Abstract: AI algorithms are increasingly used for negotiating agreements. Responses to an offer may depend not just on whether the counterparty is AI-assisted or not, but also whether the counterparty was responsible for the choice to delegate decisions to the algorithm. For example, people may interact with a chatbot in a similar way to a human in principle, but react negatively when it is salient that the chatbot has been imposed upon them. On the other hand, people may appreciate a human interaction more when they aware that a chatbot was a viable alternative.
This study uses variations on the Ultimatum Game as a framework to investigate related questions. Specifically, how do Proposers in the game delegate decisions to artificial intelligence algorithms, and how do Responders react to this? We use two alternative "algorithms", one trained to behave as a typical human (simply choosing the modal offer from the baseline treatment), and the other trained to optimize profits (choosing the profit maximizing offer based on acceptance rates from the baseline). There is common knowledge about how each algorithm makes decisions, but neither the Proposer nor Responder knows precisely which offer either will make. Five treatment variations allow us to disentangle the effects of algorithm-based offers versus delegation decisions. We also elicit beliefs about the choices of the algorithms and Proposer and Responder behaviour to pinpoint the motivations for delegation, and further investigate the causes of differential reactions by Responders across treatments.
Joint work with Francesca Marazzi, Jesper Roine, Giancarlo Spagnolo