Not legal advice. This read summarizes primary legal texts for engineers. Check the source, and talk to a lawyer before acting on it.
What the rule required
HHS certifies the software hospitals and clinics use to keep health records. In January 2024, its health IT office (ONC) added a criterion for "decision support interventions", the alerts and predictions an EHR shows clinicians, including those built on machine learning 4. From 1 January 2025, a certified EHR had to give its users a set of plain-language facts about each tool: 13 for evidence-based tools and 31 for predictive ones 2. Commentators called these model cards. The developer also had to analyze each predictive tool's risks (validity, fairness, safety, privacy and more), reduce them, and govern the data it used.
What a certified EHR had to disclose about each predictive AI tool
The 31 source attributes in 45 CFR 170.315(b)(11)(iv)(B), by group; all removed on 28 Dec 2026
Source: eCFR, 45 CFR 170.315(b)(11), issue of 29 Oct 2026, counted line by line 2
Show the data
| label | value | display |
|---|---|---|
| Quantitative performance | 5 | 5 |
| Details and output | 4 | 4 |
| Purpose | 4 | 4 |
| Development details and input features | 4 | 4 |
| External validation | 4 | 4 |
| Ongoing maintenance | 4 | 4 |
| Cautioned out-of-scope use | 2 | 2 |
| Fairness process | 2 | 2 |
| Update and revalidation schedule | 2 | 2 |
What changes on 28 December 2026
The HTI-5 final rule, published 28 October, removes three parts of the criterion 1:
- The source attributes: all 13 evidence-based and all 31 predictive.
- Users' ability to see, record and change them.
- Intervention risk management: the risk analysis, mitigation and governance duty, with the yearly upkeep duty and the certification body's job of linking each developer's risk summary.
It also narrows the data a predictive tool must be able to use. Several parts stay. Users must still be able to select which decision-support tools are active, evidence-based tools still export feedback, and the criterion stays in the definition of a certified "Base EHR". ONC says it may consider restoring transparency for evidence-based tools in a later rule 1.
The rule also renames the technology. It calls the removed provisions the "super intelligence 'model card' requirements" and uses "SI" throughout, following an executive order that tells agencies to say "Super Intelligence" instead of "Artificial Intelligence" 1. The legal definition of AI is unchanged.
ONC's reasons, and the objections
ONC's case is about use and burden. Developers reported that clinicians rarely open the attributes and lack the time to use them. ONC said it found no evidence that the attributes improved care or helped organizations evaluate AI, and that other parts of HHS, such as FDA, are better placed to oversee it. In its words, it was "not persuaded of their value" 1.
The comments ran the other way. ONC itself reports that most commenters on this proposal opposed it, and more than half of those raising transparency concerns wanted the attributes or the whole criterion kept 1. The AMA proposed a risk-tiered alternative instead of removal 5, and said after the final rule that it will keep working to preserve AI transparency 10. The AHA called the criterion one of the few standards that give hospitals AI information for buying decisions 6. Nursing-home groups asked to strengthen the cards, including validation on older and multimorbid patients 7. Even the vendors' association reported that its members were split, and that many would keep providing the information because customers ask for it 9.
From model cards to no model cards in three years
Each date is from the Federal Register or the agency's own page
- 18 Apr 2023HTI-1 proposedFirst proposes source attributes and risk management for decision support (88 FR 23746)
- 9 Jan 2024HTI-1 final ruleCreates the decision support criterion 170.315(b)(11): 13 + 31 source attributes and risk management (89 FR 1192)
- 6 May 2024Section 1557 ruleAdds 45 CFR 92.210 on patient care decision support tools (89 FR 37522)
- 1 Jan 2025Model cards required(b)(11) replaces the old CDS criterion in the Base EHR definition; upkeep duty begins
- 23 Jan 2025Executive Order 14179Later cited by ONC as a reason to drop the model cards
- 1 May 202592.210 duties applyIdentify and mitigate discrimination risk in decision support tools
- 29 Dec 2025HTI-5 proposedRemove 34 of 60 certification criteria, revise 7, including the model cards (90 FR 60970)
- 6 Jan 2026FDA CDS guidanceFinal guidance adds enforcement discretion for single-recommendation tools (re-issued 29 Jan)
- 27 Feb 2026Comments close298 unique submissions; most comments on this part opposed the cut
- 29 Sep 2026Executive Order 14434Agencies to say \Super Intelligence (SI)\"; HTI-5 uses the new term"
- 28 Oct 2026HTI-5 final rule14 criteria removed, 7 revised, 39 kept (91 FR 67916)
- 2026-11-02Today
