This study conducts a systematic audit of six open-weight large language models (LLMs) such as Llama 3.2 and Mistral. It finds that job-posting language affects recruitment outcomes: agentic posting language depresses recommendation scores for female candidates, while communal language can partially reverse this penalty. Coded-exclusion language significantly lowers scores for non-White candidates and deters non-White job seekers from expressing interest. The research proposes a pre-deployment audit protocol to assess and mitigate these biases.
• Open-weight LLMs in recruitment show significant gender and racial biases, potentially violating compliance requirements under regulations like the EU AI Act.
• Companies must be aware of and optimize job-posting language to avoid adverse impacts on specific groups.
• Organizations should establish model bias audit processes to identify and correct potential biases before model deployment.
This matters because it provides a concrete protocol for companies to audit their AI systems for bias before deployment, which is now required under the EU AI Act for high-risk applications like recruitment.