Rethinking Clinical Decision Support: Insights from Dr. Adam Rodman on the Future of AI in Healthcare
Artificial intelligence is reshaping healthcare at a pace few anticipated—but according to Dr. Adam Rodman of Harvard Medical School, we’re still at the very beginning of understanding how it should integrate into clinical practice. In a recent episode of Practical AI in Healthcare, Dr. Rodman joined us for a candid conversation about decision-making, diagnostic reasoning, and the regulatory and cultural challenges that stand between today’s tools and tomorrow’s fully realized AI-enhanced healthcare system.
From Medical Historian to AI Researcher
Rodman describes himself as “the strangest artificial intelligence researcher you will ever meet,” because he didn’t begin in computer science—he began by studying how doctors think. His work in clinical reasoning and diagnostic excellence focused on human decision-making, cognitive biases, and the influence of medical technologies over time.
Ironically, his own historical analysis led him to underestimate how fast AI would progress. In 2021, while finishing his book Shortcuts Medicine, he predicted that large language models (LLMs) were still “more than a decade” away from meaningful clinical capability—only for GPT-3.5 and GPT-4 to prove him wrong within months.
The rapid leap in performance, particularly the ability of newer models to maintain and reason over longer context windows, prompted Rodman to shift his research toward understanding how these models could augment—or fail to augment—clinical decision-making.
Why Better AI Doesn’t Automatically Improve Care
One of Rodman’s most striking findings is that simply giving clinicians access to LLMs does not reliably improve diagnostic accuracy. He has conducted randomized controlled trials showing that physician performance does not automatically increase when asked to consult AI, even when models perform well on benchmark tasks.
This echoes a longstanding truth in clinical decision support (CDS): tools succeed or fail not only on accuracy, but on the human-AI interface. Earlier systems—from Internist-1 to the classic QMR—were powerful but unbearably cumbersome. Physicians often abandoned them not because the logic was flawed, but because the workflow was.
Today’s LLMs are different: they can interpret messy, real-world text and generate rich reasoning. But they also hallucinate, overgeneralize, and behave inconsistently. Rodman emphasizes that these limitations aren’t fatal—they’re challenges of collaboration design. The real question becomes: What interaction models lead humans and AI to perform better together than either can alone?
AI as Collaborator, Not Replacement
Rodman sees tremendous potential in AI-assisted patient intake and data gathering. His lab is experimenting with models that conduct preliminary conversations with patients before a physician encounter—capturing symptoms, history, and values—and then synthesizing that information for clinicians.
The promise is twofold:
- Efficiency: Reduce the burden on clinicians who spend vast time extracting basic information.
- Quality: Enable more thorough, consistent patient interviews, especially in high-volume settings.
Some real-world urgent care companies are already doing this, such as K Health and other “digital front door” providers. These systems handle simple, low-risk cases by having patients talk to an AI first, with a physician reviewing the case before finalizing the treatment plan.
However, Rodman points out that urgent care is the “easy slice” of medicine: low complexity, minimal context, and younger, tech-comfortable users. Scaling such models to chronic disease management or acute inpatient care requires deeper breakthroughs—especially in AI’s ability to parse messy EHR data and reason across long clinical timelines.
The Toughest Barrier: Regulation and Liability
While technology is accelerating, regulation is not. Rodman is clear that neither current medical-device frameworks nor traditional drug-style clinical trials adequately capture the dynamic, adaptive nature of LLMs.
Two alternative regulatory models he discusses include:
- Benchmark-Based Validation
AI models could be approved based on standardized clinical task simulations—though defining tasks that represent the complexity of medical decision-making remains a challenge. - AI-as-Judge Oversight
Smaller, fine-tuned models could provide scalable quality monitoring for larger clinical models post-deployment. But this raises the “watching the watchers” problem—and depends on advances in AI evaluators that don’t yet exist.
Because vendors fear liability and institutions fear risk, Rodman predicts that near-term healthcare AI will remain largely human-in-the-loop, slowing widespread transformation.
A Slow Start—Then Rapid Acceleration
Despite enormous hype, Rodman does not expect healthcare to transform radically in the next three to five years. He anticipates incremental adoption of human-in-the-loop systems, followed by a “dam-burst moment” once regulatory clarity, cultural comfort, and technical capability align.
When that happens, change may come swiftly—similar to how autonomous vehicles moved slowly for a decade before suddenly operating driverless taxis in major cities.
The Big Question: Who Will AI Ultimately Serve?
Rodman offers both a hopeful and pessimistic vision:
- Pessimistic: AI tools may be optimized for hospital administrators and payers, reinforcing billing priorities rather than clinical excellence—similar to how EHRs evolved.
- Optimistic: Companies building patient-facing AI have strong incentives to prioritize patient experience and outcomes, potentially reshaping care around patient needs rather than institutional ones.
Which direction wins will shape the next era of healthcare technology.