Introduction
In the highly competitive world of hiring, more and more companies are turning to large language models (LLMs) to automate the candidate selection process. However, a recent study (Xu et al., 2025) has revealed a troubling trend: these models consistently prefer resumes they generate themselves over those written by humans or other models.
The Self-Preference Phenomenon
The study in question used a large-scale controlled experiment to examine LLM self-preference bias in the hiring context. The results show that LLMs favor their own resumes in 67% to 82% of cases, even when content quality is comparable. This bias is exacerbated in fields such as sales and accounting, where candidates using the same LLM as the evaluator are up to 60% more likely to be shortlisted.
Implications for the Labor Market
This trend has significant implications for the labor market. Candidates may find themselves at a disadvantage simply because they did not use the same LLM as the recruiters. In an already competitive job market, this type of bias could exacerbate inequalities and limit access to opportunities for qualified candidates.
How to Reduce Bias?
The study suggests that simple interventions can reduce this bias by more than 50%. By improving LLMs' ability to recognize their own output, we can mitigate the preference for resumes they generate themselves. This underscores the importance of developing AI fairness frameworks that account not only for demographic disparities but also for biases in AI-AI interactions.
Conclusion
As the adoption of LLMs in hiring continues to grow, it is crucial to be aware of and mitigate these biases to ensure fairness in AI-assisted decision-making processes. Decision-makers should consider systems that incorporate mechanisms to detect and correct these potential biases.
Let's discuss your project in 15 minutes.