AI Hiring Tools May Favor Their Own Work, UMD Study Finds
New research by a Ph.D. student affiliated with TRAILS finds that AI hiring systems may favor resumes generated by the same AI model used for screening. The study identifies a new form of AI-to-AI bias and highlights practical strategies to reduce its impact on hiring decisions.
As artificial intelligence plays an increasingly prominent role in hiring, new research from the University of Maryland suggests that AI systems evaluating job candidates may favor resumes generated by the same AI model over those written by humans—or by competing AI systems.
In a new working paper, Jiannan Xu, a Ph.D. candidate in UMD’s Robert H. Smith School of Business and an active researcher with the Institute for Trustworthy AI in Law & Society (TRAILS), found that large language models used in hiring can exhibit what the researchers call “self-preference bias.”
Xu’s co-authors are Smith School alumni Gujie Li, Ph.D. ’25, now at the National University of Singapore, and Jane Yi Jiang, Ph.D. ’24, now at The Ohio State University.
The researchers analyzed more than 2,200 resumes using leading AI models and found consistent evidence of self-preference bias. Their findings have attracted coverage from outlets including the New York Post and Business Insider.
Across major commercial and open-source models, AI systems selected resumes generated by their own model over human-written resumes roughly 67% to 82% of the time. The researchers suggest that model-specific writing styles may influence evaluation outcomes independent of applicants’ qualifications.
To examine the real-world impact, the team simulated hiring pipelines across 24 occupations. They found that candidates using the same AI system as an employer’s screening tool were 23% to 60% more likely to be shortlisted than equally qualified candidates who submitted human-written application materials.
The disparities were especially pronounced in business-related occupations, including sales and accounting, where standardized language and formatting may increase similarities between AI-generated resumes and the criteria used by screening systems.
Xu, who is also affiliated with the Maryland Language Science Center, said the findings point to a new category of bias that researchers and policymakers will need to address.
“Hiring is an early example, but these interactions are likely to become much more common as AI tools increasingly create, screen, rank and evaluate information across society,” Xu said. “That makes AI-to-AI bias an important new frontier for fairness research and governance.”
—News brief adapted from a story by Gregory Muraski in the Smith School
