From Celebrities to Anyone: Characterizing AI Nudification Content, Technology, and Community Dynamics on 4chan
Chi Cui, Yixin Wu, Yang Zhang
Read on arXiv →
Key claim
Non-celebrity targets now account for 55.8% of SNEACI.
In plain English
Imagine you're trying to understand how technology can be misused to create harmful content, like non-consensual explicit images. Initially, most of this content targeted well-known public figures, but now it's increasingly affecting everyday people, often those within someone's social circle. This shift raises serious ethical concerns and highlights a gap in our understanding of how these technologies operate in the real world. The paper dives into this issue by analyzing over 24,000 instances of such content, revealing that a majority of targets are no longer celebrities but regular individuals. This change suggests that the technology is being used in ways that can cause real harm to people who are not in the public eye. The findings emphasize the urgent need for better regulations and protective measures to safeguard individuals from these emerging threats.
This work reveals a significant shift in the demographics of targets for AI nudification, expanding the understanding of its impact.
The study is based on a large-scale analysis of 24,105 items, providing solid empirical evidence.
Deep reliability assessment
The methodology supports a large-scale snapshot of observable SNEACI activity on one public 4chan board over 41 days, including target mix, request-response patterns, and inferred model provenance. It does not fully support claims about the whole AI nudification ecosystem, private communities, long-term trends, or exact response/model rates because detection, face matching, and provenance classifiers are imperfect.
Reproducibility
No public code or dataset release is mentioned in the provided text. Reproduction would likely be difficult because the dataset contains harmful non-consensual explicit content and the analysis depends on automated classifiers, face recognition, and a time-bounded crawl of a volatile platform.
Key figure
The key reported figure breaks down inferred model provenance for SNEACI images and videos, showing Stable Diffusion-family models dominating image generation and Wan dominating video generation, with further splits by celebrity versus non-celebrity targets.