A new report published Oct. 13, 2025, in the Journal of the American Medical Association outlines how the health care sector should responsibly seize the opportunities of AI, including what must change to ensure AI adoption improves patient outcomes, not just efficiency.

Among the key recommendations in “AI, Health, and Health Care Today and Tomorrow”: expanded oversight by the Food and Drug Administration and the development of evaluation tools to measure effectiveness in clinical settings.

The report was co-authored by Michelle Mello, professor of law and of health policy at Stanford Law School and Stanford University School of Medicine, along with medicine professors Dr. Tina Hernandez-Boussard and Dr. Nigam Shah. The report grew out of the 2024 JAMA Summit on Artificial Intelligence, an invitation-only convening that brought together more than 60 leaders in medicine, law, policy, and industry to examine the opportunities and risks of AI integration in clinical care. The summit was part of an ongoing JAMA series launched in 2023 to spark cross-sector dialogue and drive practical solutions to pressing health policy challenges.

“AI is being adopted at remarkable speed in the health care sector, but our systems for evaluating and regulating it haven’t kept pace,” said Mello, a member of the National Academy of Medicine whose empirical research is focused on understanding the effects of law and regulation on health care delivery and population health outcomes. “This report identifies concrete steps that can help make AI’s integration into health care more transparent, effective, and fair.”

Mello and her co-authors emphasize that AI’s potential is vast for reducing administrative burdens, improving diagnostic accuracy, personalizing treatment, and extending care to underserved populations. But without better infrastructure, evaluation, and incentives, they write, that promise could be undercut by limited and inequitable deployment, unintended harms, and wasted resources

Four priorities for responsible integration

The authors outline a roadmap for safer and more effective AI adoption:

• Multistakeholder engagement throughout an AI tool’s life cycle, bringing together developers, clinicians, regulators, health systems, and patients to align design, deployment, and monitoring.
• Robust evaluation tools and methods to measure effectiveness in real-world settings, not just technical performance in test environments. The report calls for new ways to rapidly assess outcomes across diverse care settings and patient populations.
• National data infrastructure to support learning across systems, similar to the FDA’s Sentinel initiative, which uses large, distributed health data networks to monitor medical product safety in real time. A shared data environment would help identify both benefits and unintended harms more quickly.
• Stronger regulatory frameworks and incentives to ensure accountability and responsible use, including an expanded and better-coordinated oversight role for the FDA and other federal agencies. The authors also call for funding mechanisms, clearer rules, and aligned incentives for developers and health systems to participate in evaluation and compliance efforts.

A health system already in transition

AI tools are increasingly embedded in clinical practice, from sepsis alert systems in hospitals to mobile apps that help patients track heart rhythms or mental health symptoms. Others work behind the scenes to automate scheduling, billing, and prior authorization. Some, like AI scribes, straddle both worlds: transcribing clinical conversations while suggesting treatment options.

Yet only a portion fall under FDA oversight, and even those that do often aren’t required to demonstrate real-world effectiveness, according to the report authors. 

Tools used to support business operations, such as algorithms for prior authorization or operating room scheduling, can shape patient access to care but usually aren’t subject to FDA review. Direct-to-consumer apps, which now number in the hundreds of thousands, are typically marketed as low-risk wellness tools, meaning they can avoid regulatory scrutiny altogether. Even for clinical AI tools that do undergo FDA clearance, demonstrating improved patient outcomes is not always required.

“Hospitals are adopting AI tools faster than they can realistically evaluate them, and most don’t have the infrastructure or resources to run rigorous assessments in-house,” Mello said. “Right now, oversight is mostly about process and safety checks – like preventing algorithmic errors or meeting transparency requirements – not about whether these tools actually improve health.”

The goal, argue the report’s authors, isn’t to slow innovation but to make sure its benefits are real, measurable, and distributed fairly.