Gender Bias in LLM Hiring Decisions: Evidence from a Japanese Context and Evaluation of Mitigation Strategies

A new study has found a significant pro-female bias in hiring decisions made by five major large language models when evaluating Japanese-format resumes, replicating a pattern previously documented in Western contexts. The research, posted to arXiv on June 17, 2026, used a counterfactual resume design with 60 Japanese rirekisho-format resumes and 12 name pairs selected on linguistically grounded gender-signal criteria [1]. The five models tested were Claude Sonnet 4.6, GPT-4o, DeepSeek-V3, Gemini 2.5 Flash, and Llama 3.3 70B [1]. Across 43,200 API calls, a crossed random-effects linear mixed model confirmed a statistically significant pro-female bias in all five systems [1]. The study also evaluated two practical mitigation strategies. A prompt-level instruction to apply gender-neutrality produced no meaningful reduction in bias [1]. However, a name-reliance analysis formally identified the candidate name as the primary channel through which gender information entered the decision. Removing the name from the prompt reduced the female effect by nearly its full magnitude [1]. An unexpected technical obstacle emerged during the privacy-filter condition. An incompatibility between the name-anonymization filter and GPT-4o's content safety filter produced a 42 percent refusal rate, highlighting a practical deployment challenge for name anonymization in LLM-assisted recruitment pipelines [1]. The findings extend the evidence base on algorithmic bias beyond English-language, Western-format resumes. The ethics of artificial intelligence has long flagged algorithmic bias and fairness as central concerns, particularly where systems influence or automate human decision-making in areas such as hiring [3]. The deployment of AI in recruitment is part of a broader trend in which machine learning is used for decision-making, credit scoring, and e-commerce applications [4]. Claude is developed by Anthropic, a San Francisco-based AI company founded in 2021 by former OpenAI members with a stated focus on AI safety [5]. Gemini is developed by Google DeepMind, the British-American AI research laboratory that became responsible for Google's family of large language models following a 2023 merger with Google Brain [6]. Both organizations, along with others in the field, have stated goals that include the eventual creation of artificial general intelligence, a hypothetical type of AI that would match or surpass human capabilities across virtually all cognitive tasks [7].