AI in health care regulation: outlook to 2027

FDA's device rules are maturing slowly, ONC has cut its algorithm transparency requirements, and the binding duties on clinical AI now sit in the Section 1557 rule, Medicare Advantage coverage rules and state statutes.

The question for 2027

No single US law governs AI in health care. A health system that buys an imaging algorithm, runs a predictive model in its EHR and lets software draft patient messages answers to five rulebooks: FDA device law, ONC certification, the Section 1557 nondiscrimination rule, Medicare Advantage coverage rules and state statutes. This outlook sets out where each stands as checked on 16 November 2026 and what is likely to move through 2027. Since 2025 the federal direction has been lighter paperwork for software developers, while the rules that put a clinician in front of a coverage or care decision stay.

FDA: a growing list, one final guidance and one long-running draft

FDA keeps a public list of AI-enabled medical devices authorized for marketing in the United States; as checked on 16 November 2026, the page was content current as of September 4, 2026, says the list is updated periodically, and FDA says the list is not comprehensive, because devices were identified mainly by AI-related terms in their public summaries or classification [1]. In that version, the downloadable file holds 1,614 devices with final decision dates from 1995 to June 29, 2026 [2]. In the same version, authorizations rose from 235 in 2024 to 335 in 2025, and the first half of 2026 brought 181 against 176 a year earlier [2]. In the same version, radiology is the lead review panel for 1,230 of the 1,614, about 76 percent, and cardiovascular devices are a distant second at 154 [2]. In the same version, almost all reached the market through 510(k) clearance: 1,553 devices, against 40 De Novo grants and 21 premarket approvals [2].

The rule that governs how these devices may change after clearance is already final. FDA announced its guidance on predetermined change control plans (PCCPs) for AI-enabled device software functions on December 4, 2024 [5]; the guidance page, last updated August 18, 2025, recommends that a PCCP describe the planned modifications, the method to develop, validate and implement them and an assessment of their impact, and says FDA reviews it as part of the marketing submission so each described change does not need a new submission [4].

The broader lifecycle guidance is not final. "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations" remains a draft, marked not for implementation, with its page content current as of January 7, 2025, under docket FDA-2024-D-4488 [3]. CDRH's agenda for fiscal year 2027 puts a final version on its A-list of priority guidances, alongside final guidance on predetermined change control plans for all medical devices [8] [9]. The same list adds a draft guidance on evidence for generative AI conversational devices for mental disorders and, on the "Under Construction" list, risk assessment and evidence for generative AI-enabled devices [9]. CDRH reports that it published 7 of the 11 guidances on its FY 2026 list [8].

FDA also issued final clinical decision support software guidance in January 2026, explaining which functions fall outside the device definition under section 520(o)(1)(E) of the FD&C Act [6]. On August 18, 2026 it released a discussion paper on generative AI-enabled medical devices covering risk assessment, premarket evaluation, postmarket monitoring, foundation models and agentic systems, with comments due October 19, 2026 under docket FDA-2026-N-7874 [7].

Moraph's view

Moraph's view: FDA's own priority list makes 2027 the most likely year for the lifecycle guidance to become final, but CDRH finished about two-thirds of its list last year, so a slip into fiscal 2028 is a real possibility. For hospitals the effect is small either way: the model card and monitoring content the draft asks of manufacturers is what a careful purchasing committee already asks vendors to show.

ONC certification: HTI-1's transparency rules cut back by HTI-5

HTI-1, published January 9, 2024, replaced the old clinical decision support certification criterion with a "decision support interventions" criterion at 45 CFR 170.315(b)(11) [10]. It required certified health IT to let a limited set of users select evidence-based and predictive interventions, to support "source attributes" (categories of technical performance and quality information) and to let users see and change them, and it required intervention risk management for each predictive intervention a developer supplied [10]. The old criterion expired January 1, 2025 [10].

HTI-5, published October 28, 2026 and effective December 28, 2026 with some amendments effective January 1, 2027, removes 14 certification criteria and revises seven [11]. One of the larger revisions narrows the decision support interventions criterion: it removes the source attribute ("model card") requirements, the requirements to access and modify source attributes, and the intervention risk management requirements for predictive interventions [11]. What remains is the duty to let a limited set of users select evidence-based and predictive interventions, and ONC kept the electronic feedback data requirement for evidence-based interventions [11].

In December 2025 HHS and ASTP/ONC also asked for comment on how HHS can accelerate the use of AI in clinical care, with comments due February 23, 2026 [20].

