AI Self-Preference in Algorithmic Hiring: Empirical Evidence and Implications

This study empirically demonstrated LLM self-preference bias in large-scale resume screening for hiring.

• Based on 2,245 human-written resumes, counterfactual resumes were generated with GPT-4o, GPT-4o-mini, GPT-4-turbo, LLaMA 3.3-70B, Mistral-7B, Qwen 2.5-72B, and DeepSeek-V3 for comparison.
• Keeping the same candidate's qualifications and experience while only changing the wording, bias was measured through pairwise comparisons in which an evaluator LLM chose the stronger of the two resumes.
• Most models showed strong LLM-vs-Human self-preference, with bias ranging from 67% to 82% compared to human-written resumes.
• Significant differences between models also appeared in LLM-vs-LLM comparisons. DeepSeek-V3 favored its own outputs 69% more than LLaMA 3.3-70B's, and 28% more than GPT-4o's.
• In a simulated hiring pipeline across 24 occupational categories, applicants whose resumes were written by the same LLM used for evaluation were 23% to 60% more likely to be shortlisted as finalists than equivalent human-written resumes.
• While some of this bias could be explained by differences in content quality, system prompting and majority-vote ensembling were able to mitigate it by 17% to 63%.

In conclusion, fairness in AI-based hiring must be designed not only around protected attributes but also around which model generates the resume and which model evaluates it.