What Lies Beneath? Exploring the Impact of Underlying AI Model Updates in AI-Infused Systems

As AI models evolve, understanding the influence of underlying models on user experience and performance in AI-infused systems becomes critical, particularly while transitioning between different model versions. We studied the influence of model change by conducting two complementary studies in the context of AI-based facial recognition for historical person identification tasks. First, we ran an online experiment where crowd workers interacted with two different facial recognition models: an older version and a recently updated, developer-certified more accurate model. Second, we studied a real-world deployment of these models on a popular historical photo platform through a diary study with 10 users. Our findings sheds light on models affecting human-AI team performance, users' abilities to differentiate between different models, the folk theories they develop, and how these theories influence their preferences. Drawing from these insights, we discuss design implications for updating models in AI-infused systems.