Section 1557: the binding rule that stayed

The federal rule that speaks directly to clinical algorithms used by providers is still in force. Under 45 CFR 92.210, a covered entity must not discriminate on the basis of race, color, national origin, sex, age or disability through patient care decision support tools, has an ongoing duty to make reasonable efforts to identify tools that use input variables measuring those traits, and must make reasonable efforts to mitigate the risk of discrimination for each one identified [12]. Part 92 took effect July 5, 2024, and covered entities had to meet the identification and mitigation duties in paragraphs (b) and (c) within 300 days of that date [13]. A June 2, 2026 notice reports a federal court's October 22, 2025 judgment vacating listed provisions of the 2024 rule, including the definition of sex discrimination in section 92.101(a)(2)(iv), to the extent they expand sex discrimination to include gender identity; the notice says the other provisions remain in force, and section 92.210 is not among those it lists as vacated [14]. As checked on 16 November 2026, the eCFR shows section 92.210 with a single version dated July 5, 2024 [12].

Moraph's view

Moraph's view: with the HTI-1 source attributes leaving certification, a hospital's evidence for its section 92.210 "reasonable efforts" will rely more on what it asks vendors for in contracts and less on what its EHR is certified to display.

Medicare: guardrails by interpretation, AI by demonstration

CMS has no AI-specific rule for Medicare Advantage coverage decisions. The binding text is 42 CFR 422.101(c), which requires Medicare Advantage organizations to base medical necessity determinations on Medicare coverage and benefit criteria, on whether the item or service is reasonable and necessary, and on the enrollee's medical history, physician recommendations and clinical notes [15]. In a February 6, 2024 memo to all Medicare Advantage organizations, CMS said an algorithm or software tool may assist coverage determinations but must comply with those rules, that an algorithm deciding on a larger data set rather than the individual patient's record would not comply, and that algorithms or AI alone cannot be the basis to deny an inpatient admission or downgrade it to an observation stay [16].

CMS proposed "Guardrails for Artificial Intelligence" for Medicare Advantage, but in the Contract Year 2026 final rule of April 15, 2025 it said it did not intend to finalize them, while acknowledging broad interest and saying it would consider whether future rulemaking on AI is appropriate [17]. The Contract Year 2027 final rule of April 6, 2026 added none: AI appears in it in public comments, a cost estimate for call transcripts and a request for information asking how AI and machine learning might be used in risk adjustment [18].

In traditional Medicare, CMS is testing technology-assisted review itself. The WISeR model, effective January 1, 2026, is a 6-year model in two 3-year agreement periods in which companies with experience in technology-enhanced prior authorization review selected services in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington [19]. Review may combine technology and clinician review, and denied claims keep full appeal rights [19].

States: disclosure, human review and limits on AI therapy

States have moved faster and more specifically. California's AB 3030, approved September 28, 2024 as Chapter 848, requires health facilities, clinics and physician offices that use generative AI to generate written or verbal patient communications about clinical information to include a disclaimer that the communication was AI-generated and instructions for reaching a human, unless a licensed or certified health care provider read and reviewed it [21]. SB 1120, approved the same day as Chapter 879, requires health plans that use AI, algorithms or other software for utilization review to base decisions on the individual's clinical record, not solely on a group dataset, and bars the tool from denying, delaying or modifying care based on medical necessity: that determination must be made by a licensed physician or a competent licensed health care professional [22]. AB 489, approved October 11, 2025 as Chapter 615, prohibits terms, letters or phrases in the advertising or functionality of an AI system that indicate or imply its care or advice comes from a licensed health professional [23]. None of the three bill texts contains an urgency clause; under Government Code section 9600, a statute enacted at a regular session goes into effect on January 1 next following a 90-day period from its enactment, and urgency statutes take effect immediately [29].

Illinois's HB 1806 became Public Act 104-0054, effective August 1, 2025 [24]. It bars anyone from offering therapy or psychotherapy, including through internet-based AI, unless a licensed professional conducts it; it bars licensed professionals from letting AI make independent therapeutic decisions, interact directly with clients in therapeutic communication, generate treatment plans without their review or detect emotions or mental states, and it sets civil penalties of up to $10,000 per violation [25]. Texas SB 1188, signed June 20, 2025 and effective September 1, 2025 [26], lets practitioners use AI for diagnostic purposes if they act within their license, the use is not otherwise restricted by state or federal law, they review all AI-created records under Texas Medical Board standards and disclose the use to patients [27].

Federal preemption has not reached these laws: Executive Order 14365 of December 11, 2025 sets a policy of a minimally burdensome national framework, directs the Attorney General to set up an AI Litigation Task Force to challenge state AI laws that conflict with it, and calls for a legislative recommendation for a federal framework that preempts them [28]. As checked on 16 November 2026 we found no enacted federal law that displaces them.

Moraph's view

Moraph's view: for 2027, the state rules are the ones with fixed duties and named enforcers. A multistate system or plan needs a register mapping each AI use to the state disclosure, human-review and licensing rules that reach it.

What to watch in 2027

  • Whether FDA finalizes the AI-enabled device lifecycle guidance during fiscal year 2027, as its FY 2027 A-list plans [9].
  • The Contract Year 2028 Medicare Advantage rulemaking, the next place CMS could return to AI guardrails it set aside in 2025 [17].

