Onboarding Materials as Cross-functional Boundary Objects for Developing AI Assistants
Carrie J. Cai, Samantha Winter, David F. Steiner, Lauren Wilcox, Michael Terry
Abstract
Deep neural networks (DNNs) routinely achieve state-of-the-art performance in a wide range of tasks, but it can often be challenging for them to meet end-user needs in practice. This case study reports on the development of human-AI onboarding materials (i.e., training materials for users prior to using an AI) for a DNN-based medical AI Assistant to aid in the grading of prostate cancer. Specifically, we describe how the process of developing these materials changed the team’s understanding of end-user requirements, contributing to modifications in the development and assessment of the underlying machine learning model. Importantly, we discovered that onboarding materials served as a useful boundary object for cross-functional teams, uncovering a new way to assess the ML model and specify its end-user requirements. We also present evidence of the utility of the onboarding materials by describing how it affected user strategies and decision-making with AI in a study deployment to pathologists.