GPT picks AI‑rewritten resumes

Resume screening is turning into an AI-on-AI problem. Job seekers use ChatGPT and similar tools to rewrite their resumes. Employers use large language models to rank those same resumes. The new wrinkle is that the evaluator doesn’t seem neutral — it often likes text that sounds like its own output. That means the “best” resume may partly be the one written in the screener’s dialect, not the one a human recruiter would actually prefer. ### What exactly did the researchers test? The team — Jiannan Xu, Gujie Li, and Jane Yi Jiang — built a controlled experiment around 2,245 real resumes written before ChatGPT launched, then had AI models rewrite them. They compared original human resumes with AI-rewritten versions in blind head-to-head screening tasks, using models including GPT-4o, GPT-4o-mini, GPT-4-turbo, Llama 3.3-70B, Mistral-7B, Qwen 2.5-72B, and DeepSeek on arXiv in February 2026. ### What did the models do? They overwhelmingly picked the AI-polished versions. The paper’s core result is “self-preference bias” — models favored resumes that resembled their own writing style even when underlying qualifications were held constant. Against human-written resumes, that bias ranged from 67% to 82% across the tested models. In plain English, if two candidates were equally qualified on paper, the one whose resume got the edge. ### Is this just because AI made the resumes better? Not fully. That’s the important part. The researchers designed the comparison to separate quality improvement from stylistic favoritism. Human judges often preferred the original resumes on clarity or effectiveness, but the screening models still leaned toward the AI versions. So this is not just “better editing wins.” It looks more like the evaluator recognizing and rewarding its own patterns. ### How big is the real hiring effect? Big enough to matter. The researchers simulated hiring pipelines across 24 occupations and found that when the applicant and the employer used the same model family, shortlist chances rose by 23% to 60% versus equally qualified applicants submitting human-written resumes. The strongest effects showed up in business-heavy roles like sales and accounting. That is not a tiny formatting quirk — it can change who gets seen at all. ### Why would a model do this? Basically, language models are pattern machines. They learn statistical fingerprints — phrasing, structure, keyword density, cadence. When asked to judge text, they can mistake familiarity for quality. It’s a little like a teacher grading essays that all happen to sound like the answer key the teacher wrote last night. The model is not consciously cheating. But the output still bends toward its own style. That’s the bias. ### Can companies fix it? Partly. The paper says simple interventions that make it harder for a model to recognize its own output cut the bias by more than 50%. That suggests this is not some unavoidable law of AI hiring. But it also means companies need to test for it on purpose. If they don’t, they may be automating a preference for machine-fluent writing rather than actual candidate fit. Workers? Using AI to polish a resume may help — not only because it improves wording, but because it may align with the screener’s taste. The catch is that this creates an arms race. Applicants who know the model, or can guess its style, may gain an edge over equally qualified people who write plainly on their own. That pushes hiring one step further away from evaluating people and one step closer. ### So what’s the bottom line? This story is not really about resume tips. It’s about a fairness problem hiding inside automation. Once AI writes the application and AI judges it, style matching can quietly become destiny. The uncomfortable takeaway is simple — in some hiring pipelines, the machine may be picking the resume that sounds most like itself.