Fairness and Trust in Human Bot Interaction---Investigating the Consequence of Manipulating Norm Consensus on Trust Allocation

As machines become increasingly intelligent, many roles that have been traditionally taken by people are now supplemented or augmented by machines and algorithms. Such roles require various levels of interaction between people and machines/algorithms that make decisions, thus research into human-machine interaction is timelier than ever. Recently, research on this topic has adopted methods from behavioral economics and behavioral psychology to study how people react and interact with bots by employing incentivized games. This interdisciplinary approach is gradually recognized and increasingly adopted by the academy and industry to understand the underlying mechanisms that drive differences from platform to platform, and to forecast each platform's performance. We build on recent work by Jordan and colleagues \cite{Jordan} to investigate the behaviors of fairness, punishment and trust when people are interacting with bots. We carried out an online experiment that includes a two stage game following \cite{Jordan}; a third-party-punishment game followed by a trust game with approximately 3500 participants who interacted with one another and/or with bots that made pre-determined decisions over the summer of 2019. We are building on this prior work in this study where we now investigate the impact of manipulating norm-consensus (in the form of injunctive norms, i.e., what one should do), and their relationship to the trust one gains when being seen as behaving according to the consensus. We are conducting a within-person design where we invite participants from the first experiment conducted a year ago, and give them norm-consensus information. We expect that the information-treatment we administer will increase the trust-gain gap between those who behave according to the consensus-behavior compared to those who do not behave in such a manner as compared to the setting that participants made decisions a year ago, receiving no such information. Out of the 3 experimental conditions, we anticipate such effects where prior measurement shows lower levels of believed consensus. In the third condition, where we give participants truthful information on the basis of prior measurement we do not have clear expectations.