“Good studies come from good jokes,” Grayson Baird, PhD, says.

The joke, for years now, has been that radiologists will be replaced by artificial intelligence within five years. “That won’t happen,” Baird says. “If they got rid of radiologists, who would you sue?”

That dark punchline became, in part, the origin story of the Brown Radiology, Psychology, and Law Lab, founded by Baird and Michael Bernstein, PhD, who are both associate professors of radiology as well as psychologists. They say their training—Baird as a quantitative psychologist, Bernstein as an experimental psychologist—gives them an unusual advantage in a medical world increasingly preoccupied with AI but unsure how to think about its human consequences. As Baird puts it, psychologists know it doesn’t take much to move a person’s judgment—a number on a screen, a flag, a threshold are just stimuli. And AI, however sophisticated underneath, is stimulus too.

That deceptively simple insight has guided a body of research that asks not whether AI works, but what its presence does—to radiologists, patients, jurors, and the medical system itself.

One of the lab’s recent studies, published in NEJM AI, tested how ordinary Americans judged radiologists in hypothetical malpractice cases. Participants read vignettes about a radiologist missing either a brain bleed or lung cancer. Sometimes they were told no AI was involved. In other conditions, AI either agreed with the radiologist and missed the abnormality too, or disagreed by correctly flagging it. The result was striking: When AI found the pathology and the radiologist did not, participants were much more likely to hold the radiologist liable. If a doctor disagrees with AI and turns out to be wrong, that disagreement itself appears damning.

“So you have this AI penalty,” Bernstein says, “but what’s interesting is that this AI penalty can be mitigated by informing people what the AI error rate is. Because to most people, AI feels like magic, and once you can show people that it’s not, they start to think in more realistic ways—and that makes them become more sympathetic with a radiologist who misses it.”

“Our remedy, so to speak, is to provide context,” Baird adds. When physicians know the AI’s error rate and can relay this information in an accessible way to patients (for example, that half of its positive flags were actually false alarms), physicians are able to anchor against automation bias, the machine loses its aura of omniscience, and jurors become more sympathetic to the radiologist.

This concern has proved important on home ground. When Brown University Health briefly deployed a clinical AI tool for detecting large vessel occlusion, radiologists shut it down within three weeks. The system was, in a narrow technical sense, highly accurate. But it was also constantly “overcalling” or flagging abnormalities that were not actually there. This problem can be explained by prevalence: When a condition is rare, even a highly sensitive, highly specific test can generate a flood of false positives. An AI system may look excellent on paper and still produce a false discovery rate high enough to rattle clinicians, encourage defensive medicine, and distort judgment. Baird adds, however, that he and Bernstein “weren’t saying that AI is bad. Just that if you’re going to use it, provide context.”

In medicine, evidence usually arrives wearing a familiar white coat: randomized trials, chart reviews, epidemiologic models. But the Radiology, Psychology, and Law Lab asks questions that can’t be framed by those rows and columns. While important opinion pieces have begun to circulate and interrogate these questions, few approach them empirically. So Baird and Bernstein brought in “bread and butter” psychology methods—mock juries, vignette experiments, decision science—to study how people behave when medicine collides with law, technology, and fear.

Modern medical problems are not only biological, but also psychological, legal, statistical, and social, and thus require an interdisciplinary approach. Similarly, AI in radiology is not merely an engineering question; it is a human-factors question, a communication question, and increasingly, a liability question. If medicine is to navigate these new dilemmas well, it will need more than technical expertise. It will need a greater willingness to welcome other disciplines into the room.