Investigating choices regarding the accuracy-transparency trade-off of AI-based systems across contexts

Abstract

Artificial intelligence (AI) is increasingly used in decision-making. However, choosing between different algo rithmic methods underlying AI-based systems involves trade-offs. The accuracy-transparency trade-off is one of the most prominent: the most accurate approaches are often the least transparent, and the most transparent ones are the least accurate. This study examined how individuals navigate this trade-off from a deployer perspective. In an experimental between-participants online study (N = 468), we examined how framing (framing the system performance as accuracy rate vs. error rate), accountability (being able to justify a decision vs. no need to justify a decision), and the context of use (medicine, hiring, finance, law) affect choosing between different versions of systems underlying the accuracy-transparency trade-off. We also investigated whether the experimental ma nipulations and system choice affected trustworthiness and trust perceptions. Regarding the system choice (i.e., a preference for accuracy at the expense of transparency or a preference for transparency at the expense of ac curacy), framing and accountability did not affect system choice. As expected, participants favored high per formance in medicine compared to the other contexts. The results also supported the expected relationship between system choice and perceptions of different system trustworthiness facets, as well as framing effects on perceived trustworthiness and trust. We conclude that the context of use is critical for deployer preferences regarding system accuracy and transparency. Additionally, we identified person-related factors influencing such choices. Furthermore, a simple change in wording (i.e., without changing the system properties) can affect in dividuals’ perceived trustworthiness of AI-based systems.