Rewards and punishments help humans overcome biases against cooperation partners assumed to be machines

High levels of human-machine cooperation are required to combine the strengths of human and artificial intelligence. Here, we investigate strategies to overcome the machine penalty, where people are less cooperative with partners they assume to be machines, than with partners they assume to be humans. Using a large-scale iterative public goods game with nearly 2,000 participants, we find that peer rewards or peer punishments can both promote cooperation with partners assumed to be machines but do not overcome the machine penalty. Their combination, however, eliminates the machine penalty, because it is uniquely effective for partners assumed to be machines and inefficient for partners assumed to be humans. These findings provide a nuanced road map for designing a cooperative environment for humans and machines, depending on the exact goals of the designer.