I propose to investigate how roleplay between humans and large language model (LLM)–based personas can support productive, societally beneficial, and cooperative negotiation agreements between members of distinct social groups under conditions of social inequality. I ask – How do authority levels and social group membership affect goal achievement, collective benefits, and fairness of communication in negotiations with both competing and cooperative objectives? And, to what extent can this be early detected by LLMs? Successful negotiations require human participants to perceive LLM personas as trustworthy, reliably and credibly playing their assigned roles with appropriate appearances and communication styles that do not exhibit biased behavior, enabling a greater immersion experience and improved training effectiveness. By identifying competitive and cooperative language, I aim to develop communication models that optimize for multiple negotiation objectives that align with human values, and mediator interventions that mitigate unconscious bias. My approach will assume (and I will investigate whether) people are inherently utility seeking as well as trusting and altruistic toward others, society and the earth. Specifically, my framework will support active LLM mediator intervention in a two player negotiation training game, featuring a human participant in a role with higher authority and an LLM persona in a lower authority role, to mitigate the often unconscious tendencies toward bias as well as support cooperation.This project unifies studies of algorithmic bias in modern AI systems with understanding communicative behavior, a combination that far outstrips the specific test cases I explore.