Reproduce “Multimodal datasets: misogyny, pornography, and malignant stereotypes” 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 Uday Prabhu, Emmanuel Kahembwe
An audit of the LAION-400M multimodal dataset reveals widespread presence of misogynistic, pornographic, and stereotypical content, demonstrating the inadequacy of automated filtering and challenging the 'scale beats noise' approach to data curation. The findings illustrate how harmful content is embedded and amplified within large-scale web-scraped datasets used to train AI models.
Key insights
- Automated filtering mechanisms, such as CLIP-based similarity, are insufficient and prone to bias, failing to adequately remove problematic content from large web-scraped datasets.
- The 'scale beats noise' paradigm for dataset creation externalizes significant ethical harm and responsibility, as problematic content is embedded and amplified rather than averaged out.
- Systemic asymmetries exist in AI development, with disproportionate effort allocated to data aggregation over harm reduction, and a lack of accountability for data-driven harms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.