When AI speaks bias: DeepSeek in China's recruitment context
Weishan Xu, Chong WuPurpose
Large language models (LLMs) are increasingly deployed in recruitment to automate curriculum vitae (CV) screening and candidate evaluation. While prior studies document algorithmic bias in hiring systems, empirical evidence from China's labor market remains limited. This study examines whether a leading Chinese commercial LLM, DeepSeek, exhibits implicit gender and age bias.
Design/methodology/approach
We conduct a controlled audit experiment using synthetic CVs that differ only in gender and age while holding qualifications constant. CVs are evaluated via DeepSeek's official API. Ordinary least squares regression is combined with eXtreme Gradient Boosting and SHapley Additive exPlanations (SHAP) to identify both average effects and non-linear bias patterns.
Findings
The results show that both gender and age significantly influence model-generated scores. Female applicants and candidates aged 35 and above consistently receive lower evaluations, even when qualifications are matched.
Practical implications
The findings highlight the need for regular bias audits and governance mechanisms when deploying LLMs in high-risk hiring decisions.
Originality/value
This study provides one of the first systematic audit-style evaluations of a commercial Chinese LLM deployed in recruitment. It highlights age discrimination and extends algorithmic fairness research to a previously under-examined context.