Setting the Stage
Moderator Dr. Carmen Sceppa, Dean of the Bouvé College of Health Sciences, opened by taking a quick pulse of the room. A show of hands revealed the audience was divided, some in favor of AI as doctor, many opposed, many unsure. By the end of the morning, more hands went up for “yes” than at the start.
Arguing for AI as doctor was Isaac (Zach) Kohane, MD, PhD, the inaugural Chair of the Department of Biomedical Informatics at Harvard Medical School and Editor-in-Chief of NEJM AI, a physician and computer scientist who has spent three decades at the intersection of data and care.
Arguing against was Eugene (Gene) Tunik, PT, PhD, Senior Associate Dean of Research and Innovation at Bouvé and Director of the AI + Health & Human Performance division at Northeastern’s Institute for Experiential AI, a researcher whose work spans movement neuroscience, rehabilitation technology, and human-centered AI.
Both acknowledged up front that their assigned positions might not fully represent their own views. What followed was a debate designed to stress-test the premise from every angle.
The Case For: AI as First-Line Doctor
Kohane opened by reframing the question. He wasn’t arguing for AI to replace physicians entirely, he was arguing that AI should be “the always-on, first-line clinical layer” working alongside human doctors. His reasoning: the status quo is already broken.
The U.S. faces a projected shortage of up to 86,000 physicians by 2036, he noted, and training more doctors is slow and bottlenecked by limited residency slots. The average wait time for a new primary care appointment is 30 days; for psychiatry, it stretches to 90. Even physicians at Boston’s most prestigious hospitals often lack their own primary care doctors.
He pointed to a study published in Science the week of the event showing that physicians at Beth Israel were outperformed by a 2024 version of GPT on diagnostic tasks, as judged by external doctors. Diagnostic errors, he argued, cause approximately 800,000 serious harms annually in the U.S., including around 400,000 deaths. A continuously available AI that could catch even a fraction of those errors, he said, would be “unethical not to use.”
Kohane illustrated his point with a personal story: a colleague who died of a heart attack after being misdiagnosed with asthma, and a former student at UCSF who received a wrong cancer diagnosis, one that an early version of GPT-4 flagged as suspicious before any human clinician did. “Not using AI as your first-line doctor,” he said, “is a mistake.”
He closed on equity: without AI, the gap between those who can afford concierge medicine and those who cannot will only widen. AI, he argued, is the only scalable path to “concierge-light” care for everyone.
The Case Against: AI Is a Tool, Not a Doctor
Tunik acknowledged everything Kohane said, and then pushed back on what he saw as a categorical confusion.
“Doctors don’t just do pattern recognition,” he said. “They do judgment under conditions of uncertainty. That’s medicine.”
The EKG vignette that Dr. Sceppa presented, in which an AI embedded in an electronic health record flagged a subtle cardiac pattern that a busy physician had missed, potentially preventing a fatal arrhythmic event, proved his point rather than Kohane’s, Tunik argued. The AI flagged a concern. The doctor ordered the follow-up echo. The AI didn’t bear the weight of that decision, and it wouldn’t have if the outcome had been different.
He cited the 2026 State of Clinical AI report, which found that even top-performing large language models make omission errors at rates as high as 20%, failing to recommend follow-up tests or emergency procedures when warranted. Deploying those models widely, he warned, risks creating “a regression to the mean” where AI performs at the lower end of physician performance, not the top.
On the question of accountability, Tunik was direct: “By definition, clinicians are doctors because they deliver the care and bear the responsibility. AI does none of those things.” He outlined how liability would need to be distributed across physicians, hospitals, and AI developers, and argued that until that framework exists, AI cannot and should not be treated as a doctor.
His closing argument was compact: “Medicine is not just detection. It is judgment under uncertainty. Patients live in the gray zone of ambiguity. This is what humans are good at.”
Where They Actually Agreed
Despite the spirited sparring, both debaters found significant common ground, and were candid about it.
They agreed that younger, less experienced clinicians are currently struggling to identify AI errors, which risks perpetuating mistakes rather than catching them. Both called for AI to be integrated into clinical training early, treated like any other instrument, a stethoscope, a radiological scan, with competencies built through simulation and repetition.
They agreed that medical schools are badly behind. Kohane was characteristically blunt: “Our medical schools are in no position to train our doctors how to use AI to avoid deskilling.” Tunik called for apprenticeship-style models and clinical simulations that help trainees detect AI errors before they encounter them in the real world.
They agreed that the fractured, fee-for-service U.S. healthcare system, and its global counterparts, is a major obstacle to responsible AI deployment. The data silos, misaligned incentives, and administrative weight of the current system make it hard to train, validate, and monitor AI models at scale.
And they agreed, most pointedly, that this debate needs to be happening far more broadly. “We have not had this debate at Harvard Medical School,” Kohane noted. “This is a debate our healthcare system should be having right now.”
Questions from the Room
The audience, a mix of physicians, pharmacists, physical therapists, nurses, engineers, students, and researchers, pushed the conversation into territory the debaters hadn’t fully mapped.
One attendee raised the growing burden on physicians, who have already absorbed roles as researchers, billers, and accountable care managers. Adding AI stewardship to that list, she asked, is it even fair? Kohane acknowledged the problem squarely: if healthcare institutions don’t step up, platform companies like Amazon’s One Medical will fill the gap, and may do so on their own terms.
The question of insurance and predictive data drew pointed exchanges. If AI can predict a patient’s disease risk decades in advance, who gets that information, and what can insurers do with it? Kohane drew on the precedent set by the Genetic Information Nondiscrimination Act, which prohibits health insurers from using genetic data in pricing, while noting that life insurers are not bound by the same rules. The underlying principle, he said, is that insurance only works under conditions of partial ignorance: “We cannot allow a pure information risk calculation.”
Mental health came up too, with a question about high-profile cases in which AI systems, by being sycophantic, failed to redirect users in crisis. Kohane didn’t minimize the concern but placed it in context: clinicians make serious errors in mental health care too, at a scale that goes largely undocumented. The answer, he argued, is not to exclude AI but to align it better toward patient interests, which requires the kind of societal and regulatory clarity the debate itself was trying to generate.
Tunik pointed to emerging state-level regulation as one near-term brake on irresponsible deployment. New York has a pending bill that would prohibit AI from making health-related recommendations; California and Texas are taking different approaches. “I don’t know how it will cross state lines,” he said, “but those guardrails matter.”
The conversation also touched on how developing nations might actually leapfrog the entrenched inertia of Western healthcare systems, just as much of the world skipped landlines and went straight to mobile, and build more effective AI-augmented care on a cleaner foundation.
What Comes Next
The debate ended where good debates should: with more questions than answers, and a room full of people thinking harder about something that will affect every one of them.
What both Kohane and Tunik made clear is that the real question is no longer whether AI will transform healthcare, it’s how, for whom, and on whose terms. The answers to those questions will depend on decisions being made right now, by institutions like Bouvé, by healthcare systems, by policymakers, and by the clinicians and researchers training the next generation of providers.
Bouvé’s “Should AI Be Your Doctor?” debate is the first in what promises to be an ongoing series exploring AI’s role across healthcare disciplines. Future events will continue to bring leading voices together to examine questions that matter, not just for practitioners, but for every patient.
Our next event will take place this October at Northeastern University Miami. Stay tuned for details.
Participants received 1.5 hours of CME/CNE/CPE continuing education credit. For questions about CE accreditation, contact the Office of Experiential and Continuing Professional Education at [email protected].
