Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment

Job posting language directly influences AI hiring bias: The study found that "agentic" language (assertive, competitive terms) in job postings reduced AI recommendation scores for female candidates by a substantial margin (correlation of 0.309, statistically significant), while "communal" language (collaborative, supportive terms) partially reversed this penalty—demonstrating that how companies write job descriptions directly shapes discriminatory outcomes.

Coded exclusionary language creates large-scale discrimination against non-White candidates: Language designed to subtly exclude minorities suppressed recommendation scores for non-White candidates at "large effect sizes" (0.646-0.758 correlation), and notably also deterred non-White job seekers from even applying—creating what researchers call a "chilling effect" that compounds discrimination beyond the hiring algorithm itself.

The bias stems from demographic labels, not just general model limitations: A controlled experiment isolating demographic persona labels revealed these are the primary driver of bias, suggesting the models are directly responding to racial and gender information rather than simply reflecting general knowledge gaps—a finding with important implications for how companies should (or shouldn't) structure candidate data.

Multiple popular open-weight models show consistent bias patterns: Six widely-used models (Llama, Mistral, Gemma, Qwen, Phi, DeepSeek) all exhibited these biases, indicating this is a systemic problem across the open-source AI ecosystem rather than isolated to one vendor.

These findings create regulatory exposure: The discriminatory patterns identified align with violations under the EU AI Act's high-risk classification and U.S. EEOC adverse-impact standards, meaning companies deploying these systems without mitigation face legal liability.