Gender Stereotype Threat in AI-Assisted Negotiation

Abstract

Researchers have long used negotiation as a testbed for intelligent virtual agents – studying social intelligence, creating training paradigms for human negotiators, and leveraging IVAs as methodological tools to study bias. In this study, we analyze a collection of negotiation dialogues, wherein participants negotiated against an opponent presenting as male or female; further, we manipulated whether participants were told their counterpart used AI to assist them. We investigate whether AI-assistant use may impact stereotype threat in three ways: 1) Through self-reported feelings; 2) In the language they use against their opponent; and 3) Whether their objective performance at the task shows signs of threat. First, through self-reports, we see evidence consistent with stereotype threat, where women report higher discomfort than men. Next, a textual analysis of the dialogues suggests the AI-assistance intervention, gender matching, and one’s own gender interact to influence the language one uses during the task – specifically, influencing whether they speak more or less cautiously. Lastly, we find an interaction between whether one’s opponent purportedly used AI and one’s gender, reflected in their negotiated agreement.