Design and Effects of Co-Learning in Human-AI Teams

Abstract

The rapid progress of artificial intelligence (AI) will increase opportunities for humans and AI-driven technology to collaborate as teammates. This requires both partners to learn from interactions about the task, each other and the team (co-learning). Co-learning can be supported by enabling partners to share knowledge and experiences on the task and team level. This paper first analyzes the requirements regarding tasks and environments for co-learning. These requirements were subsequently implemented in a testbed: a human and intelligent robot jointly conducting an urban search and rescue task in a simplified task environment. We designed Learning Design Patterns (LDPs): interaction sequences intended to initiate and facilitate co-learning. Effects of LDPs on collaboration, knowledge and understanding, and team performance were experimentally evaluated using the testbed. In comparison to a previous study, participants appreciated the robot more, had more interaction and displayed more commitment. Results show evidence that the LDPs, in comparison with no interventions, initiated and improved learning of the human team member, in particular on knowledge development and understanding the partner. Better knowledge and understanding did, however, not also lead to better team performance. Implications for co-learning in human-AI teams and for learning-supporting interventions are discussed.