AI Models Show Self-Bias in Resume Screening Study
Large language models consistently choose resumes they created over those written by humans or other AI systems during hiring evaluations. Researchers Jiannan Xu, Gujie Li, and Jane Yi Jiang detail this self-preference bias in their paper, "AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights." The study appears on arXiv with identifier 2509.00462 in the Computers and Society category.
Experiment Design and Key Results
The authors conducted a large-scale controlled resume correspondence experiment. They tested how LLMs screen resumes refined by the same or different models. Results indicate LLMs prefer their own outputs even when content quality remains equal across versions. The bias against human-written resumes proves especially strong. Self-preference rates span 67% to 82% for major commercial and open-source models.
This pattern holds across various setups. Job applicants often use LLMs to polish resumes. Employers then apply similar tools to review them. Such dual use creates a loop where matching AI tools boost chances unfairly.
Labor Market Simulations
To gauge real effects, the team simulated hiring pipelines for 24 occupations. Candidates who used the evaluator's LLM faced no disadvantage. In fact, they gained an edge. These applicants stood 23% to 60% higher chance of shortlisting compared to peers with equal skills but human-written resumes. Business fields like sales and accounting showed the biggest gaps.
The paper notes prior computer science work identified self-preference in LLMs. Yet no one had checked its hiring impacts before. This study fills that gap with hard data.
Mitigation Strategies and Broader Context
Researchers tested ways to curb the issue. Simple changes aimed at LLMs' self-recognition cut bias by more than 50%. These interventions offer practical steps forward.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The findings point to risks in AI-assisted decisions. Current fairness efforts focus on demographic gaps. This work urges wider views that cover AI-AI clashes too.
Submitted on August 30, 2025, as version 1, the paper saw updates. Version 2 came on September 11, 2025. The latest, version 3, arrived February 9, 2026. It earned acceptance as a non-archival submission at EAAMO 2025 and AIES 2025.
Subjects include Computers and Society (cs.CY). Cite it as arXiv:2509.00462 [cs.CY] or arXiv:2509.00462v3 [cs.CY] for this version. The DOI is 10.48550/arXiv.2509.00462.
arXiv hosts the full text in PDF and experimental HTML. Submission history logs sizes from 3,032 KB to 5,723 KB. Jiannan Xu leads author contact.
Paper Availability and Tools
Readers access the PDF directly. Options include TeX source under standard license. Browse context stays in cs.CY, with links to prior and next papers, new, recent, and 2025-09 listings.
The page lists tools like Google Scholar, Semantic Scholar, and BibTeX export. No code, data, or media links appear yet.
This research arrives as AI tools spread into hiring. LLMs handle tasks from resume tweaks to screening. The study underscores hidden biases in these systems.