With the dramatic progress of artificial intelligence algorithms in recent times, it is hoped that algorithms will soon supplant human decision-makers in various fields, such as contract design. We analyze the possible consequences by experimentally studying the behavior of algorithms powered by Artificial Intelligence (Multi-agent Q-learning) in a workhorse \emph{dual contract} model for dual-principal-agent problems. We find that the AI algorithms autonomously learn to design incentive-compatible contracts without external guidance or communication among themselves. We emphasize that the principal, powered by distinct AI algorithms, can play mixed-sum behavior such as collusion and competition. We find that the more intelligent principals tend to become cooperative, and the less intelligent principals are endogenizing myopia and tend to become competitive. Under the optimal contract, the lower contract incentive to the agent is sustained by collusive strategies between the principals. This finding is robust to principal heterogeneity, changes in the number of players involved in the contract, and various forms of uncertainty.

翻译：随着近期人工智能算法的显著进步，人们期望算法将很快在合同设计等各个领域取代人类决策者。我们通过实验研究基于人工智能（多智能体Q学习）算法在经典双重委托-代理问题的“双重契约”模型中的行为，分析了其可能产生的后果。我们发现，人工智能算法无需外部指导或相互通信，便能自主学会设计激励相容的契约。我们强调，由不同人工智能算法驱动的委托人可能表现出混合策略行为，如合谋与竞争。研究发现，更聪明的委托人倾向于合作，而较不聪明的委托人则内生化短视行为，倾向于竞争。在最优契约下，委托人之间通过合谋策略维持对代理人的较低契约激励。这一结论对委托人异质性、参与契约的玩家数量变化以及各种不确定性形式具有稳健性。