AI, Radiology Workflow, and Liability: Michael Bernstein, MD

As artificial intelligence (AI) becomes more embedded in diagnostic imaging, questions are emerging about how AI affects physician decision-making, patient safety, and legal liability.¹ New research suggests that seemingly small workflow choices—such as whether radiologists review images once or twice when using AI—can shape how jurors interpret responsibility when diagnostic errors occur.

In this interview with The American Journal of Managed Care^® (AJMC^®), Michael Bernstein, MD, an experimental psychologist and assistant professor in the department of diagnostic imaging at the Warren Alpert Medical School, discusses how AI-assisted workflows influence liability perceptions, cognitive bias, and the future of radiology practice.

This transcript was lightly edited for clarity.

AJMC: Your study suggests that workflow, specifically whether radiologists interpret images once or twice when using AI, can influence how jurors perceive liability. How might these findings inform best practices for integrating AI into radiology workflows in real-world clinical settings?

Bernstein: These findings point to a likely trade-off that radiology practices should consider. While having radiologists interpret imaging twice will increase read time, our work suggests that it will decrease legal liability when the radiologist ultimately fails to catch a pathology.

AJMC: The results show that jurors were significantly more likely to side with the plaintiff when radiologists performed a single read after seeing AI feedback. What does this suggest about how clinicians should balance efficiency with safeguards against automation bias when using AI-assisted diagnostics?

Bernstein: Yes, this is exactly what I was getting at above. However, the question of how to balance these factors is a consideration that can only be resolved by individual practices. The pros and cons of weighing legal liability with efficiency will vary from one setting to another, and each must consider those trade-offs independently.

AJMC: AI is often promoted as a tool to improve diagnostic accuracy and reduce missed findings. Based on your findings, do you think certain workflow models, such as independent reads before consulting AI, could also help mitigate diagnostic risk and improve patient safety?

Bernstein: AI is indeed often discussed in that manner, and there is emerging evidence to suggest that radiologists have greater diagnostic accuracy, in totality, when using an AI. Interestingly, though, our work showed that in the circumstances where AI makes a mistake, it can cause radiologists to also make an error when they were correct without the use of AI.

More research is needed to examine this and related questions empirically. However, based on decades of research in cognitive science, it is likely the case that specific AI workflow models could be optimized to improve a radiologist's diagnostic performance.

AJMC: Public trust plays a critical role in the adoption of AI in health care. Do you think patients may view AI-assisted diagnostics as increasing accuracy and safety, or could high-profile malpractice cases involving AI undermine confidence in these technologies?

Bernstein: Some of our other experiments found that when people were aware of the AI error rate, they were more sympathetic to a radiologist who incorrectly disagreed with the AI. As we allude to in the Globe piece, this may be due to the fact that people have "magical thinking about AI." In other words, with the understandable hype of AI, people are vulnerable to believing that AI is more accurate than it really is. If true, then patients might have unrealistically high expectations about what AI can accomplish, which could help explain why radiologists are viewed as more liable when they contradict AI in making a false negative interpretation (FN) relative to making an FN interpretation absent AI.

However, informing people of the error rates of AI may reduce this inflated view of what AI can accomplish. In other words, I certainly think it is the case that patients may view AI-assisted diagnostics as increasing accuracy and safety, but it's important for everyone to understand the limitations of AI so that expectations can be appropriately calibrated. And our research suggests that doing so (by informing people of the error rates, as mentioned above) mitigates the radiologist's legal liability.

AJMC: The study highlights the potential for AI to influence physician decision-making and cognitive biases such as anchoring or automation bias. What safeguards (eg, technological or procedural) do you think are most important to ensure AI supports, rather than overrides, clinical judgment?

Bernstein: This is an ongoing area of investigation. The study obviously points to one possible solution: having radiologists interpret images twice: once before, then with AI. Another study we recently published suggests that if AI is going to be used in a triage setting for detecting abnormalities with a low prevalence rate, like screening mammography, we should pay close attention to the triage threshold. The threshold refers to the severity score that is used to distinguish which images are deemed as negative for pathology and do not require radiologist interpretation vs those that are deemed as higher risk for pathology and do require radiologist interpretation.

This is yet another trade-off that practices must consider for themselves—more conservative thresholds will result in fewer instances of AI incorrectly determining a case is negative (and thus no radiologist interpretation is needed), but at the cost of more total cases a radiologist will need to interpret. Other, related projects on this topic are underway, and I'm more than happy to discuss those results once published.

AJMC: Looking ahead, what additional research is needed to understand how AI integration in radiology affects not only liability but also real-world patient outcomes, such as diagnostic accuracy, treatment timing, and long-term health results?

Bernstein: Examining how AI integration impacts patient outcomes often requires that investigators utilize what is called a multi-case multi-reader design. In this design, radiologists interpret the same set of cases under different AI integration models so that we can compare diagnostic performance. These studies can be time- and resource-intensive but are the best way to examine how a radiologist's performance differs between various AI-integrated workflow options.

Another area that is worth exploring is AI integration into a patient's online portal. For instance, we conducted one study that looked at (among other outcomes) how the inclusion of AI reports into a patient portal for mammography would impact a patient's follow-up behavior (eg, seeking a second opinion).

References

1. McCrear S. AI double reads may reduce radiologist malpractice liability. AJMC^®. March 11, 2026. https://www.ajmc.com/view/ai-double-reads-may-reduce-radiologist-malpractice-liability