- 28 Dec 2026Model cards removedSource attributes, access/modify and risk management deleted from (b)(11)
Source: Federal Register, healthit.gov and fda.gov 1438
Show the data
| date | human | label | detail | tone |
|---|---|---|---|---|
| 2023-04-18 | 18 Apr 2023 | HTI-1 proposed | First proposes source attributes and risk management for decision support (88 FR 23746) | 0 |
| 2024-01-09 | 9 Jan 2024 | HTI-1 final rule | Creates the decision support criterion 170.315(b)(11): 13 + 31 source attributes and risk management (89 FR 1192) | 0 |
| 2024-05-06 | 6 May 2024 | Section 1557 rule | Adds 45 CFR 92.210 on patient care decision support tools (89 FR 37522) | 2 |
| 2025-01-01 | 1 Jan 2025 | Model cards required | (b)(11) replaces the old CDS criterion in the Base EHR definition; upkeep duty begins | 0 |
| 2025-01-23 | 23 Jan 2025 | Executive Order 14179 | Later cited by ONC as a reason to drop the model cards | 1 |
| 2025-05-01 | 1 May 2025 | 92.210 duties apply | Identify and mitigate discrimination risk in decision support tools | 2 |
| 2025-12-29 | 29 Dec 2025 | HTI-5 proposed | Remove 34 of 60 certification criteria, revise 7, including the model cards (90 FR 60970) | 0 |
| 2026-01-06 | 6 Jan 2026 | FDA CDS guidance | Final guidance adds enforcement discretion for single-recommendation tools (re-issued 29 Jan) | 2 |
| 2026-02-27 | 27 Feb 2026 | Comments close | 298 unique submissions; most comments on this part opposed the cut | 0 |
| 2026-09-29 | 29 Sep 2026 | Executive Order 14434 | Agencies to say \Super Intelligence (SI)\"; HTI-5 uses the new term" | 1 |
| 2026-10-28 | 28 Oct 2026 | HTI-5 final rule | 14 criteria removed, 7 revised, 39 kept (91 FR 67916) | 0 |
| 2026-12-28 | 28 Dec 2026 | Model cards removed | Source attributes, access/modify and risk management deleted from (b)(11) | 0 |
What still applies
- Section 1557 (45 CFR 92.210). Hospitals and other covered entities must make reasonable efforts to identify decision-support tools that use race, color, national origin, sex, age or disability as inputs, and to reduce the risk of discrimination from them 3. HTI-5 itself points readers to it. Arbiter's agent keeps this inventory and will not start without it (
S1557-INVENTORY). - State laws. California's AB 3030 still requires a disclosure on AI-written patient messages about clinical information, unless a clinician reviewed them (rule
CA-AB3030-DISCLOSURE). More state laws start on 1 January 2027. - FDA, for software that is a medical device. Its January 2026 guidance on clinical decision support widened the room for tools that make a single recommendation 8.
- Contracts. With no federal requirement, a hospital that wants a tool's training data, validation and fairness testing now has to ask for them in procurement. The 31 attributes are still a useful checklist 2.
Show your work
How this was done
Everything here is from primary text read on 2 November 2026: the full HTI-5 final rule (FR Doc 2026-21943) and proposed rule, the HTI-1 final rule, the 2024 Section 1557 rule, the current eCFR text of 45 CFR 170.315 and 92.210, ONC's own pages, and each organization's comment letter. The 31 and 13 attribute counts were made by counting the regulatory text line by line. The comment count is ONC's own (298 unique submissions). The docket shows 305 posted comments, and regulations.gov reports 6,459 received, which probably includes duplicates. Page numbers for each fact are in the research notes kept with the site.
What this does not tell you
We did not find a statement on HTI-5 from the Coalition for Health AI (CHAI). Epic's and Oracle's own comments were not read, and HHS's press release returned an error, so its wording is not quoted. The rule's text is inconsistent in one place: one passage says 5 criteria were revised where the rest of the rule and ONC's fact sheet say 7. We use 7. The Section 1557 compliance date (1 May 2025) is calculated from the rule's 300-day window. This read does not cover FDA's draft guidance on AI-enabled devices.
Data
Every number in the charts, as CSV:
Agent policy
This read is evidence for these rules in the agent's policy:
Sources
- ONC, "Health Data, Technology, and Interoperability: ONC Deregulatory Actions To Unleash Prosperity" (HTI-5 final rule), 91 FR 67916, 28 Oct 2026
- eCFR, 45 CFR 170.315(b)(11), decision support interventions
- eCFR, 45 CFR 92.210, nondiscrimination in the use of patient care decision support tools
- ONC, HTI-1 final rule, 89 FR 1192, 9 Jan 2024
- American Medical Association, letter on the HTI-5 proposed rule, 27 Feb 2026
- American Hospital Association, comments on the HTI-5 proposed rule, 27 Feb 2026
- AHCA/NCAL, comments on the HTI-5 proposed rule, 24 Feb 2026
- FDA, Clinical Decision Support Software final guidance (January 2026)
- EHR Association, comments on the HTI-5 proposed rule, 27 Feb 2026
- AMA, National Advocacy Update, 30 Oct 2026
- ONC, HTI-5 final rule overview
Corrections
None so far. Found an error? Tell me and it will be corrected here, with the date.