# Gender Bias in Multimodal Models: A Transnational Feminist Approach Considering Geographical Region and Culture

**Authors:** Abhishek Mandal, Suzanne Little, Susan Leavy  
**arXiv:** [2309.04997](https://arxiv.org/abs/2309.04997)  
**Date:** 2023-09-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.04997v1/

## Key points

- Authors address gender bias in deep learning based visual-linguistic multimodal models
- Models like CLIP contain gender and social biases
- Proposed methodology for auditing multimodal models with a focus on gender considerations
- Inspiration drawn from transnational feminism concepts including regional and cultural dimensions
- Significant gender bias found in CLIP across different global regions
- Harmful stereotypical associations identified related to visual cultural cues and labels
- Higher levels of bias found in regions from the Global South
- Importance of addressing gender bias in multimodal models for fairness and equity in AI systems
- Research contributes to developing more inclusive AI systems by shedding light on gender bias in CLIP and providing a framework for auditing biases informed by transnational feminist perspectives.

## Summary

In the paper titled "Gender Bias in Multimodal Models: A Transnational Feminist Approach Considering Geographical Region and Culture," authors Abhishek Mandal, Suzanne Little, and Susan Leavy address the issue of gender bias in deep learning based visual-linguistic multimodal models. These models, such as Contrastive Language Image Pre-training (CLIP), have gained popularity and are used in text-to-image generative models like DALL-E and Stable Diffusion. However, it has been discovered that these models contain gender and other social biases that can be perpetuated through AI systems. To tackle this problem, the authors propose a methodology for auditing multimodal models with a focus on gender considerations. They draw inspiration from concepts of transnational feminism which include regional and cultural dimensions. The aim is to identify and mitigate biases present in these models. The study specifically examines CLIP and uncovers significant gender bias with varying patterns across different global regions. Additionally, harmful stereotypical associations related to visual cultural cues and labels such as terrorism were also identified. Interestingly, the levels of gender bias within CLIP align with global indices of societal gender equality; higher levels of bias were found in regions from the Global South. The findings highlight the importance of addressing gender bias in multimodal models to ensure fairness and equity in AI systems. By incorporating concepts from transnational feminism and considering geographical region and culture, this research provides valuable insights into understanding and mitigating biases within these models. Overall, this paper contributes to ongoing efforts towards developing more inclusive AI systems by shedding light on the presence of gender bias in multimodal models like CLIP and providing a framework for auditing these biases informed by transnational feminist perspectives.