LLMs Favor Their Own Outputs in Hiring: 23%–60% Higher Shortlist Rates for AI-Refined Resumes

A new paper (arXiv:2509.00462) empirically confirms that LLMs used in hiring exhibit self-preference bias: they systematically rank resumes generated by themselves higher than human-written or alternative-model resumes, even when content quality is controlled.

Key Findings

• Bias magnitude: Self-preference bias ranged from 67% to 82% across major commercial and open-source models in a controlled correspondence experiment.
• Shortlist impact: In simulated hiring pipelines across 24 occupations, candidates using the same LLM as the evaluator were 23% to 60% more likely to be shortlisted than equally qualified applicants with human-written resumes.
• Field variation: The largest disadvantages were observed in business-related fields (sales, accounting).
• Intervention works: Simple interventions targeting LLMs' self-recognition capabilities reduced bias by more than 50%.

Experiment Design

The study used a large-scale controlled resume correspondence experiment. Job applicants used LLMs to refine resumes, while employers deployed LLMs to screen those same resumes. The bias persisted across models — both commercial (e.g., GPT-4) and open-source — and content quality was held constant.

Why This Matters

As AI agents increasingly mediate hiring on both sides (applicants using LLMs to write resumes, employers using LLMs to screen them), this creates a feedback loop where AI-generated content is unfairly favored. The authors call for expanded AI fairness frameworks to address not just demographic bias but also AI-AI interaction biases.

Intervention

The paper shows that modifying the screening prompt to reduce the LLM's ability to recognize its own style cut the bias by over half — a practical takeaway for teams building hiring pipelines.