AI Hiring Tools May Favor Their Own Work, Smith Study Finds

As artificial intelligence becomes embedded across hiring processes, new research from the University of Maryland’s Robert H. Smith School of Business reveals an unexpected source of bias: AI systems favoring their own outputs.

In a working paper, “AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights,” Smith Ph.D. candidate Jiannan Xu, along with Smith PhD graduates Gujie Li ’25  (National University of Singapore) and Jane Yi Jiang ’24 (The Ohio State University), document how large language models (LLMs) used in hiring may systematically prefer resumes generated by the same model.

The findings, which have drawn coverage from outlets such as The Register, El Confidencial, the New York Post and Business Insider, point to a previously under-examined risk in AI-assisted decision-making—bias emerging not from human demographics, but from interactions between AI models themselves.

A New Form of Bias in the Hiring Pipeline
Today, job applicants increasingly use AI tools to refine resumes and cover letters, while employers deploy similar technologies to screen large pools of candidates. This dual use of AI led the researchers to examine whether these systems favor content that resembles their own output.

Drawing on a large-scale, controlled resume experiment involving more than 2,200 resumes and multiple leading AI models, the study finds consistent evidence of “self-preference bias.” The researchers show that when evaluating candidates, LLMs rate resumes they generated more favorably than those written by humans—or even by competing AI models—despite equivalent quality.

Across major commercial and open-source models, the study reports self-preference rates ranging from roughly 67% to 82% when comparing AI-generated resumes with human-written ones. The researchers interpret this pattern as evidence that model-specific writing styles may influence evaluation outcomes, independent of applicant qualifications.

Real-World Consequences for Job Seekers
To assess operational impact, the researchers simulate hiring pipelines across 24 occupations. Their results indicate that candidates using the same AI system as an employer’s screening tool may gain a measurable advantage—between 23% and 60% higher likelihood of being shortlisted—compared with equally qualified candidates submitting human-written materials.

The study finds that disparities are especially pronounced in business-related roles such as sales and accounting, where standardized language and formatting may amplify similarities between AI-generated resumes and evaluation criteria.

The researchers also identify a secondary effect: LLMs tend to favor their own outputs over those generated by other models, though this “LLM-vs-LLM” bias is smaller and less consistent than the gap between AI-generated and human-written resumes.

Simple Fixes Show Promising Results
Importantly, the study identifies practical interventions that reduce self-preference bias with minimal cost. The researchers highlight two approaches:

• System prompting: Directing AI evaluators to disregard the origin of a resume and focus strictly on skills and qualifications
• Model diversity (majority voting): Using multiple models in screening decisions to offset the influence of any single system

In their experiments, these strategies reduced self-preference bias by more than half, bringing outcomes significantly closer to neutral evaluation.

Implications for Businesses and Policymakers
The findings extend current discussions of AI fairness beyond demographic bias to include interactions between AI models. The researchers argue that such interactions can shape outcomes in subtle but consequential ways, even when candidate qualifications are equivalent.

For organizations, the study underscores the need to audit AI-driven hiring processes and design safeguards that prevent unintended advantages tied to specific tools.

For policymakers, the researchers suggest expanding transparency and accountability frameworks. Potential measures include requiring disclosure of AI use in resume screening and incorporating self-preference metrics into third-party audits of hiring systems.

As AI adoption accelerates, interactions between automated systems are likely to become more common across domains such as hiring, education, and content moderation. The researchers emphasize that understanding how AI models evaluate and respond to their own outputs will be critical for ensuring fair outcomes.

Xu, who is affiliated with UMD’s Institute for Trustworthy AI in Law & Society and the Maryland Language Science Center, says the study highlights “a new form of bias that arises from AI-to-AI interactions.”

“The central contribution is to show that bias can emerge not only from human data or human decisions, but also from the way AI models respond to one another,” Xu says. “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. That makes AI-to-AI bias an important new frontier for fairness research and governance.”

Read the study, "AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights," via SSRN.