# On Hate Scaling Laws For Data-Swamps

**Authors:** Abeba Birhane, Vinay Prabhu, Sang Han, Vishnu Naresh Boddeti  
**arXiv:** [2306.13141](https://arxiv.org/abs/2306.13141)  
**Date:** 2023-06-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.13141v2/

## Key points

- Prevailing sentiment in generative AI is to scale model, data, and GPU-farms
- Gap in understanding the impact of scaling datasets, especially visio-linguistic datasets from CommonCrawl dump
- LAION-400M project aimed to create open-source variants of opaque datasets and models like CLIP and DALL.E
- Open-sourcing initiatives are vital for accountability and trustworthy AI by enabling auditing and evaluation
- Concerns with web-sourced data necessitate continual evaluation and audit for ethical practices
- Emphasizing scalability over societal equity often aligns with Big Tech's profit-driven objectives at the expense of individual rights and societal impacts
- Auditing multimodal datasets like LAION-400M presents challenges due to hateful content existing in text and image modalities
- Poor data quality issues such as duplicates can significantly impact image generation processes at scale
- Presence of hateful content increased by nearly 12% in larger datasets, exacerbating societal biases and negative stereotypes
- Ongoing evaluation and audit of large-scale datasets are crucial for ensuring ethical AI practices

## Summary

In the world of generative AI, the prevailing sentiment is to scale not only the model and data but also the GPU-farms. While model scaling has been extensively studied, there remains a gap in understanding the impact of scaling datasets, especially in visio-linguistic datasets sourced from the CommonCrawl dump on the World Wide Web. This large-scale data dump serves as a crucial resource for training large generative models despite its drawbacks. The LAION-400M project aimed to create open-source variants of opaque datasets like WIT and models like CLIP and DALL.E. Open-sourcing initiatives are vital for accountability and building trustworthy AI by enabling auditing and evaluation. However, concerns arise with web-sourced data, necessitating continual evaluation and audit of large-scale datasets and models for ethical practices. , emphasizing scalability over societal equity or systemic change, often aligns with Big Tech's profit-driven objectives at the expense of individual rights and societal impacts. Recognizing the negative downstream effects of scale underscores the importance of evaluating models and datasets as an intervention mechanism. Auditing multimodal datasets like LAION-400M and LAION-2B-en presents challenges due to hateful content existing in both text and image modalities. Poor data quality issues such as duplicates can arise with scale, significantly impacting image generation processes. The audits focused on comparing LAION-400M containing 400 million samples with LAION-2B-en containing 2 billion samples to evaluate their impact on hateful content and racial bias in visio-linguistic models. The presence of hateful content increased by nearly 12% in larger datasets, exacerbating societal biases and negative stereotypes. As scale increased, model tendencies towards associating images with offensive classes decreased significantly. Overall, this study highlights the need for ongoing evaluation and audit of large-scale datasets to ensure ethical AI practices. The open-sourcing movement plays a crucial role in providing access to datasets for independent auditing and evaluation to address concerns surrounding dataset curation practices effectively.