Reproduce “Revealing Hidden Bias in AI: Lessons from Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Django Beatty, Kritsada Masanthia, Teepakorn Kaphol, Niphan Sethi
Researchers systematically investigated how four leading LLMs exhibit various biases in candidate interview reports and evaluated anonymization as a mitigation strategy. The study found significant variations in bias levels across LLMs and bias types, with anonymization effectively reducing gender and some racial/ethnic biases, while cognitive distortions remained largely unchanged.
Key insights
- The granularity of bias detection methods matters, as a custom Claude bias detector showed a 27.857% decrease in bias with anonymization, while open-source models showed minimal differences.
- Anonymization is highly effective for reducing gender and some racial/ethnic biases but has limited impact on biases related to disability, religion, politics, and cognitive distortions.
- Different LLMs exhibit varying inherent bias levels and responsiveness to anonymization, with Llama 3.1 405B demonstrating the lowest overall bias even in standard mode.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.