation isn't finished.
Write the equity check into your AI policy now before the certification requires it. The systems that build it in early will be the ones ready when the rules tighten.
Don't let governance ride on one volunteer. If equity oversight is somebody's unpaid side duty, it will be the first thing dropped on a busy week.
For patients and families:
You can ask. "Is any part of my note or care plan drafted by AI? Who checks it?" You have the right to know what's shaping your record.
Read your visit summary. If something's wrong or doesn't sound like your visit, say so errors in an AI-drafted note follow you until someone corrects them.
The bottom line
The first national rulebook for hospital AI is here, and it deserves credit for arriving at all. But a standard that asks whether AI works without requiring anyone to ask whether it works equitably leaves the same communities exposed that every prior wave of medical technology left exposed. The fix isn't to slow the standard down. It's to close the gap before it hardens into the norm to make equity a line item in every validation, not a value someone gets to skip.
We'll keep reading the rulebooks the way the people in the exam room have to live with them. That's the work.
We challenge. We investigate. We restore truth. · Journalism for the culture.
Not a brand. A blueprint.
MB AI Strategy is a separate advisory practice operated by Melanin Bliss Media. No organization named in this piece is a current or recent MB AI Strategy client. MB's editorial work is independent and walled from its advisory work.