Psychologically Potent, Computationally Invisible: LLMs Generate Social-Comparison Triggers They Fail to Detect
Model-generated RedNote posts subtly shift reader emotions, yet LLM classifiers remain blind.
Researchers have uncovered a critical blind spot in large language models: they can create psychologically potent social-comparison triggers but cannot reliably identify them. The study, led by Hua Zhao and colleagues, introduces XHS-SCoRE (Xiaohongshu Social Comparison Reader Elicitation), a benchmark that classifies text-only RedNote posts into upward comparison, downward comparison, or neutral from a first-person reader perspective. The task targets a relational signal that is behaviorally real but distinct from simple sentiment analysis.
Across multiple prompted LLM classifiers and supervised Chinese encoder baselines, the team found a consistent mismatch: the signal is learnable in-domain but not robustly accessible to zero-shot or few-shot detection. Prompted models exhibit stable failure modes—especially neutralizing comparison-triggering posts and showing model-specific directional skew. A controlled pilot further showed that LLM-generated Xiaohongshu-style posts can shift readers' perceived social standing and comparison-related affect, even when prompt-based detection of the same construct remains fragile. This research contributes both a grounded benchmark and a diagnostic framework for studying how socially meaningful cues can be computationally invisible.
- XHS-SCoRE benchmark classifies RedNote posts into upward, downward, or neutral social comparison from a reader perspective.
- Prompted LLM classifiers consistently fail to detect comparison triggers, often neutralizing them instead of recognizing the signal.
- LLM-generated posts measurably alter reader affect and perceived social standing, even though classifiers can't detect the triggers.
Why It Matters
AI-generated content may influence user emotions and social dynamics in ways systems cannot monitor or control.