When AI Flags a Brain Bleed and the Radiologist Disagrees, Jurors Notice

Picture a radiologist at 2 a.m., scanning a head CT for signs of bleeding. An AI system flags the image as abnormal. The radiologist looks, sees nothing alarming, and moves on. Weeks later, the patient has irreversible brain damage. A jury convenes. And here is the question that now matters more than whether the AI was right: how many times did the radiologist look?

That scenario, drawn from a new study published March 10 in Nature Health, captures a legal dimension of medical AI that few hospitals have reckoned with. The research, led by Michael Bruno at Penn State College of Medicine alongside colleagues from Brown University and Seton Hall University School of Law, found that the structure of a clinician's workflow when using AI tools has a measurable effect on how mock jurors assign blame.

One look versus two: a 22-percentage-point swing

The team recruited 282 participants and randomly assigned them to read one of two versions of a hypothetical malpractice case. In both, a radiologist failed to detect a brain bleed on a CT scan that an AI system had correctly flagged as abnormal. The patient suffered irreversible brain damage.

The difference between the two scenarios was small but consequential. In the first, the radiologist reviewed the scan once, after the AI alert. In the second, the radiologist read the scan twice: once independently, then again after receiving the AI feedback. In both cases, the radiologist reached the same wrong conclusion.

The results were stark. Nearly 75% of mock jurors found the radiologist failed their duty of care in the single-review scenario. That figure dropped to 53% when the radiologist had reviewed the scan twice. Same error, same outcome for the patient, but a 22-percentage-point difference in perceived liability.

The compliance trap

The study did not explore why jurors reacted this way, but the implication is clear: reviewing a scan twice signals diligence. It tells a jury that the physician exercised independent judgment rather than passively deferring to a machine. And that perception matters enormously in a legal system where malpractice cases hinge on whether a doctor met the standard of care.

But there is a cost to building workflow safeguards around legal perception. Grayson Baird, associate professor of radiology at Brown University and a co-author, pointed to a troubling incentive structure already forming around AI in clinical practice. Radiologists face mounting pressure not to disagree with AI systems. If a physician overrides an AI recommendation and turns out to be wrong, that decision becomes a powerful weapon in litigation. The result is a subtle bias toward compliance, where doctors follow the algorithm not because they trust it more than their own judgment, but because the legal cost of dissent is too high.

That compliance has downstream effects. Patients may undergo unnecessary follow-up imaging, biopsies, or monitoring. Healthcare costs rise. And the physician's role as an independent decision-maker erodes.

Why radiology is the test case

The researchers chose radiology deliberately. AI integration in imaging is further along than in most other medical specialties, making it a plausible scenario for jury evaluation. And since most malpractice cases settle out of court or take years to resolve, using hypothetical cases was the only way to study jury reasoning in real time.

This study builds on earlier work by the same team. In a previous experiment using the same hypothetical case, they found that mock jurors were less likely to find a radiologist liable when the physician agreed with the AI interpretation versus when they disagreed. Liability perceptions also shifted when jurors were told about AI error rates compared to when error rates were unknown.

Taken together, the findings paint a picture of a legal landscape that is already responding to AI in medicine, even before most courts have heard actual cases.

What hospital administrators should be weighing

For hospital systems deciding whether to adopt AI diagnostic tools, the study offers a concrete variable to consider. Brian Sheppard, professor of law at Seton Hall and a co-author, framed it in terms of cost-benefit analysis: stakeholders purchasing AI products, designing clinical workflows, or settling malpractice claims can now weigh these perceptual dynamics with actual data.

The practical takeaway is that workflow design matters as much as the technology itself. A system where the radiologist reviews imaging independently before receiving AI input may carry less legal risk than one where the AI alert comes first. But dual-review workflows also take more time in a field already strained by volume and burnout.

The limits of a mock courtroom

Several caveats apply. The study used mock jurors, not actual trial participants, and the scenario was hypothetical. Real malpractice cases involve far more complexity: expert testimony, cross-examination, institutional policies, and the emotional weight of a living plaintiff. The sample of 282 participants, while adequate for statistical analysis, cannot capture the full range of jury demographics and attitudes toward technology.

The study also measured only one specialty and one type of error. Whether these findings generalize to other AI applications in medicine, such as pathology, cardiology, or clinical decision support, remains an open question.

And perhaps most importantly, this research measured perception, not legal outcomes. How actual judges and juries will treat AI-related malpractice cases as they begin appearing in court is still largely unknown.

A moving target

Michael Bernstein, associate professor of radiology at Brown and the study's corresponding author, emphasized that public perception of AI is shifting rapidly alongside the technology itself. What jurors believe today about physician responsibility when using AI may look different in five years, as familiarity with these tools grows and legal precedent accumulates.

For now, the study offers one of the clearest signals yet that the legal risks of medical AI are not just about whether the algorithm works. They are about how the human uses it, and how that use looks to a room full of people deciding who is at fault.