Chinese social media turns LLMs into fortune tellers for love and gacha
23,000+ divination posts analyzed; users act as prompt engineers for AI readings.
A systematic study published on arXiv (2606.12418) examines the viral phenomenon of using large language models for divination on Chinese social media, specifically on Xiaohongshu. Researchers analyzed 23,000+ posts and conducted 32 semi-structured interviews with both casual users and professional diviners. The practice, termed LLM-mediated Xuanxue, covers pragmatic concerns: romantic relationships, career decisions, exam outcomes, and even in-game gacha pulls. Users typically engage through two pathways: trend-driven curiosity (viral exposure and zero cost) and event-driven anxiety (uncertainty about a specific outcome). A defining feature is collaborative prompt refinement, where users actively engineer prompts to get more satisfying readings, effectively turning each consultation into a co-creative process.
Among commenters who expressed a clear stance, perceived efficacy skewed positive. "Accuracy" was often justified through biographical fit and retrospective confirmation, consistent with the Barnum effect and confirmation bias. Users developed verification strategies such as repeated trials with the same prompt and cross-model comparisons (e.g., trying ChatGPT vs. DeepSeek). Professional diviners, by contrast, pushed back, arguing LLMs lack "spiritual power" — a stance the researchers interpret as both ontological commitment and economic boundary-work to protect their craft. The study situates the practice within anthropological theories of divination, noting that LLMs preserve core functions while introducing scalability, repeatability, and prompt-driven co-production that reshape how authority is constructed and evaluated.
- Researchers analyzed 23,000+ Xiaohongshu posts and interviewed 32 users and professional diviners about LLM-based divination.
- Top query topics include romantic relationships, careers, exams, and in-game gacha draws — all pragmatic concerns under uncertainty.
- Users actively refine prompts and verify readings through repeated trials and cross-model comparisons, acting as prompt engineers.
- Perceived accuracy is often justified via Barnum effect and confirmation bias; professional diviners reject LLMs as lacking spiritual power.
Why It Matters
LLMs are reshaping ancient divination practices at scale, blending AI with cultural superstition and challenging traditional authority.