Artificial Agents Mitigate the Punishment Dilemma of Indirect Reciprocity

Altruistic cooperation is socially desirable yet costly, thereby challenging to promote in multiagent systems. Indirect reciprocity (IR), where the decision to cooperate or defect is based on reputations, serves as a key mechanism to elicit cooperation among selfish agents. However, IR faces challenges under private assessment, due to the so-called punishment dilemma: without mechanisms forcing reputation consensus, disagreements will emerge, resulting in apparently unjustified defections which are punished. Following the increasing prevalence of hybrid systems, where artificial agents (AAs) coexist with humans, we aim to understand the role of AAs in alleviating IR's punishment dilemma and improving cooperation. We develop an analytical evolutionary game-theoretical model to study cooperation under IR with private assessment. A fixed-strategy AA is embedded within an adaptive population, the latter simulating a population of humans adapting over time. We show that limited interactions with the AA are sufficient to impact the distribution of reputations in a population, allowing justified defection to be widely recognized and fostering cooperation. This work highlights the potential of using artificial agents, even with simple fixed strategies, to impact humans' moral assessments, generate reputation consensus and promote cooperation.