“How do we make AI worth its cost? The answer is simple: We fix the world,” said Senior Research Scientist Leo Celi, clinical research director of the Laboratory of Computational Physiology at Massachusetts Institute of Technology,
Celi’s bold proclamation was part of his keynote address at the second annual Responsible AI for Health Symposium, hosted by Johns Hopkins Carey Business School’s Center for Digital Health and AI. The event, held at the Johns Hopkins University Bloomberg Center in Washington, D.C., was designed to examine the growing role of artificial intelligence in healthcare through a lens of responsibility, ethics, and the preservation of human expertise.
"RAIHS highlights the critical role of human factors in AI success, which closely aligns with CDHAI’s mission,” said Gordon Gao, professor at Carey Business School and co-director of the Center for Digital Health and AI. “Just as road safety requires both better cars and better drivers, responsible AI requires not only responsible technology, but also responsible and well-prepared users.”
As it becomes increasingly clear that AI is not going away, conversations emphasize the importance of caution in the development and deployment of AI while harnessing its potential to better serve, support, and elevate humanity. David Rhew, global chief medical officer and vice president of Health and Life Sciences at Microsoft, outlined gaps in FDA-cleared AI for clinical use, citing a lack of important information such as demographic data, bias assessments, and training sample sizes. He also pointed to evidence showing that approximately 50% of AI systems perform worse in real-world settings than in testing, underscoring the need for continuous testing and oversight.
Rhew also offered an industry perspective on the FDA’s recent updates to the Clinical Decision Support guidance, released in January 2026. The new guidelines reduced pre-market oversight for low-risk AI and wearable devices, expanded general wellness, and shifted the emphasis to real-world performance. This shift from pre-market regulation to post-market monitoring leads to more AI being deployed with less FDA gatekeeping, turning health systems into the primary safety net. As a result, responsibility increasingly falls on health systems to ensure continued safety monitoring, performance evaluation, bias assessment, and AI governance.
According to Celi, medical practitioners and the broader technology ecosystem are currently out of sync on AI, potentially creating gaps between human-centered models and those designed primarily to benefit companies and investors. The lack of guardrails, continuous quality assurance, and guaranteed human oversight in pre-market settings may have detrimental effects in healthcare environments using AI.
Atul Gawande, the John and Cyndy Fish Chair in Surgery at Brigham and Women’s Hospital, and the Samuel O. Thier Professor of the Practice of Surgery at Harvard Medical School, raised concerns over the limitations of AI models that lack the context and trade-offs necessary to truly support decision-making for healthcare workers.
“Impact depends on both the technology itself and how it is implemented within the primary healthcare system,” said Gawande.
Across sectors, many conversations also focused on the human implications of AI systems and pointed to knowledge disruption that happens as entry-level tasks increasingly move into AI systems.
Retaining knowledge and resisting over-reliance
“AI displacement of entry-level jobs is consistently framed as efficiency, as progress, as inevitable. But efficiency for whom?” Celi asked.
Over-reliance on AI can be seductive for workers who experience burnout, and over time, this choice can lead to deskilling or the erasure of knowledge and experience. The practice of “vibe-coding,” for example, is creating new knowledge gaps as developers rely on AI tools to keep up with industry demands, often bypassing the learning of programming fundamentals.
Celi noted that scientists—like coders—are already leaning into a similar approach (“vibe-science”) due to pressure for faster turnaround times and quick results. The practice could erode the specific rigor involved in research. If machines remove rigor, the short-term goals are met faster, but scientists who lack fundamental expertise could eventually lose the ability to challenge hallucinations or detect errors in the AI-generated outputs.
Celi took this argument further, framing responsible AI as a social issue. Using healthcare as a concrete example, he argued that unchecked AI doesn’t necessarily benefit humanity. By automating documentation and diagnostic support, efficiency gains may primarily benefit technology companies rather than clinicians and those providing care, leaving healthcare workers “relegated to being assembly line workers,” and diminishing their agency.
This process could create broader societal consequences as the automation of entry-level knowledge eliminates not just those jobs, but also what Celi referred to as the “onramp to the middle class.” If current trends continue, Celi warned, systems of inequality may deepen, ultimately concentrating the benefits of AI among those who already possess inherited access and wealth.
The human-centered path forward
RAIHS convened researchers and experts to engage with these concerns and consider the best path forward to ensure that AI systems in healthcare remain human-centered. Discussions centered on how to steer development away from unethical uses, unchecked development, and blind implementation, factors that could lead to over-reliance on AI and create uncertainty around its reliability and safety. This intended outcome looks at the human factor surrounding conversations on AI rather than the technology on its own.
Rhew argued that it is important to accept the evolving technological landscape but change the paradigm, readjusting how we view this advancing technology. How can AI benefit humanity? How can it meaningfully improve healthcare?
“What gives me optimism is not the sophistication of the models or technologies, but the growing recognition that responsibility must be engineered into healthcare AI from the outset. That shift in mindset may be the most important signal of progress,” said Ritu Agarwal, Wm. Polk Carey Distinguished Professor at Carey Business School and co-director of the Center for Digital Health and AI.
The symposium was supported in part by a Johns Hopkins NEXUS Award, a university-wide initiative launched in 2023 to support convening, research, and teaching at the Bloomberg Center. RAIHS was led by Professor Ritu Agarwal from the Carey Business School and organized by an interdisciplinary team of faculty from across the Johns Hopkins ecosystem, including Professor Gordon Gao from the Carey Business School, Professor Emily Haroz from the Bloomberg School of Public Health, Professor Kadija Ferryman from the Berman Institute of Bioethics, Drs. Risa Wolf and Brian Hasselfeld from the School of Medicine, and Professors Mark Dredze and Rama Chellappa from the Whiting School of Engineering.
