Does AI roleplay that softens prejudice work differently on hot-button issues like immigration than on calmer topics?

This explores whether AI-simulated intergroup contact, where you talk with an AI playing a member of another group to soften prejudice, has weaker, stronger or different effects when the topic is emotionally charged, like immigration. The short answer is that none of the retrieved notes covers synthetic contact or its results broken down by topic. Nothing here tells us whether immigration conversations behaved differently. Any direct answer to the question would be invented. What the collection does offer is a set of nearby findings. They suggest why charged topics might behave differently, and they point to what to look for in the original study.

The most relevant clue is about how AI models handle emotion and sensitive subjects. GPT-4 tends to pull negative prompts back toward neutral or positive replies, an 'emotional rebound'. That pattern switches off on sensitive topics, where safety training takes over from tone (Does emotional tone in prompts change what information LLMs provide?). For synthetic contact, this means the AI 'contact partner' may itself act differently on immigration than on low-stakes topics. It might hedge more, show less of a real personality, or feel more scripted. So a difference in outcomes could come from the AI's behavior as much as from the human participant's feelings.

A second angle asks whether AI-voiced contact can have real social effects at all. One argument says readers bring the same interpretive habits to AI text as to human text, so AI text can shape attitudes in the same ways (Does AI text affect readers the same way human text does?). Research on social presence finds that one strong cue, like a voice or a face, does more to make an AI feel like a social partner than many weak cues (Do more social cues always make AI feel more present?). Together these suggest synthetic contact could plausibly work. They also say nothing about whether a charged topic weakens that effect.

The emotion-focused notes add a caution that matters more on hot-button issues. Empathetic AI that soothes negative feelings can strip away the information those feelings carry: what you value, and what you signal to others (What information do we lose when AI soothes emotions?, Does soothing AI empathy actually harm what emotions teach us?). On immigration, anger or fear is often tied to deeply held values. A contact partner that smooths those feelings over might lower stated hostility without the person actually changing their mind. That is a different outcome from real attitude change, and worth checking for in any study claiming success. Relatedly, AI-written text has been shown to shift how a writer comes across toward extremism and confidence (Does AI writing assistance change how readers perceive the writer?). That is a reminder that the AI-played 'outgroup member' may not represent the real group faithfully.

To answer the question properly, the collection would need the synthetic contact paper itself, or its note, with results split by topic. Until then, the useful takeaway is about how to read such a result: if effects differ on immigration, ask whether the AI partner behaved differently on that topic, and whether people actually changed their minds or just calmed down.