Abstract

Abstract for: Framing the commons - How collective decision changes by Generative AI on a common pool resource problem

Introduction Shared resource problems are difficult to manage because decisions that seem reasonable for each actor can still damage the whole system by degrading the common pool resource over time.
Approach This paper examines whether generative AI agents make different group decisions when the same common pool resource management problem is presented in different ways. We analyze eight multi-agent discussion runs designed around a shared agricultural river basin. Across the runs, the overall discussion structure was similar, but the role descriptions and the way information was presented were changed.
Results In seven of the eight scenarios, the agents supported the low restricted policy option, even when evidence related to water scarcity such as drought risk, water shortages, crop losses, price increases, and higher pumping costs were clearly described. Only the framing centered on responsibility for the long-term health of the basin led the full group to support the more protective and restrictive option.
Discussion These results suggest that group decisions by AI agents can be sensitive to framing, and that some forms of framing matter more than others. This study presents an early step toward understanding how generative AI agents may behave when used as decision makers in policy making in the context of shared-resource problems, with possible relevance for policy exploration, decision support, and social simulation.

Introduction	Shared resource problems are difficult to manage because decisions that seem reasonable for each actor can still damage the whole system by degrading the common pool resource over time.
Approach	This paper examines whether generative AI agents make different group decisions when the same common pool resource management problem is presented in different ways. We analyze eight multi-agent discussion runs designed around a shared agricultural river basin. Across the runs, the overall discussion structure was similar, but the role descriptions and the way information was presented were changed.
Results	In seven of the eight scenarios, the agents supported the low restricted policy option, even when evidence related to water scarcity such as drought risk, water shortages, crop losses, price increases, and higher pumping costs were clearly described. Only the framing centered on responsibility for the long-term health of the basin led the full group to support the more protective and restrictive option.
Discussion	These results suggest that group decisions by AI agents can be sensitive to framing, and that some forms of framing matter more than others. This study presents an early step toward understanding how generative AI agents may behave when used as decision makers in policy making in the context of shared-resource problems, with possible relevance for policy exploration, decision support, and social simulation.