To test their hypotheses, three researchers asked respondents to participate in a scenario-based survey experiment on Mechanical Turk (MTurk), a crowdsourcing marketplace. Data from 655 MTurk respondents were used to collect the sample. Bansal, Matta, and Díaz-Ordonez used attention checks and asked follow-up questions to confirm the reliability and validity of the data.
As expected, we found that transparency is essential in the patient-provider relationship in relation to the use of AI, supporting our main hypothesis. Transparency in the use of AI is important, and it has not only increased trust in the AI used by doctors, but also in the healthcare providers who use it. Essentially, researchers have found that it’s true that if you trust your provider, you can trust the tools they use.
Even more surprising was the response to the role of accuracy in AI diagnosis and its impact on the patient-provider relationship. Increased transparency predictably increased trust, but as AI accuracy increased, trust actually decreased or stagnated. Bansal considers several possible explanations for why this is the case.
“People fear that if AI becomes too accurate, doctors will no longer make important decisions themselves and will be outsourced to AI,” he said. “Especially in primary care. I think that’s where the fear is being expressed.”
Matta said these findings could change the whole game in terms of how we think about the relationship between artificial intelligence and trust.
“The reason this is not just a big problem, but a really big problem, is that it goes against the idea that accuracy increases trust,” Matta said. “The implications are huge. Without this discovery, everyone would have feared that doctors would be replaced by AI, but now we can say, ‘Maybe not so soon.'”
These findings were presented at the Midwest Association for Information Systems (MWAIS) conference in May 2025 in Oklahoma. This year’s MWAIS 2026 conference was held at Ohio University. The research is currently available through conference proceedings, and Bansal and Matta are also working on publishing it in a journal.
The results are being borne out. Matta and Bansal have seen a number of recent papers arguing that AI-generated output is less valuable than human-generated output. Bansal agrees that these results essentially reshape the common understanding of how AI accuracy is perceived, but says how these results are interpreted is critical.
“We need to be careful how we interpret these findings because they are in the interest of transparency and accuracy,” Bansal explained. “We should not generalize that accuracy is not important in AI healthcare. In the context of transparency and accuracy, transparency is more important, and when transparency exists, accuracy plays a smaller role.”
Bansal and Matta also want to make clear that just because AI accuracy isn’t important in a primary care setting doesn’t mean it won’t be important in other medical settings as well. They hope to extend this study to specialty care to compare the findings with other medical fields.
“One study is never enough,” Bansal says. “We discovered something counterintuitive, which means we need to study it further.”
