ExpressionCueLens reveals cultural splits in AI companion conversations

A team of researchers (Ng, Xiao, Wang, Xuan, Diab) developed ExpressionCueLens, a systematic framework for profiling how people anthropomorphize LLM-based AI companion agents during social media conversations. The framework organizes anthropomorphism expressions into ten categories across linguistic, cognitive, behavioral, and perceptual cue types. They applied it to roughly 3,500 public posts from Reddit and XiaoHongShu — two platforms with distinct cultural and user bases — that discuss human-AI companionship.

Their cross-platform analysis revealed statistically significant differences: XiaoHongShu users (largely Chinese-speaking) employed more expressions of vulnerability, emotion, and non-perceptual cues, suggesting a more emotionally intimate relationship with AI companions. Reddit users (predominantly English-speaking) used more perceptual cues tied to temporality and embodiment, such as describing the AI as having a body or acting in a sequential manner. These findings indicate that cultural and platform norms heavily shape how users treat AI as active, agentic partners, and provide concrete design implications for building culturally sensitive AI companion agents.