Data and method

Device counts come from FDA's AI-Enabled Medical Devices List, downloaded as CSV from the list page on November 16, 2026 [2]; the page was content current as of September 4, 2026 [1]. The file has one row per device with the columns Date of Final Decision, Submission Number, Device, Company, Panel (Lead) and Primary Product Code [2]. We counted rows with a short Python script: totals by the year in Date of Final Decision; the first-half comparison counts decision dates from January through June of 2025 and of 2026; the panel figures group the Panel (Lead) column as written; the pathway figures use the first letter of the Submission Number (K for 510(k), DEN for De Novo, P for premarket approval) [2]. The radiology share is 1,230 divided by 1,614 [2]. FDA says devices whose decision summaries were not published within a data collection period are added in a later update, so recent months may be undercounted, and counts from a later version of the list will differ [1].

What this means for you

  • Health systems: Keep the section 92.210 inventory of decision support tools current; the identification and mitigation duties have applied since 300 days after July 5, 2024 [13], and HTI-5 removes the certified source attribute display that could have supported it [11]. In California, check which generative AI patient messages go out without clinician review, since those need the AB 3030 disclaimer [21].
  • Payers: Document that coverage tools rest on the individual's record, as 42 CFR 422.101(c) and CMS's 2024 memo describe [15] [16], and in California that a licensed clinician makes every medical necessity denial, delay or modification [22].
  • Device and software makers: Use a PCCP to plan model updates under the final guidance [4], and track the lifecycle guidance, which CDRH plans to finalize in FY 2027 [9]. Generative AI products should follow the FDA discussion paper docket [7] and avoid language implying licensed clinician advice in California [23].
  • EHR developers: Plan certification updates for the narrowed decision support interventions criterion, effective December 28, 2026 [11], while noting that hospital clients still carry their own section 92.210 duties [12].

Moraph predictions

Every prediction is Moraph's view, not a fact. Each has a date, a confidence level (70% means we expect to be wrong about three times in ten) and a test anyone can apply. We score them on this page when the date comes and never remove one. How we make and score predictions.

  1. P1 By Confidence 60% Open

    FDA issues final guidance on marketing submission and lifecycle management recommendations for AI-enabled device software functions.

    How we will check: FDA's guidance database or a Federal Register notice of availability shows a final version of the AI-enabled device software functions lifecycle guidance (docket FDA-2024-D-4488) issued on or before September 30, 2027.

    Basis: The draft has been unfinished since January 7, 2025 [3]; CDRH lists a final version on its FY 2027 A-list [9] and published 7 of 11 guidances on its FY 2026 list [8].

  2. P2 By Confidence 65% Open

    The latest version of FDA's AI-Enabled Medical Device List published on or before December 31, 2027 lists at least 1,900 devices.

    How we will check: Take the most recent version of the list page published on or before December 31, 2027 (the live page if its "Content current as of" date is on or before that day, otherwise an Internet Archive copy captured on or before it) and count the device rows in its CSV download or on-page table; the call is right if there are at least 1,900. If FDA has published no new version after the one current as of September 4, 2026, the count is 1,614 and the call is wrong.

    Basis: The version content current as of September 4, 2026 held 1,614 devices with decisions through June 29, 2026, with 335 added for 2025 and 181 for the first half of 2026 [1] [2]; FDA says it updates the list periodically and adds authorizations whose summaries were published late in later updates [1].

  3. P3 By Confidence 75% Open

    CMS does not finalize a Medicare Advantage regulation that expressly governs the use of artificial intelligence or algorithms in coverage or medical necessity determinations.

    How we will check: No final rule published in the Federal Register on or before December 31, 2027 adds regulatory text to 42 CFR part 422 that expressly refers to artificial intelligence or algorithms in coverage or medical necessity determinations.

    Basis: CMS declined to finalize proposed AI guardrails in April 2025 [17], the Contract Year 2027 rule added none [18], and CMS relies on 42 CFR 422.101(c) and its 2024 memo [15] [16].

  4. P4 By Confidence 75% Open

    45 CFR 92.210 remains in the eCFR without amendment.

    How we will check: On December 31, 2027 the eCFR version history for 45 CFR 92.210 shows no amendment dated after July 5, 2024 (editorial, non-substantive versions do not count), and the section has not been removed.

    Basis: The section has one version, dated July 5, 2024 [12], and the June 2026 vacatur notice leaves the rule's other provisions in force [14].

Facts on this page link to their sources. Text marked "Moraph's view", and every prediction, is our own analysis, not a fact or advice. Methodology.

Spotted an error? Tell us and we will check it and correct the piece where needed.

  • artificial intelligence
  • FDA
  • AI-enabled devices
  • predetermined change control plans
  • HTI-1
  • HTI-5
  • Section 1557
  • Medicare Advantage
  • prior authorization
  • state AI laws