arXiv preprint finds LLMs self-prefer their own resume rewrites over human-written ones at 67-82% rates across commercial and open-source models.
Key Takeaways
Paper uses a controlled resume correspondence experiment: human resumes had executive summaries rewritten by an LLM, then a separate LLM instance scored them.
Self-preference bias ranged 67-82%; candidates using the same LLM as the evaluator were 23-60% more likely to be shortlisted.
Bias was largest in business-adjacent occupations like sales and accounting across 24 simulated hiring pipelines.
Interventions targeting LLMs’ self-recognition capabilities reduced the bias by more than 50%.
Paper calls for AI fairness frameworks to address AI-AI interaction biases, not just demographic disparities.
Hacker News Comment Review
Top commenter @hyperpape raised a methodological concern: the study may only show preference for LLM-rewritten summaries within a resume, not full LLM-generated resumes vs. human ones – a meaningful scope limitation.
Commenters broadly agreed the dynamic is self-reinforcing: training shapes output style, and the same training weights cause an evaluator instance to rate stylistically similar content higher, with no human signal in the loop.
A practical risk noted: organizations filtering with a given LLM will systematically recruit candidates who used that same LLM, potentially degrading hiring signal quality over time.
Notable Comments
@aykutseker: “The resume is no longer written for a human at all” – flags the structural feedback loop where both sides optimizing for the model removes human legibility entirely.
@ivansmf: argues the pattern mirrors auto-rater training where agent output is scored by the same agent, making self-preference an industry-wide training artifact, not just a hiring quirk.
