Learning with Conflicts of Interest

Financial, social, and political factors often prevent the in- terests of the owners of ML systems and their users from being perfectly aligned. These systems often produce biased information that can influence users to make decisions that are not in their best interest, and vice versa. Current solu- tion approaches require ML systems to implement protocols to mitigate their biases, or attempt to compel users to report data honestly. However, system owners and users usually have little incentive to implement these protocols. We believe that a successful solution to this problem must recognize the con- flict of interest between the systems and their users, and use this information to protect both systems and users against in- formation that adversely influences their decisions. Toward this end, we propose a game-theoretic framework that models the interaction between supervised learning systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired in- formation and actions and minimize biased and manipulative actions in interaction with supervised learning systems.