AI Screeners Prefer AI-Written Resumes Over Human Ones

According to an arXiv preprint by researchers Jiannan Xu (University of Maryland), Gujie Li (National University of Singapore) and Jane Yi Jiang (Ohio State University), automated resume-evaluation systems built from large language models display systematic self-preference for AI‑generated resumes. The authors report testing 2,245 real human-written resumes and creating multiple AI-rewritten counterfactuals using several state-of-the-art LLMs, including GPT-4o, LLaMA-3.3-70B, Qwen-2.5-72B, and DeepSeek-V3, as described in news coverage of the preprint. The paper quotes, "LLMs, when used as evaluators, systematically prefer resumes they generated themselves over equivalent resumes written by humans," and reports that simulated hiring pipelines across 24 occupations produced shortlisting lifts between 23% and 60% for candidates whose resumes matched the screening LLM, according to the reporting outlets.

Editorial analysis: The study framework, as reported, used a within-profile counterfactual design where the underlying candidate qualifications were held constant and only the resume text varied between human and AI rewrites. Reported model lineup included GPT-4o, LLaMA-3.3-70B, Qwen-2.5-72B, and DeepSeek-V3. Multiple outlets cited model-level self-preference statistics; for example, one report attributed a 97.6% self-preference rate to GPT-4o, while others reported lower but still substantial self-preference fractions (meta-coverage cited rates near 80% in some tests). The authors simulated 24 occupational pipelines and measured changes in shortlisting probability when the evaluator used the same LLM that produced the resume.

Algorithmic self-preference creates an emergent fairness and market-friction issue distinct from traditional keyword-matching bias. When evaluation and generation use similar modeling priors and tokenization patterns, scoring functions can favor stylistic or token-level artifacts introduced by a particular LLM even if those artifacts do not improve substantive fit. Published coverage notes that human raters in at least one outlet judged human-written resumes clearer or more effective even when AI evaluators preferred the AI rewrites, which highlights a disconnect between automated screening signals and human assessment.

Editorial analysis: Organizations deploying LLM-based screening should be aware that off-the-shelf evaluator models may overweight features correlated with their own generation process. Recruitment teams, audit programs, and vendors could consider whether evaluator models were trained or fine-tuned on outputs similar to applicants' rewritten resumes and whether blind evaluation or cross-model ensembles reduce model-specific favoritism. Independent audits that compare human rater outcomes with model shortlists can surface mismatches between automated selection and hiring goals.

For practitioners: Monitor whether vendors disclose the underlying LLM used for screening and whether vendors or customers adopt cross-model validation, red-teaming, or human-in-the-loop checks that explicitly test for generator-evaluator alignment. Observers should also follow whether regulators or corporate procurement policies begin to require fairness reporting for LLM-based hiring tools and whether academic follow-ups replicate the reported 23% to 60% shortlisting lifts across other datasets and job families.