Humans Forgo Reward to Instill Fairness into AI

Abstract

In recent years, artificial intelligence (AI) has become an integral part of our daily lives, assisting us with decision making. During such interactions, AI algorithms often use human behavior as training input. Therefore, it is important to understand whether people change their behavior when they train AI and if they continue to do so when training does not benefit them. In this work, we conduct behavioral experiments in the context of the ultimatum game to answer these questions. In our version of this game, participants were asked to decide whether to accept or reject proposals of monetary splits made by either other human participants or AI. Some participants were informed that their choices would be used to train AI, while others did not receive this information. In the first experiment, we found that participants were willing to sacrifice personal earnings to train AI to be fair as they became less inclined to accept unfair offers. The second experiment replicated and expanded upon this finding, revealing that participants were motivated to train AI even if they would never encounter it in the future. These findings demonstrate that humans are willing to incur costs to change AI algorithms. Moreover, they suggest that human behavior during AI training does not necessarily align with baseline preferences. This observation poses a challenge for AI development, revealing that it is important for AI algorithms to account for their influence on behavior when recommending choices.