The Study's Findings
The experimental design is clever: candidate materials of equal qualifications, one group keeping the human original and one polished by an LLM, are then screened by a model. The result: the screening model systematically prefers the AI-polished version, even when the substantive content is identical. The mechanism isn't hard to guess—a model naturally gives high scores to text that's "its kind": fluent, tidy, with a high-frequency-word distribution. The paper calls this self-preference and provides cross-model evidence: every vendor's model has this flaw, and is even more biased toward text generated by itself.
What It Means for Both Sides of Hiring
This finding pushes the absurdity of the hiring arms race into the open: a job seeker who doesn't use AI to polish their résumé takes a hidden hit in the algorithmic first pass; if everyone uses it, the screening layer is only comparing whose prompt is better, and the résumé loses its signal value entirely. The warning for companies is more serious: automated screening has not only the familiar demographic bias but also this brand-new text-source bias, and the compliance-audit checklist gains another line. The pragmatic advice in the short term is not to let a model make elimination decisions alone—lower the weight of the first pass, keep human review, and at least wait until the bias has a fix.
via: Hacker News