Gender Disparities in StackOverflow's Community-Based Question Answering: A Matter of Quantity versus Quality

TL;DR

This study examines gender disparities in Stack Overflow, finding that differences in reputation scores are mainly due to activity levels rather than answer quality, highlighting the need for fairer scoring mechanisms.

Contribution

It uniquely combines human and AI evaluations to assess answer quality and reveals that gender bias does not significantly affect answer quality or selection, but activity levels do.

Findings

No significant gender difference in answer quality.

Gender disparities in reputation are driven by activity levels.

Reputation systems emphasizing activity may worsen gender inequality.

Abstract

Community Question-Answering platforms, such as Stack Overflow (SO), are valuable knowledge exchange and problem-solving resources. These platforms incorporate mechanisms to assess the quality of answers and participants' expertise, ideally free from discriminatory biases. However, prior research has highlighted persistent gender biases, raising concerns about the inclusivity and fairness of these systems. Addressing such biases is crucial for fostering equitable online communities. While previous studies focus on detecting gender bias by comparing male and female user characteristics, they often overlook the interaction between genders, inherent answer quality, and the selection of ``best answers'' by question askers. In this study, we investigate whether answer quality is influenced by gender using a combination of human evaluations and automated assessments powered by Large Language Models. Our findings reveal no significant gender differences in answer quality, nor any substantial influence of gender bias on the selection of ``best answers." Instead, we find that the significant gender disparities in SO's reputation scores are primarily attributable to differences in users' activity levels, e.g., the number of questions and answers they write. Our results have important implications for the design of scoring systems in community question-answering platforms. In particular, reputation systems that heavily emphasize activity volume risk amplifying gender disparities that do not reflect actual differences in answer quality, calling for more equitable design strategies.