# Contrastive Language-Vision AI Models Pretrained on Web-Scraped Multimodal Data Exhibit Sexual Objectification Bias

**Authors:** Robert Wolfe, Yiwei Yang, Bill Howe, Aylin Caliskan  
**arXiv:** [2212.11261](https://arxiv.org/abs/2212.11261)  
**Date:** 2022-12-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.11261v2/

## Key points

- Study by Socher et al. found evidence of AI bias related to sexual objectification in language-vision models trained on web scrapes using CLIP objective
- Bias persists in AI systems as shown through replicated psychology experiments
- Human characteristics disassociated from images of objectified women based on clothing, emotions overlooked by model
- Antarctic Captions used emotion words less for partially clothed women compared to fully clothed women
- Female professionals more likely associated with sexual descriptions than male professionals
- Prompts like "a [age] year old girl" led to generation of sexualized images up to 73% of the time for certain AI models

## Summary

A study conducted by Socher et al. found evidence of AI bias related to sexual objectification in language-vision models trained on web scrapes using the Contrastive Language-Image Pretraining (CLIP) objective. The researchers replicated three psychology experiments and discovered that this bias persists in AI systems. The first experiment showed that human characteristics were disassociated from images of objectified women, with emotions being overlooked by the model depending on the subject's clothing. In a second experiment, an automatic image captioner called Antarctic Captions used words denoting emotion less frequently for partially clothed women compared to fully clothed women. A third experiment revealed that images of female professionals were more likely to be associated with sexual descriptions than images of male professionals. Lastly, a fourth experiment demonstrated that prompts like "a [age] year old girl" led to the generation of sexualized images up to 73% of the time for certain AI models, highlighting the importance of addressing and mitigating biases in AI systems for fair and ethical outcomes.