Pro-Female Hiring Bias Persists Across Five LLMs: Removing Candidate Name Kills It

A study of 43,200 API calls across 5 LLMs using Japanese resumes finds consistent pro-female bias that a simple name-anonymization eliminates, but GPT-4o's safety filter rejects 42% of name-removed prompts.

gender biasllm hiringjapanese contextgpt 4oclaude sonnet 4 6deepseek v3

43,200 API calls across five state-of-the-art LLMs reveal a stubborn pro-female bias in hiring decisions that a simple gender-neutrality instruction cannot budge. The study uses 60 Japanese rirekisho-format resumes, 12 name pairs selected on linguistically grounded gender-signal criteria, and five models: Claude Sonnet 4.6, GPT-4o, DeepSeek-V3, Gemini 2.5 Flash, and Llama 3.3 70B.

The Bias Is Real and Resistant to Admonishment

A crossed random-effects linear mixed model confirms a statistically significant pro-female bias across all five models, replicating Western findings in a non-Western context. A prompt-level instruction to be gender-neutral produces no meaningful reduction in that bias. Telling an LLM to be fair does not make it fair.

The Name Is the Leak

A name-reliance analysis formally identifies the candidate name as the primary gender channel. Removing the name from the prompt reduces the female effect by nearly its full magnitude. This is the single most actionable finding: if you want to de-bias an LLM hiring pipeline, strip the name before the model sees the resume.

GPT-4o's Safety Filter Creates a Practical Trap

The study tested a privacy filter that removes the candidate name from the resume text. For GPT-4o, this filter collided with the model's built-in content safety filter, producing a 42% refusal rate. The system refused to process nearly half the anonymized resumes. That is not a theoretical concern; it is a deployment blocker for any team trying to implement name anonymization in production.

What This Means for LLM-Assisted Hiring

Prompt engineering is not enough. Name removal works but triggers model-specific safety guardrails that must be tested per model. The next step is to evaluate whether other personal identifiers (address, educational institution) act as similar bias channels in non-Western resume formats, and whether safety filter incompatibilities can be bypassed without reintroducing bias.

Source: Gender Bias in LLM Hiring Decisions: Evidence from a Japanese Context and Evaluation of Mitigation Strategies
Domain: arxiv.org