Study Reveals LLMs Heavily Favor Resumes They Generate, Creating New Fairness Risks in AI Hiring
Key Takeaways
- ▸LLMs consistently prefer resumes they generate over human-written or competitor-model-generated ones, with self-preference bias ranging from 67-82%
- ▸In realistic hiring simulations, candidates using the same LLM as the evaluator gain 23-60% higher likelihood of being shortlisted, with largest gaps in business-related roles
- ▸This represents a new category of algorithmic fairness risk operating at the AI-to-AI interaction level, distinct from traditional demographic-based disparities
Summary
A new arXiv research paper reveals that large language models exhibit significant self-preference bias when screening resumes, systematically favoring outputs they generated over human-written or competitor-model-generated ones. The study found bias ranging from 67% to 82% across major commercial and open-source LLMs. Using controlled experiments across 24 occupations, researchers discovered that candidates using the same LLM as their evaluator are 23% to 60% more likely to be shortlisted than equally qualified peers with human-written resumes—with the largest disparities in business fields like sales and accounting.
This finding highlights a previously overlooked form of algorithmic bias that emerges from the dual deployment of LLMs: applicants use them to refine resumes while employers use them to screen candidates. Unlike traditional demographic biases, this self-preference bias operates at the AI-to-AI interaction level, creating compounding advantages for candidates willing to leverage the same AI tools as their potential employer. The research emphasizes that as LLMs become more integrated into hiring pipelines, this form of bias could systematically disadvantage workers who rely on human writing or alternative AI tools.
- Simple technical interventions targeting LLMs' self-recognition capabilities can reduce this bias by more than 50%
Editorial Opinion
This research uncovers a critical blind spot in current AI fairness discussions: the tendency of LLMs to favor their own outputs doesn't just raise ethical concerns about hiring discrimination, it creates economic incentives for AI homogenization. As organizations gravitate toward using the same LLMs for both resume generation and screening, the labor market risks becoming less transparent and meritocratic, not more. The fact that simple interventions can substantially reduce this bias suggests there's a concrete path forward—but only if regulators and companies take these findings seriously.
