Co-Overcooked: Cognitive Constraints on Partner Modeling in Human-AI Team Composition

Effective collaboration requires accurate mental models of one’s partners. Current AI agents, however, may not model humans the way humans model each other under real-time constraints, creating an asymmetry in human-AI teamwork. Prior research has focused on one-human-one-AI settings, leaving open how coordination changes when multiple humans work with multiple AI agents simultaneously. We developed Co-Overcooked, a four-player cooking game where teams of humans and LLM agents must coordinate in real time. In a within-subjects experiment (N=40), participants experienced four team compositions varying in human-to-AI ratio: four humans (H4A0), three humans with one AI (H3A1), two humans with two AIs (H2A2), and one human with three AIs (H1A3). Results revealed non-linear patterns: H3A1 and H1A3 teams performed worse than pure AI teams, while H2A2 teams showed better performance through spontaneous one-human-one-AI pairing strategies. We identified expectation conflict, where multiple humans hold divergent models of a shared AI partner, as a coordination difficulty specific to hybrid teams, and propose cognitive ownership as a design principle for human-AI team configuration.