Abstract Models of Human-AI Teams
Proceedings of the First WI-IAT 2025 International MASST Initiative Workshop
2025
Abstract
Understanding the effectiveness of a human-AI team design before deploying it typically requires building candidate AI agents and running experiments with human users, a process that is time consuming and expensive. We instead propose creating abstract computational models of human-AI team behavior to explore and evaluate the space of alternative interaction designs. Building on Ohlsson and Jewett's notion of abstract models of behavior, we develop such models using an extension of Petri Nets built on the Machinations framework, representing each team member as a token that flows through a sequence of tasks informed by a cognitive task analysis. We illustrate the approach with a model of the cooperative game Dice Adventure, showing how simulation can predict a team's completion time and probability of success without implementing AI teammates or running human-subjects studies. We discuss related approaches, including GOMS and joint activity graphs, and outline plans to extend the framework to larger teams, agent communication, and organizational structures involving authority relationships.
BibTeX
@inproceedings{langley-maast-2025,
title = {Abstract Models of Human-AI Teams},
author = {Langley, Pat and MacLellan, Christopher J.},
booktitle = {Proceedings of the First WI-IAT 2025 International MASST Initiative Workshop},
year = {2025},
}