AI must outperform physicians
Both physicians and the general public want AI in healthcare to be more accurate than humans. At the same time, trust in the technology is only moderate. That is the finding of a study from the University of Gothenburg.
In spring 2025, a survey was sent to 1,000 randomly selected people across Sweden, half of them physicians. It was completed by 45 percent of the physicians and 31 percent of the general public. Participants were asked to assess different healthcare scenarios and indicate to what extent it is acceptable for an AI system to miss or misjudge a case, compared with the current level of accuracy in healthcare.
Higher standards
The results show that expectations for AI are high, especially in serious situations such as chest pain. Many members of the public want no cases to be missed, while physicians accept a small margin of error, reflecting different ways of weighing the risk of missing something against the need to avoid unnecessary diagnostic workups.
![[none]](/sites/default/files/styles/50_10_6_xmedium_1x/public/2026-04/Rasmus%20Arvidsson%20foto%20Oskar%20Arvidsson%2016x9.jpg?h=9fac7430&itok=Mu26b8h7)
Rasmus Arvidsson, a doctoral student at the Sahlgrenska Academy, University of Gothenburg, and a specialist in general practice, is one of the researchers behind the study:
“It is easy to build an AI that never misses a serious disease: just make it say that everyone is sick! But then it becomes useless. The challenge is to also identify the healthy people. The balance between false positives and missed cases requires an open discussion among developers, medical staff, and patients to find the right balance,” says Rasmus Arvidsson.
Moderate trust
Many respondents said they were already using AI, but trust in the systems was only moderate. Few respondents reported high trust. Among physicians, trust in chat-based AI tools was roughly the same as trust in the AI-based systems that are now often used to interpret ECGs. Just over seven in ten physicians had tried chat-based tools, but few used them in clinical decision-making. In the population as a whole, about one in ten people had used AI for health advice.
The study is published in BMJ Health Care Informatics. The response rate is in line with similar studies, but it does entail some uncertainty about how well the results reflect the population as a whole.
Article: Acceptable accuracy for medical AI: a survey of physicians and the general population in Sweden https://doi.org/10.1136/bmjhci-2025-101899
