Imagine that the government announced tomorrow that packaged food no longer needs a Nutrition Facts panel. The cereal would be the same cereal. The sugar would still be in it. What would change is who has to find out how much. Before, the manufacturer printed it on the box because the law said so. After, every shopper, school cafeteria and hospital kitchen that cares would have to ask, and the manufacturer could decide how to answer.
On October 26 the federal health IT office did something similar for artificial intelligence. Its HTI-5 rule, published on October 28 and taking effect on December 28, removes the requirement that certified electronic health record software provide what amounted to nutrition labels for predictive AI: disclosures about how a model was trained, what it is for, how it performs and what risks it carries, along with proof that the vendor manages those risks. The rule has received almost no mainstream attention.
My argument is that provider organizations should treat HTI-5 not as relief and not as a scandal, but as a transfer. The question the label answered still needs answering. The only change is that a label printed by law must now become a term written into a contract. Organizations that make that switch in the next few months will lose very little. Organizations that wait will discover that the AI in their clinical software has quietly become something they are responsible for and know nothing about.
Model Facts
Predictive decision support · per certified module
Source attributes required under 45 CFR 170.315(b)(11)
Not required after Dec 28rewrite it here
EXHIBIT C
AI feature disclosures · to master services agreement
What the rule did and did not do
HTI-5 removes 14 of the 60 criteria that health IT products must meet to be certified, and revises seven more. That is far fewer than the 34 removals originally proposed, which tells us the final rule was negotiated down. HHS estimates the savings at about $1.55 billion in present value, or roughly $108 million a year, mostly in reduced compliance work for software developers. Thirteen changes take effect on December 28 and seven more on January 1.
The piece that matters most for anyone deploying AI is the change to the decision support criterion. Certified products no longer have to expose source attributes for predictive tools or apply risk management to them. The regulators said they might restore older transparency requirements for evidence-based, non-AI decision support. They said nothing similar about predictive AI.
Just as important is what survived. The privacy and security criteria, including multifactor authentication and audit reports, stay for now, though the regulators plan to revisit them. The rules against information blocking, which forbid interfering with the flow of patient data, were left untouched. And the whole program is being reset around FHIR, the modern standard for health data interfaces, with future rules promising interfaces that can write data as well as read it. The regulators even named Model Context Protocol, the emerging standard for connecting AI tools to data, as a possible future route.
Why AI needs a label at all
To see what was lost, it helps to ask what labels are for. In 1970 the economist George Akerlof published a paper about used cars that later helped win him a Nobel Prize. When sellers know more about quality than buyers do, he showed, buyers assume the worst and pay accordingly, good sellers leave the market, and quality falls. He called it the market for lemons. Disclosure rules, warranties and brand reputations exist largely to close that gap.
Economists later sorted goods by how hard the gap is to close. Phillip Nelson distinguished search goods, whose quality you can judge before buying, from experience goods, which you can judge only after using them. Michael Darby and Edi Karni added a third and most troublesome category: credence goods, whose quality you may never be able to judge, even after use. A dental chair is a search good. A practice management screen is an experience good; a week of use tells you whether it is clumsy. An AI model that flags early decay on a radiograph is a credence good. If it misses lesions more often on certain images or certain patients, no clinician using it day to day is likely to notice. The errors look exactly like the absence of disease.
This is why the model card mattered. For a credence good, the only practical defense is information supplied by the party that has it, which is the developer. Take that disclosure away and the buyer faces Akerlof's problem in its purest form: a product whose flaws are invisible in use, sold by someone with every reason to describe it favorably.
The label was thinner than it looked
Before mourning the label, though, a dental or specialty provider should check whether it ever had one. The decision support requirements applied only to certified health IT, and only to predictive tools the developer supplied through its certified product. Much dental practice software was never certified, or was certified only for older editions. Anyone can check on the government's Certified Health IT Product List. For many dental service organizations, the honest summary is that HTI-5 removes a benchmark more than a protection. That makes the case for a self-written standard stronger, not weaker: if the federal label never reached your software, the contract was always the only label you were going to get.
Where the duty went
Deregulation rarely deletes an obligation. More often it changes who holds it. Several duties that touch clinical AI were untouched by HTI-5, and most of them sit with the provider.
