"They didn't listen to A.I. when A.I. told them things they didn't agree with," said Adam Rodman, an expert in internal medicine at Beth Israel Deaconess Medical Center who helped design the study.

Going forward, Goh said those factors, such as a lack of appropriate training or bias, should be studied to determine whether they make a difference in doctors' diagnoses with AI technology. There are also more questions for physicians to answer after a diagnosis, and AI could potentially help in the future.

"What are the correct treatment steps to take?" Goh said. "What are the tests and such to order that would help you guide the patient towards what to do next?"

2 ways to prepare doctors for an AI-assisted future

By John League and Sarah Roller

Headlines that say "AI defeats doctors" ignore what is really going on in this study and the obstacles to realizing AI's clinical value for physicians.

Yes, physicians struggled to use AI to improve their unassisted diagnostic performance, and the use of AI didn't show any time savings. But those aren't permanent conditions. The looming challenges of physician shortages, data and information deluge, and patient access only amplify the need for better understanding of how physicians can make best use of this technology.

This study is somewhat surprising because it conflicts with previous research on the performance of physicians assisted by AI to create differential diagnoses. A 2023 study showed that when using a large-language model (LLM) adapted for diagnostic reasoning, clinicians assisted by the LLM had 51.7% diagnostic accuracy, significantly outperforming clinicians who used standard search engines (44.4%) and unassisted clinicians (36.7%). Clinicians who used the LLM also generated "more comprehensive differential lists than those without its assistance."

1. Physician training must include training on the use of AI.

One of the biggest keys to the performance of a LLM on any task is the strength of the entry prompt. In this case, the prompt given to the LLM when measuring its standalone diagnostic performance was tailored to increase the chances that the LLM would perform well.

But that's not how the doctors in the study used the LLM. Most physicians in the study treated the LLM like a search function, much as they would if searching on Google or UptoDate. Rather than copy-and-pasting the entire vignette — a best practice that was included in the standalone prompt — physicians prompted the LLM to respond to questions about individual characteristics extracted from each vignette. This limited the value of the LLM to physicians, and likely contributed to the lack of time savings between the two physician groups.

This underscores the need for physician education on how to best leverage LLM technology — both in prompt design and efficiency. We would never expect someone who didn’t know how to use a tool to create quality work with that tool, much less a tool used in a dynamic and complex clinical setting. The study authors suggest that organizations should "invest in predefined prompting for diagnostic decision support integrated into clinical workflows and documentation, enabling synergy between the tools and clinicians" — and we agree.

Importantly, we see a role for organizations across the healthcare ecosystem to play a role in this training. It should begin in medical education but must continue into physician practice, as the technology will continue to evolve. This creates an opportunity for digital health and health IT vendors to partner with provider organizations to better train physicians on their tools.

2. Physicians must adopt a broader mindset.

Even with prompt engineering training, significant adaptive challenges requiring a broader mindset shift remain to get physicians to accept — or even to engage with — recommendations from AI.

Medical education often emphasizes diagnosis as a primary role of physicians — and one that distinguishes the physician from others on the care team. Over time, this creates a cycle where physicians' experience and intuition rise above other factors in making diagnoses. We see this play out in this study. When offered an alternative diagnosis that conflicted with their own diagnosis, physicians "didn't listen to AI when AI told them things they didn't agree with, " according to The New York Times.

At the end of the day, embracing assistive technology like AI will require a broader identity shift from physicians around their role and their limitations. Many of the AI applications currently in use and development focus on easing or even eliminating administrative tasks. Those have value, but the ceiling on their ability to help organizations confront workforce and access challenges at scale is relatively low.

There is a much higher ceiling on the potential for AI to augment and improve clinical decision making, especially for physicians. Healthcare approaches to AI should focus on finding the upper limit of that ceiling, both technically and adaptively. And provider organizations must start building the infrastructure now to support their physicians in embracing these shifts.