Reproduce “Into the LAIONs Den: Investigating Hate in Multimodal Datasets” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Abeba Birhane, Vinay Prabhu, Sang Han, Vishnu Naresh Boddeti, Alexandra Sasha Luccioni
Researchers investigated the impact of scaling on the prevalence of hateful content in large vision-language datasets, specifically LAION-400M and LAION-2B-en. The study found that scaling these datasets leads to a quantifiable increase in hate speech within the alt-text and that current image-based safety filtering is inadequate for mitigating textual harms.
Key insights
- Unconstrained dataset scaling results in a measurable and statistically significant increase in hateful, targeted, and aggressive content in the alt-text.
- Image-based NSFW filtering is largely ineffective for identifying and mitigating textual harms, demonstrating only a slight correlation with alt-text toxicity.
- Rigorous, continuous, and truly multimodal auditing is essential for large-scale datasets, moving beyond simplistic binary safety labels to provide granular toxicity information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.