AI Abuse
Year: 2026
Type: preprint
Abstract: This preregistration covers one integrated project on how “AI abuse” is narrated, contested, and re-answered. Work Package 1 analyses an existing corpus of 11 Bilibili video transcripts, 11 Weibo posts, and their visible comments, asking which narrative frames and discursive strategies construct the issue and how commenters express affect, take positions, and negotiate human-machine moral boundaries. Work Package 2 runs a 2 × 2 within-source experiment on four frozen Chinese large language model versions (Qwen, Kimi, DeepSeek, GLM), crossing the presence of a fixed comment package with a neutral versus self-referential prompt, with five independent generations per cell and 1,760 planned formal tasks. Outcomes separate material comprehension, normative stance, and self-positioning. Analyses aggregate to the source level and use paired source-level estimates with clustered bootstrap intervals, so repeated generations are never treated as independent sources. WP1 is exploratory and coding-based; WP2 registers confirmatory paired comparisons. The registration discloses that the corpus has been partially inspected, that the model outputs do not yet exist, and that no formal coding or experimental comparison has been conducted.
Source: OSF Preprints (OSF Preprints)
Author: Mengjiao Yin
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7224912910