The most important is Section 1557 of the Affordable Care Act. Its regulation at 45 CFR 92.210 requires covered providers, which include any organization receiving federal funds such as Medicaid payments, to make reasonable efforts to identify and mitigate the risk of discrimination in the patient care decision support tools they use. The law firm McGuireWoods notes that the regulators reminded providers of exactly this duty in the HTI-5 rule itself. You cannot identify a tool's discrimination risk without knowing what it was trained on and how it performs across groups, which is to say, without the model card. HIPAA's security rule still requires a risk analysis of any system that handles patient data, AI features included. States are adding duties too: Colorado's rewritten AI law, effective January 1, 2027, requires HIPAA-covered providers there to give patients a general notice of their use of advanced technologies. And the FDA still reviews AI that functions as a medical device, which covers many radiograph detection tools but not most treatment suggestion or scheduling models.
Inter-office · Do not seal
Unchanged: vendor still certifies MFA and audit reports. FDA still reviews device AI. HIPAA risk analysis still yours.
The rule crossed out the vendor.
It never crossed out the question. That came down the hall to you.
The legal scholar Guido Calabresi offered a useful principle for deciding who should bear a burden: assign it to the party that can avoid the harm, or produce the information, most cheaply. For an AI model, that party is plainly the developer, who already holds the training data, the validation results and the error analysis. Writing those disclosures into a contract does not invent a new burden. It returns the work to the party that can do it for almost nothing, at the moment the buyer has the most leverage, which is at signing and renewal.
The best case for HTI-5
A fair argument has to take the other side seriously, and it has a strong one. The legal scholars Omri Ben-Shahar and Carl Schneider spent a book, More Than You Wanted to Know, documenting how mandated disclosures fail: privacy policies, loan terms and consent forms that almost nobody reads. Model cards may have been another example, produced at real cost and filed unread. HHS's savings estimate is not trivial, the FDA already reviews the highest-risk imaging tools, and a certification checklist can slow innovation by forcing every product into one template.
I think the critics are largely right about mandates, and that is precisely why the contract approach is better rather than worse. Disclosure fails when it is aimed at people who did not ask for it. A provider organization negotiating a renewal is the opposite: a sophisticated buyer who asked, who has a legal reason to read the answer, and who can tailor the questions to the tools it actually uses. Disclosure written on demand into a contract is the version of disclosure that the research suggests actually works.
Proportion, not paperwork
None of this requires a large team. The trick is to match the scrutiny to the consequence, so that the full set of questions goes only where the stakes are highest.
Features that influence clinical decisions without FDA review, such as treatment suggestions or risk scores, deserve the full disclosure package and a clinician's sign-off before they are switched on. FDA-cleared detection tools deserve a check that the clearance covers the way you use them, plus subgroup performance data, since clearance does not guarantee equal accuracy across your patients. Administrative tools, such as no-show prediction or coding assistance, need a lighter review and a place in the HIPAA risk analysis. Most organizations will find that the top-left box holds only a handful of features, and that is where nearly all of the effort should go.
The other half of the rule
It would be a mistake to read HTI-5 only as a loss. The reset around FHIR, with future interfaces that write as well as read and a nod to Model Context Protocol, points toward a world in which AI tools plug into health data through open standards rather than through each vendor's private add-ons. For a provider organization building its own data platform, that is good news: it lowers the cost of using the AI you choose rather than the AI your vendor bundles.
It also raises the stakes of everything argued above. An AI tool that can only read a chart can mislead a clinician. An AI tool that can write to the chart can change the record itself. The more open the interfaces become, the more the buyer's own standards become the real safety system.
Deregulation does not delete a question. It changes who has to ask it.
The calendar
The dates make the case for acting now. Vendors lose the obligation on December 28. Every contract renewal between now and then is a chance to keep the disclosure before "we are no longer required to" becomes the standard answer.
Four principles for buying AI after HTI-5
- Ask before it switches on. Require notice before any AI feature is enabled, so nothing arrives in your clinics by software update alone.
- Ask for the card anyway. Write the old source attributes into the contract for every predictive feature that touches clinical decisions.
- Tier by consequence. Spend scrutiny where clinical stakes are high and federal review is absent, and keep the rest light.
- Keep the off switch. Reserve the right to disable any AI feature without losing the rest of the product, and lock in today's certified functions at today's price.