Can Conversational AI loosen Us-Versus-Them Boundaries? The Effects of Common, Dual, and Separate Identity Framings on Pro-Immigrant Intergroup Helping
A preregistered experiment with 658 non-Latine White U.S. adults tested whether conversational AI can shift intergroup attitudes toward Latine immigrants using different identity framings GPT-4o was instructed to frame immigrants through common ingroup identity (shared American identity), dual identity (both Latine and American), separate identity (distinct cultural boundaries), or unrelated control topics across five dialogue rounds Common and dual identity conditions significantly reduced sepa
Analysis
TL;DR
- A preregistered experiment with 658 non-Latine White U.S. adults tested whether conversational AI can shift intergroup attitudes toward Latine immigrants using different identity framings
- GPT-4o was instructed to frame immigrants through common ingroup identity (shared American identity), dual identity (both Latine and American), separate identity (distinct cultural boundaries), or unrelated control topics across five dialogue rounds
- Common and dual identity conditions significantly reduced separate categorization and increased willingness to act, with dual identity also raising dual categorization
- Semantic similarity analysis confirmed participants' language converged with assigned narratives, and this convergence correlated positively with willingness to act
- The study reveals a gap between cognitive recategorization and actual behavior, though willingness to act was meaningfully influenced
Why It Matters
This research demonstrates that conversational AI can serve as a scalable intervention tool for reducing intergroup bias, addressing a critical need as traditional bias-reduction programs face scaling and policy constraints. For AI practitioners and researchers, it provides empirical evidence that LLM-mediated dialogues can meaningfully influence social categorization and prosocial intentions, opening new avenues for AI-driven social impact applications.
Technical Details
- Sample: Quota-representative national sample of 658 non-Latine White U.S. adults
- Model: GPT-4o used as the conversational AI agent across five dialogue rounds per participant
- Experimental design: Four conditions—common ingroup identity, dual identity, separate identity, and control (unrelated topic)—based on the Common Ingroup Identity Model
- Analysis methods: Path modeling to test indirect associations, semantic similarity analysis of transcripts to verify narrative tracking, and moderation analysis across need for closure, openness to experience, and political orientation
- Key findings: Common and dual identity conditions reduced separate categorization (vs. control); dual identity uniquely increased dual categorization; willingness to act was higher in superordinate identity conditions; indirect effects operated through reduced separate categorization
Industry Insight
- AI-powered interventions could offer a cost-effective, scalable alternative to traditional bias-reduction programs, particularly valuable in politically constrained environments where institutional interventions face resistance
- The gap between cognitive recategorization and behavior suggests that while AI can shift attitudes, additional mechanisms (e.g., commitment devices, follow-up actions) may be needed to translate willingness into actual prosocial behavior
- The consistency of effects across political orientation and personality moderators indicates broad applicability, but practitioners should consider that separate-identity framing conversations may inadvertently reinforce boundaries and should be avoided in intervention design
Disclaimer: The above content is generated by AI and is for reference only.