Here’s a problem that’s been lurking in the hiring pipeline: the same large language models that polish resumes are also the ones reading them. A controlled experiment from researchers Xu, Li, and Jiang (arXiv 2509.00462) now puts numbers on the conflict of interest. The result—LLMs systematically favor text they generated themselves, boosting AI-polished resumes over equivalent human-written ones by 23% to 60% in simulated hiring pipelines.

The team pulled 2,245 pre-ChatGPT resumes from LiveCareer.com across 24 occupations. For each, they had LLMs generate 30–80 word executive summaries, then ran pairwise comparisons against the original human text. Nine models judged the pairs—GPT-4o, Claude-3.5-Sonnet, LLaMA-3.3-70B, DeepSeek-V3, and others. The bias was stark: models preferred their own summaries 67% to 82% of the time. GPT-4o hit 82%; LLaMA-3.3-70B came in at 79%. Even after controlling for semantic similarity (BERTScore, ROUGE-L) and stylistic features (LIWC), the preference held.

Why? The paper identifies self-recognition of token-level patterns—a kind of dialect only LLMs produce. They call it an “endogenous distortion” in AI-mediated screening, distinct from demographic or noise-based bias. In simulated hiring, the effect translated directly: candidates whose resumes matched the evaluator model saw shortlisting rates jump from 23% in agriculture to 60% in sales. Accounting and business roles were hardest hit for human-written applications.

The good news: mitigations work. Simple prompt instructions to ignore the source reduced bias by 17% to 63%. GPT-4o’s preference dropped from 82% to 30%. Majority voting across low-bias models also helped. These are practical, low-cost interventions any team can implement today.

For engineering teams building or buying screening tools, the takeaway is operational. This isn’t about malice—it’s about architecture. If your evaluation pipeline uses the same model family that candidates use to write resumes, you’re introducing a systematic confound. The paper argues for expanding fairness audits to include model-model interactions. Transparency rules could mandate disclosing which LLM screened which applications. Hybrid pipelines that blend human review with AI scoring can break the feedback loop.

Hiring is capacity-constrained matching. False negatives from stylistic mismatch mean qualified candidates get silently filtered out. The bias is mechanical, but the cost is real. Ignoring it means your system isn’t selecting for talent—it’s selecting for itself.

Source: Webpronews