The following version of the public comment letter was submitted to the Department of Health and Human Services on February 23, 2026.
On behalf of the Reason Foundation, we respectfully submit these comments in response to the Request for Information (“RFI”) issued by the Office of the Assistant Secretary and Assistant Secretary for Policy (ASTP) and the Office of the National Health Information Technology Coordinator (ONC) on Accelerating the Adoption and Use of Artificial Intelligence as Part of Clinical Care.
The Reason Foundation is a national 501(c)(3) public policy research and education organization with expertise across a wide range of policy areas, including artificial intelligence policy.
Our comments address two specific questions included in the RFI:
- What are the biggest barriers to private sector innovation and adoption and use of AI for healthcare in clinical care?
- What regulatory, payment policy, or program design changes should HHS prioritize to encourage the effective use of AI in health care, and why? What regulations, policies, or programs at HHS could be revisited to enhance its ability to develop or use AI in clinical care? Please provide specific changes and citations to the applicable Code of Federal Regulations.
1. What are the biggest barriers to private sector innovation in AI for healthcare and its adoption and use in clinical care?
Regulatory uncertainty is a major barrier to private sector innovation and adoption of artificial intelligence (“AI”) in clinical care. This uncertainty surrounds the boundary between regulated and unregulated software that assists clinicians in making medical decisions. While medical devices that autonomously perform diagnosis and, in some cases, treatment, are fully regulated by the Food and Drug Administration (“FDA”) and are subject to lengthy clinical trials, other software, known as clinical decision support (“CDS”), is informational and therefore not fully regulated.
CDS software is typically integrated into hospital workflows and analyzes patient data to provide alerts, risk assessments, and recommended next steps to treating clinicians. The already blurred boundaries between CDS recommendations and medical device diagnostics are a source of concern for developers who preemptively strip valuable functionality, such as time-sensitive alerts, from products and redesign what their software displays to clinicians in order to circumvent device classification and related regulations. As a result, clinicians are left with weaker and less useful tools than current technology can provide, reducing diagnostic accuracy and worsening patient outcomes.
In practice, CDS developers avoid certain features, such as providing specific probabilities for expected risks, as shown in 2025. JAMA Health Forum study. For example, a CDS tool designed to detect early signs of sepsis may suggest multiple treatment options even though some options are known to be much more relevant than others. Ranking of these options may be avoided to avoid certification as a medical device. Developers are motivated to warn clinicians that a patient has a “20-30% chance of developing sepsis within 24 hours” and present an undifferentiated, unranked list of options rather than recommending a specific treatment such as “immediate broad-spectrum antibiotics.” Dilution of CDS functionality reduces expected clinician time savings from 35% to less than 15% and diagnostic accuracy improvement by 22%. These changes nullify AI’s core value of augmenting human judgment with accurate, actionable insights at the point of care.
This appears to be exactly what Congress sought to avoid with the 21st Century Cures Act, which provided legal exemptions for certain clinician CDS that support professional judgment and allow for independent review. This exclusion applies if the following four conditions apply:
- The software provides support or recommendations to medical professionals.
- The software analyzes patient-specific medical information.
- Experts should independently consider the basis for their recommendations. and
- The software will disclose any limitations and known failure modes.
Congress designed this safe harbor to encourage non-device CDS tools that enhance, rather than replace, professional judgment, and to avoid FDA’s lengthy premarket reviews of assistive technologies while maintaining safety through transparency and reviewability.
Nevertheless, from 2022 to January 2026, FDA guidance repeatedly reinterpreted and reinterpreted the boundaries between regulated device software and non-regulated CDS tools. The agency first issued draft CDS guidance in September 2022, proposing four criteria for distinguishing between regulated devices and treatment exemptions. We then released an explanation in January 2023, further reinterpreting the exclusions for capture tools with probabilistic outputs and action prioritization. The 2024 update further tightens this standard, and many assistive CDS features, such as risk scores and ordered options, will be considered “device-like” output even if reviewed or overridden independently by a clinician.
With boundaries tightened over the years, developers now face a more rigid and unpredictable environment that slows the adoption of AI in clinical care. Hospitals may be less likely to implement CDS tools at scale if they cannot reliably predict whether normal workflow features will trigger device classification and cannot determine who will bear the burden of compliance such as validation, maintenance, and adverse event reporting.
