WP4 systematically analyze the affective framing inherent in existing user prompts, providing a rigorous diagnosis of the current state of affective polarization as reflected in these inputs. It then develop a real-time, "user-in-the-loop" prompt suggestion design based on these findings. This intervention aims to mitigate affective polarization and explore the potential of AI to facilitate inclusive interpersonal communication. To empirically verify the effects of the proposed prompt suggestion intervention framework, WP4 relies on experimental research, exploring the potential for users to refine their messages—maintaining the substance of their core arguments while reducing expressive hostility—through real-time, opt-in rewording suggestions and dialogue-enhancing prompts. The primary objective is to evaluate the efficacy of a "user-in-the-loop" prompt design that preserves user autonomy. Rather than having the AI unilaterally modify or replace content, this framework empowers users to actively calibrate their own communication, ensuring they remain the central agents in the message-generation process.

• RQ1: In what ways is affective framing manifested within existing user-generated prompts?
• RQ2: How do affective responses toward ingroups and outgroups vary according to the intensity of prompt suggestions (e.g., full generation, rewording suggestions, or no suggestion)?
• RQ3: How is the efficacy of prompt suggestions moderated by the output modality of generative AI (i.e., text, image, or video)?

To analyze the types and patterns of affective framing in existing prompts, WP4 will first collect and examine open-source prompt datasets from platforms such as Kaggle and Hugging Face. This analysis will be conducted using the LLM-based emotion classifier developed in WP1.

Second, WP4 employs a 3 by 3 between-subjects experiment to test how prompt intervention intensity and output modality (text/image/video) influence intergroup affect. By providing real-time, inclusive feedback, the study empirically measures the potential for AI-driven prompt design to reduce affective polarization.