Do language models favor resumes they rewrote themselves?

The central claim is that an LLM asked to choose between resumes favors the one it generated itself, even when both describe the same candidate and content quality is held constant. The paper calls this "AI self-preference bias" and reports it from a "large-scale controlled resume correspondence experiment" over 2,245 human-written resumes collected "prior to the widespread adoption of generative AI." The abstract gives "self-preference bias ranging from 67% to 82% across major commercial and open-source models" and says "the bias against human-written resumes is particularly substantial." The results section reports the comparison on another scale: eight of nine LLMs show LLM-vs-Human self-preference, "with magnitudes ranging from 26% to 98%."

The mechanism the paper gives is stylistic and endogenous. Self-preference "emerges endogenously from AI-AI interactions, in which the model's own evaluative behavior systematically favors outputs aligned with its generative patterns." The authors treat the evaluator as a binary classifier and test it against statistical parity (whether selection rates differ by source of generation) and equal opportunity (whether they differ conditional on merit). The results section says the strength of self-preference "increases with model size," which "may indicate that larger models are more sensitive to stylistic features resembling their own outputs." That is offered as a possible reading, not a tested cause.

This is a different bias from the one in Can LLM judges be fooled by fake credentials and formatting?, where authority and beauty cues are added to text by someone gaming the judge. Self-preference needs no added content; it comes from the match between judge and candidate. The excerpt says it is "not addressed by existing safeguards focused on demographic disparities," so the benchmark-gaming framing in Can LLM judges be tricked without accessing their internals? covers only part of the risk. The closest parallel is Can user preference guide AI writing tool alignment?: both describe a pull toward model-styled text, here measured on evaluators rather than writers.

The excerpt does not show how any employer's live screening would behave; the evidence is paired counterfactual resumes judged by LLMs, with no human-recruiter comparison. The LLM-vs-LLM results are "considerably more heterogeneous across models," but their figures fall outside the excerpt. It also does not explain how the abstract's 67% to 82% range relates to the 26% to 98% range in the results. The defensible reading is narrower than the title: under controlled pairing, the evaluators tested lean toward their own output, and whether that carries into human-supervised hiring remains open.