Former FDA Commissioner Scott Gottlieb and Sen. Bill Cassidy (R-Louisiana) directly challenged these interpretations in an April 2024 letter to FDA leadership. They called for evidence-based justification for expanding regulatory scope beyond the letter of the law. To date, the agency has not provided a clear answer or cited new security data that warrants border enforcement. Rather, the precautionary logic leads to the classification of merely auxiliary tools as medical devices. This logic risks hindering the rapid adoption of AI tools that have the potential to reduce clinician burnout by 30% and diagnostic errors by 20%, according to peer-reviewed pilots. Without an HHS-led course correction, the FDA’s unilateral reinterpretation risks ceding U.S. clinical AI leadership to a less risk-averse market.
After several years of progressive boundary tightening, FDA issued new CDS guidance in January 2026 that partially loosens FDA’s previously very restrictive treatment of clinician software. It more clearly links the boundaries between device and non-device with the four statutory criteria for treatment, and recognizes that patient-specific practical recommendations can qualify as non-device CDS as long as they support rather than replace professional judgment and allow for independent review of the underlying rationale. This guidance consolidates previous documents and replaces some of the strictest “any instructional language = device” interpretations with a more nuanced focus on intent and reliability. We also use an expanded example to demonstrate that certain high-value CDS features do not need to automatically trigger full device regulation.
Despite these recent course corrections, significant barriers remain. The boundaries between exempt CDS and regulated device software remain complex, multifactorial, and highly dependent on agencies’ evolving “current thinking” rather than a broad and unambiguous safety zone. Innovators still have to navigate fine distinctions such as whether a tool’s language “supports” or “drives” clinician decisions, whether medical predictions are too “time-sensitive” to be considered independently, or whether a single recommendation is the only “clinically appropriate” decision, all under the shadow of discretionary enforcement. This uncertainty disproportionately harms small businesses and startups, which lack the resources to continually reinterpret guidance, absorb regulatory risk, and adjust products late in development. All of this is discouraging the adoption of more innovative AI and instead leading smaller companies to settle for more cautious, low-impact designs.
Additionally, FDA’s new guidance leaves broader structural barriers unaddressed. Many of the most clinically valuable AI applications, such as early deterioration detection and sepsis prediction, remain effectively guided by device status, hindering investment and slowing adoption. Responsibility environments still favor traditional manual workflows, and with little clear AI-specific direction for generic models or patient-facing tools, developers are unsure which regulatory regime applies. Together, these factors continue to steer innovation toward what regulators are most comfortable allowing, rather than what clinicians and patients will voluntarily adopt in a more predictable and innovation-friendly framework.
Even with non-device CDS clearing FDA boundaries, hospitals face serious liability gaps that slow deployment. Questions remain about who is responsible for things like clinician training and incident reporting. Although the 2026 FDA guidance clarifies the classification, it omits this important post-adoption clarity, leaving providers to navigate fragmented accreditation standards and state regulations through lengthy legal review. This uncertainty predisposes risk-averse institutions to reject high-value AI, even when pilots show clear improvements in efficiency and accuracy, further directing innovation toward tools that circumvent device classification rather than those that provide the greatest clinical value.
2. What regulatory, payment policy, or program design changes should HHS prioritize to encourage the effective use of AI in clinical care and why? Which regulations, policies, or programs at HHS could be revisited to enhance its ability to develop or use AI in clinical care? Please provide specific changes and citations to the applicable Code of Federal Regulations.
HHS should direct FDA to codify broad and binding safe harbors for intended uses of devices and regulations for premarket approval procedures in 21 CFR parts 801 and 806 that fully implement the four CDS standards for therapeutics. Doing so would replace the current patchwork of non-binding guidance with clear rules that protect probabilistic outputs, prioritized recommendations, and time-critical tools, explicitly including high-value applications such as sepsis detection, when clinicians retain review authority.
To fill the accountability gap, HHS should revise the Centers for Medicare and Medicaid Services’ Participation Regulations and the Office of the National Health Information Technology Coordinator’s Certification Standards to establish a simple, voluntary framework for assigning responsibilities that allows for rapid contracting without legal paralysis. These responsibilities include:
- Developer discloses inputs/restrictions.
- Hospitals will verify on-site and train staff. and
- The clinician overrides and reports the incident.
These targeted steps will create predictability, foster competition, and promote clinician freedom, thereby steering innovation toward patient-chosen solutions rather than regulatory comfort zones.
