With the leadership transition, the Trump Administration has a chance to put FDA’s clinical AI policy in line with its rhetoric
It’s budget season in Washington, and the Trump Administration’s FY2027 FDA request contains a welcome idea: a Clinical Trial Notification Pathway, a risk-based, expedited alternative for some early-stage studies. FDA says the current process is duplicative and time-consuming, and has helped push early-stage preclinical and Phase 1 work to China and Australia. Give the Administration credit. It has recognized that outdated regulation can cost the U.S. competitiveness in strategically important areas.
But the same competitiveness logic applies to another frontier that matters at least as much for American medicine: clinical AI.
No doubt AI has received some ugly headlines in recent months. It also has the potential to widen access and lower costs for patients. Utah this year launched a first-of-its-kind partnership with Doctronic allowing an autonomous AI system to participate in prescription renewals for chronic conditions. That may sound minor if you live in a doctor-rich ZIP Code. It isn’t when more than 40 million rural Americans live in areas with too few primary care providers.
For many Americans, the promise of clinical AI isn’t futuristic. It means a working parent getting a refill approved without missing half a day of work, an ER patient getting a scan interpreted faster at 2 a.m., and a small-town clinic having specialist-level support it otherwise wouldn’t have. Used well, these tools can help catch disease earlier, shorten waits, and make scarce doctors and nurses go further.
Until his resignation, Marty Makary had seemed to understand at least part of this challenge. At the Consumer Electronics Show in January, he announced revised FDA guidance on low-risk general-wellness products and clinical decision support software. Useful steps, but modest ones. More important were his promises: a “new regulatory framework for AI” that would be “smarter and more forward-thinking,” would “modernize the agency and facilitate the AI revolution,” and would help make the U.S. “the best place for capital investment in AI.” Welcome words.
Then came the follow-through, or lack of it. We saw FDA’s stance on AI firsthand when it rejected a citizen petition from Harrison.ai -- a company we know through board service and collaborative research– that proposed an alternative pathway for certain radiology AI devices. Drafted with the help of two former senior FDA officials, including a former device-policy chief and a former top regulator of radiological devices, the petition argued that current U.S. pathways are “burdensome and ill-suited to the rapidly evolving nature of AI” and highlighted an innovation gap between America and what is available in the U.K., Europe and Asia.
The evidence is striking. One review found that 97% of ophthalmology AI devices were available in the European Union, compared with only 8% in the U.S. A separate review found that just one of 26 digital pathology AI devices available in Europe and Britain had been authorized by FDA. DermaSensor, the first FDA-authorized AI for skin-cancer detection in primary care, received authorization in Europe and Australia three years before FDA gave the go-ahead. The Harrison.ai chest X-ray AI algorithm can identify 124 findings abroad but is authorized for only five in the U.S. For a country that leads in frontier AI models, America has fallen far behind in bringing AI to the clinic.
Harrison.ai’s proposal would likely need refinement before implementation. But that is no reason to shut down the conversation and insist that everything is basically fine.
The core problem is that FDA still treats AI largely as just another medical device. That made sense when algorithms were narrow and static, such as software that performed a single radiologic task. It makes much less sense for systems that can be retrained and updated in real time. Nor does it make sense for systems with far more varied outputs than a pacemaker or CT scanner.
The goal isn’t less oversight. It is better-calibrated oversight, combining premarket review with postmarket monitoring. FDA’s device framework is still heavily front-loaded because it was built for products that have fixed and limited outputs. That can make premarket review feel like the only real chance to give a device a red or green light. Unsurprisingly, the bias is often toward red.
But doctors themselves are not assessed once and then left alone. They are licensed, credentialed, and reevaluated throughout their careers. AI systems that diagnose disease or influence prescribing decisions should be governed similarly. Even Congress appears to sense the mismatch. Senate appropriators last year asked FDA to assess whether it has the statutory tools it needs to properly regulate clinical AI.
And while Washington debates whether FDA should move faster, the rest of the world is speeding ahead. That means the benefits of clinical AI—faster routine care, earlier detection and broader access to specialist-level support—may reach patients abroad sooner than they reach patients here.
This pattern has become familiar. On rare disease, FDA often talked flexibility and delivered delay. On AI, Makary began the year promising a more modern framework, less rote regulation and more room for innovation. Yet when presented with a petition arguing that the current system is mismatched to next-generation AI, FDA chose to defend that system.
The Trump Administration’s proposal for a Clinical Trial Notification Pathway is a welcome sign that it understands outdated regulation can drive innovation abroad. The same logic applies to clinical AI, and FDA should move beyond a regulatory approach that isn’t keeping pace with the technology.
Dr. Dreyer is the Vice Chairman of Radiology and Chief Data Science and Imaging Information Officer at Massachusetts General Hospital and Associate Professor of Radiology at Harvard Medical School.
Dr. Hillis is Attending Neurologist at Massachusetts General Hospital, Director of Clinical Operations at Mass General Brigham AI, and Assistant Professor of Neurology at Harvard Medical School.
