Designing Interaction with AI for Human Learning: Towards Human-Machine Teaming in Radiology Training

Abstract

We explore the design of systems that enable humans and machines to operate as teams, exercising their different and complementary abilities to work and learn together. Machine Learning (ML) is now widely used in diverse applications such as medical image reading and autonomous vehicles, but typically, ML systems are not designed with human learning in mind, sometimes eroding or supplanting human skills, creating a whole that is less than the sum of its parts. We propose a new approach to ML/AI system design to foster human-machine mutual learning: synergistic interactions in which machines help people think critically and gain wisdom, while people help improve machine models by reframing ML tasks and immersing them in human-machine-human systems which provide feedback to the AI model while helping humans to learn. By explicitly aiming to increase human skill and wisdom, teaming goes beyond "human-in-the-loop"approaches where humans serve primarily to enhance machine performance. We contribute a conceptual model for human-machine teaming design and use a case study in radiology training to identify five critical considerations for interaction design and for how to make AI interactive: (1) human-machine dialogue (2) labelling and attention (3) problem framing (4) biases, values and affect (5) ethics, agency and human choice.