How should artificial intelligence be used in breast screening? Women's reasoning about workflow options

Studies show that breast screening participants are open to artificial intelligence (AI) in breast screening, but hold concerns about AI performance, governance, equitable access, and dependence on technology. Little is known of consumers' views on how AI should be used in breast screening practice. Our study aims to determine what matters most to women regarding AI use in the workflow of publicly funded breast screening programs, and how women choose between workflow options. We recruited forty women of screening age to learn about AI, the Australian breast screening program, and four possible workflows that include AI - one where AI works alone, and three different combinations of humans and AI. Participants then joined one of eight 90-minute dialogue groups to discuss their normative judgements on workflow options. Women proposed four conditions on AI deployment: preserving human control, evidence to assure AI performance, time to become familiar with AI, and clearly justifying the need for implementation. These informed women's unified rejection of AI working alone, and divided preferences across the other three workflows, as they traded off workflow attributes. Current evidence on AI performance convinced some women, but not others. Most women believed humans mitigate risk the best, so workflows should continue to be designed around them. Public breast screening services are trusted and valued by women, so significant changes require careful attention to outcomes relevant to women. Our results - women's detailed judgements on workflow design options - are new to the research literature. We conclude that women expect that AI only be deployed to do tasks it can do well, only where necessary, and only to fill gaps that radiologists cannot meet. Advancements in AI accuracy alone are unlikely to influence all women to accept AI making final decisions, if clinicians are available to perform the same task.