Thoughts on the HHS RFI on Artificial Intelligence: How we get to innovation, trusted adoption, and equitable impact

Start at: Equitable Access and Training
If we want all organizations to have an opportunity to even begin their AI journey, we need to make sure they can get to the starting line. That means providing support for workforce training programs to ensure all healthcare professionals have access to role-based, foundational AI education. It’s equally important that additional private sector accreditation and certification programs remain affordable to ensure that all individuals and organizations have access to advanced training. Whether you are able to be seen by a provider or organization with specific AI accreditation and certification shouldn’t depend on your zip code.

The Rural Health Transformation Program (RHTP) provides potential opportunities for organizations to leverage federal funding to support healthcare AI adoption in their communities. If distributed fairly, smaller organizations could benefit by using funds to jumpstart their AI education, adoption, and use. But the financial support and technical resources needed by smaller organizations to effectively integrate AI tools require a more dedicated source of funding. AI adoption doesn’t stop with initial implementation; it requires ongoing evaluation and maintenance to identify and mitigate risks related to model drift, bias, and the complex interplay of technology and human behavior. Even with the RHTP, additional federal grant programs to provide technical assistance will be needed. Bridging the gap between early adopters and the rest of the industry requires more than just access; it requires a national baseline of AI operational readiness and commitment to more diverse sources of formal AI evaluations across different patient populations and settings. A collective knowledge base enables a more comprehensive understanding of AI limitations, risks, and which applications drive value.

Watch for Forks in the Road: Defining Transparency
The concept of a model card — a short document intended to increase transparency by communicating basic information about the development processes for an AI tool — has gained some traction in the industry, but it’s not universal yet. Current HTI-1 regulations require a set of 31 Source Attributes, detailing information ranging from how an AI intervention was developed to its intended use, known risks, and limitations. The key is as much about the content as it is how it is conveyed. Helpful details for one organization might be overload for another. There is a growing need for functional transparency — and what information that encompasses differs depending on the audience. For health IT leaders and data science teams, more concrete performance metrics may be expected, but for end users, who need to grow to trust and understand the AI tool within practice, providing too much information can be overwhelming at the point of care. We need to achieve greater consensus around what the most essential information is for users based on organizations’ varied experiences with assessing and implementing AI to date.

Avoid Dead Ends: Under- or Over-Regulation
In 2025, over 250 bills related to AI in healthcare were proposed by 47 states. Early evidence from 2026 legislative trackers shows that trend will continue, with numerous healthcare-related AI bills already introduced or advancing across states this year. Having widely varied regulations across state lines makes it challenging for developers and multi-state healthcare organizations to coalesce their approach to governance. We need a harmonized approach to federal and state regulations that ensures a nationwide minimum level of comfort with AI adoption. Over-regulation stifles innovation; under-regulation threatens equity and trust. Either outcome is equivalent to a dead end. On the one hand, we could end up with very little new or competitive innovation to drive the industry forward, make new discoveries, and identify the most impactful AI solutions. On the other hand, we could have limitless AI solutions and vendors in the market, unchecked and with little to no ability to compare capabilities and performance, ultimately leading to a lack of trust and stalled adoption. Neutral sandboxes could provide a testing environment that would allow prospective assessment of an AI solution’s safety, impact, and viability. This would allow real-world insights to guide policy and regulatory action, and ground a risk-tiered approach to governance and oversight in evidence. And could even be used to help define outcomes and benchmarks that support reimbursement pathways for proven AI technologies.

Stay Straight on: Interoperability Advancements
MEDITECH has been a staunch advocate for interoperability for decades. As interoperability leaders, we strongly support the continued advancement of the TEFCA framework, CMS-aligned data exchange networks, and standards-based APIs. The next revolution of data is underfoot, and knowing that quality data is the very lifeblood of AI models, the industry needs to continue to support industry collaboratives such as the Sequoia Project’s Taking Root Movement to ultimately reduce deduplication efforts and model bias. We need to double down on the concept that data should follow the patient wherever they are in their care journey. When a patient is seen, and the clinician uses AI to summarize their most recent visit, or utilizes a predictive algorithm to assess risk, the model should have access to the most current and accurate patient data, regardless of where the visit happens. Taking an even more forward-looking perspective, we need to start preparing now for the need for an expanded data standard that defines how AI output is formatted in FHIR.