New study finds pro-female bias in LLM hiring decisions across five models in Japan

Researchers from multiple institutions investigated gender bias in LLM hiring decisions, extending previous Western-focused studies to a Japanese corporate context. Using 60 counterfactual resumes in the Japanese rirekisho format with 12 linguistically gendered name pairs, they tested five state-of-the-art models: Claude Sonnet 4.6, GPT-4o, DeepSeek-V3, Gemini 2.5 Flash, and Llama 3.3 70B. A crossed random-effects linear mixed model analysis of 43,200 API calls confirmed a statistically significant pro-female bias across all models, replicating earlier Western findings.

Importantly, simply adding a gender-neutrality prompt instruction produced no meaningful bias reduction. The researchers pinpointed the candidate's name as the primary channel for gender signals—removing the name nearly eliminated the bias entirely. However, in practice, name anonymization faces hurdles: GPT-4o's content safety filter conflicted with the privacy filter, leading to a 42% refusal rate for anonymized resumes. This highlights a critical tension between bias mitigation and system reliability in LLM-assisted recruitment pipelines.