WP5 examines how generative AI influences polarization when individuals encounter differing identities and opinions in their offline daily lives. It focuses on face-to-face conversations and the exchange of viewpoints across lines of difference. WP5 first categorizes how individuals utilize generative AI to support inclusive communication where disagreement is anticipated; it then tests which patterns of use contribute most effectively to inclusive practices and group deliberation. In this WP, we address the following research questions:

• RQ1: How do individuals utilize generative AI when those with differing social identities and political orientations must discuss issues where disagreement is anticipated?
• RQ2: How do varying patterns of generative AI utilization influence individual communication behaviors, the effectiveness of inclusive group deliberation, and deliberative outcomes—specifically regarding perspective shifts, assimilation, and the understanding and acceptance of outgroups?
• RQ3: What patterns of generative AI engagement promote inclusive communication, and how can choice architecture be designed to nudge users toward these prosocial interaction strategies?

Towards these goals, WP5 employs a mixed-methods approach, combining offline experiments with qualitative data analysis. The core design consists of an experiment where participants engage in group discussions on topics where disagreement is anticipated. We compare a treatment group, granted unrestricted access to generative AI (e.g., ChatGPT) prior to the discussion, with a control group where AI use is restricted. The study evaluates the efficacy of collaborative communication and deliberation, as well as post-discussion attitude shifts.

For the AI-enabled group, all prompts and AI-generated responses from a 10-minute session are collected as data. This qualitative data is subjected to content analysis to categorize primary usage patterns, which are then analyzed in relation to quantitatively measured dependent variables. Once AI usage patterns that positively influence collaborative communication and deliberative outcomes are identified, subsequent experiments will test strategies for nudging these behaviors.