# User Acceptance of Gender Stereotypes in Automated Career Recommendations

**Authors:** Clarice Wang, Kathryn Wang, Andrew Bian, Rashidul Islam, Kamrun Naher Keya, James Foulds, Shimei Pan  
**arXiv:** [2106.07112](https://arxiv.org/abs/2106.07112)  
**Date:** 2021-06-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.07112v2/

## Key points

- Fair AI and ML research is gaining momentum to mitigate discriminatory bias in algorithms along lines of gender, age, and race.
- Developing a fair AI algorithm may not be enough to achieve intended results in the real world.
- A recent study built a fair AI career recommender using gender debiasing machine learning techniques.
- Offline evaluation showed that the debiased recommender made fairer career recommendations without sacrificing accuracy.
- An online user study involving over 200 college students revealed that participants preferred the original biased system over the debiased system due to perceived gender disparity being a determining factor for recommendation acceptance.
- Accurately capturing users' interests and preferences is critical to building an effective recommender system.
- The study highlights the need for further research to understand and address underlying factors contributing to gender disparity in career recommendations.

## Summary

The field of fair Artificial Intelligence (AI) and Machine Learning (ML) research is gaining momentum as researchers aim to mitigate discriminatory bias in AI algorithms, particularly along lines of gender, age, and race. However, a recent study has shown that developing a fair AI algorithm may not be enough to achieve its intended results in the real world. The study focused on career recommendation as a case study and built a fair AI career recommender using gender debiasing machine learning techniques. The offline evaluation showed that the debiased recommender made fairer career recommendations without sacrificing accuracy. However, an online user study involving over 200 college students revealed that participants preferred the original biased system over the debiased system. Perceived gender disparity was found to be a determining factor for recommendation acceptance. To accurately capture users' interests and preferences, researchers used dimension reduction techniques to group 140K items into categories and selected representative items from each category. The questionnaire used in the study included five sections: demographics, user interests, beliefs about gender-bias in career choice, recommendation acceptance, and general system usability. Participants were also asked two questions to help understand their beliefs about gender roles in career selection. Overall, responses from 202 participants were received through an online survey with minimal personal information collected such as gender identification and academic standing. While 48% of participants identified as female and 49.5% identified as male there were also non-binary individuals who participated in the study. Accurately capturing users' interests and preferences is critical to building an effective recommender system. In this study researchers employed various techniques such as Latent Dirichlet Allocation (LDA) analysis to identify latent topics related to user interests and preferences. In conclusion while developing fair AI algorithms is crucial for mitigating discriminatory bias addressing gender bias in humans is equally important. The study highlights the need for further research to understand and address the underlying factors that contribute to gender disparity in career recommendations.