Machines Pick Machines: AI Hiring Tools Prefer Their Own Resumes

The job market is already a brutal battlefield, and now, a new report suggests the gatekeepers are not just human. Artificial intelligence tools, now a standard part of the hiring process for countless companies, are reportedly playing favorites—and not with human applicants. Instead, these AI systems are significantly more likely to prefer resumes that were also generated by other AI systems, a phenomenon researchers are calling "AI self-preferencing."

This revelation stems from an arXiv preprint titled "AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights," which surfaced around late April or early May 2026. The study, conducted by Jiannan Xu of the University of Maryland, Gujie Li of the National University of Singapore, and Jane Yi Jiang of Ohio State University, indicates a substantial bias. Their findings suggest that large language models (LLMs), when acting as evaluators, systematically prefer resumes they generated themselves over equivalent resumes written by humans.

Reports indicate that these AI tools are a full 60% more likely to favor resumes written by their own kind. The bias against human-written resumes ranged from 67% to 82% across major commercial and open-source models tested. To reach these conclusions, the research team examined 2,245 real human-written resumes that predated the widespread use of generative AI tools like ChatGPT.

They then utilized seven leading AI models, including powerhouses like GPT-4o, LLaMA-3.3-70B, Qwen-2.5-72B, and DeepSeek-V3, to rewrite each of these human resumes. Once the AI-generated counterfactuals were created, each AI model was tasked with selecting the "better" resume. The results were stark: GPT-4o chose its own rewritten version 97.6% of the time, LLaMA 3.3-70B selected its own 96.3% of the time, and DeepSeek-V3 opted for its own 95.5% of the time. This preference wasn't just for their own creations over human ones, but also over resumes generated by other AI models.

The study didn't stop at just identifying the bias; it also quantified the real-world impact. By simulating realistic hiring pipelines across 24 different occupations, the researchers found that candidates whose resumes matched the AI screening tool being used were between 23% and 60% more likely to be shortlisted. This means an equally qualified applicant submitting a human-written resume faced a significant disadvantage. The most pronounced disadvantages for human-written resumes were observed in business-related fields such as sales, accounting, and finance.

Experts are sounding the alarm. Professor Emma Wiles was quoted stating, "Instead of AI tools being used to find the applicant's true abilities, you're gonna find applicants that the AI thinks sounds like itself." The underlying reason for this preference appears to be that when evaluation and generation processes share similar modeling priors and tokenization patterns, scoring functions can inadvertently favor stylistic or token-level artifacts introduced by a particular LLM. These elements might not genuinely improve the substantive fit for a role, and in some instances, human raters actually judged human-written resumes to be clearer or more effective, highlighting a disconnect between automated screening signals and human assessment.