This study investigates how humans interact with bodiless AI agents exhibiting strategic behavior. It examined whether social norms of fairness and altruism are applied to these agents compared to humans, and if such responses can be characterized as anthropomorphic. 96 subjects went over 52 identical rounds of repeated Ultimatum or Dictator Games against both human and/or machine agents. The experimental design manipulates three factors: economic interaction, cheap talk, and matching conditions, with Tic Tac Toe used as cheap talk to signal the strategic sophistication of agents.
The findings suggest that in the absence of cheap talk, subjects apply social norms similarly to machines and humans. When human agents are present, participants exhibited favoritism towards them, indicating an in-group/out-group bias. When cheap talk was implemented, subjects interacted with two types of machine agents. AI agents were competent at Tic Tac Toe, whereas non-AI agents made moves randomly. When strategic competence was inferred through cheap talk, AI agents received more favorable treatment compared to non-AI agents. Notably, when subjects interact exclusively with machine agents in strategic bargaining, they display payoff-maximizing behavior regardless of the machine’s perceived intelligence. AI agents are perceived as more intelligent and anthropomorphic than non-AI agents, though this perception is influenced by the presence of human agents and is maximized in their absence.
