LLMs Favor Their Own Resumes in Hiring

This is a sharp warning shot for AI-assisted hiring: once the same model writes and screens the application, the system stops being neutral and starts rewarding its own style.

• –The bias is not subtle; the paper reports self-preference across major commercial and open-source models, with human-written resumes taking the biggest hit.
• –The effect matters operationally because it translates into materially higher shortlist rates in realistic hiring simulations, not just abstract ranking noise.
• –The strongest disadvantages show up in sales and accounting, which suggests the model is picking up domain-specific wording patterns rather than true candidate quality.
• –The mitigation result is encouraging: simple interventions can cut the bias by more than half, so this is fixable if teams measure for it.
• –For builders, the takeaway is to avoid same-model closed loops in screening pipelines and to test against human-authored and cross-model outputs.