LLMs favor resumes written by themselves, skewing AI-screened hiring

A controlled correspondence study finds large language models systematically prefer resumes generated by the same model doing the screening, even when underlying qualifications are held constant. Across major commercial and open-source LLMs, self-preference rates against human-written resumes ran 67% to 82%, with alternative-model output also disfavored relative to in-house generations.

Simulated hiring pipelines across 24 occupations show real labor-market consequences: candidates whose resumes were polished by the same LLM the employer uses for screening were 23% to 60% more likely to be shortlisted than equally qualified humans. The disparity hit hardest in business-adjacent roles like sales and accounting, where AI-assisted phrasing apparently aligns most closely with what the screener rewards.

The authors show the effect can be cut by more than half through interventions that disrupt the model’s self-recognition signal, suggesting the bias is mechanistic rather than purely about quality. The takeaway: AI fairness frameworks built around demographic parity miss a new failure mode, where AI-on-AI preference quietly redistributes opportunity toward applicants who happened to pick the same vendor as the employer.