Exploiting the machine: Human cooperation with artificial agents in prisoner's dilemma

We conduct a laboratory experiment using one-shot Prisoner's Dilemma (PD) and finitely Repeated Prisoner's Dilemma (RPD) games to study cooperation with an artificial agent that both mimics human behavior and acts on behalf of a human stakeholder. Controlling for beliefs, emotions, and personal characteristics, we investigate the underlying drivers of cooperation differences between a human baseline (“Humans”) and the artificial agents (“AA”) treatment. We find that the likelihood of cooperation does not depend on whether the counterpart is human or artificial in the one-shot PD games. In the RPD games, however, cooperation is less likely when participants play with an artificial agent than when they play with other humans. By combining individual decisions and beliefs, we identify an “exploiting-the-partner” behavior: choosing to defect while expecting the partner to cooperate. Our results suggest that participants are persistently more likely to exploit artificial agents than human partners when they anticipate cooperation, including across repeated interactions and even if the artificial agent is acting on behalf of a human counterpart.