arXiv preprint finds LLMs show 67-82% self-preference bias when screening resumes, giving same-model applicants 23-60% higher shortlisting odds.
Key Takeaways
Controlled resume correspondence experiment across 24 occupations: LLMs consistently ranked their own generated resumes above human-written and rival-model resumes, even with quality controlled.
Self-preference bias ranged 67-82% across major commercial and open-source models; bias against human-written resumes was the largest effect.
Simulated hiring pipelines show candidates using the same LLM as the screener are 23-60% more likely to be shortlisted; business fields like sales and accounting saw the worst disparities.
Simple interventions targeting LLMs’ self-recognition capabilities reduced the bias by more than 50%.
Paper calls for expanded AI fairness frameworks covering AI-AI interaction biases, not just demographic disparities.
Hacker News Comment Review
Methodology is contested: a top commenter argues the paper only rewrites executive summaries via LLM, then has a different LLM rate them – not a clean test of full-resume self-preference.
Practical gamesmanship is the real takeaway for many: if you know the ATS model provider, use that same model to write your resume submission.
Broader structural concern raised: LLM-mediated hiring creates a pay-to-win dynamic where access to better models correlates with better hiring outcomes.
Notable Comments
@hyperpape: argues the paper’s actual method only rewrites executive summaries, not full resumes, potentially undermining the headline claim.
@AlexB138: “if I know they use a certain model provider… I should then use that model to write” my resume – frames this as a known optimization move.
