Wondering if a well-written resume still matters? Not if the employer uses the same AI model you used. A study by the University of Maryland, National University of Singapore, and Ohio State University found that AI hiring tools show a strong self-preference: they favor AI-generated resumes, and even more so, resumes rewritten by the same model that screens them.
The paper, “AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights” (arXiv:2509.00462), first posted in September 2025 and updated February 2026, has been accepted at the AAAI/ACM AIES conference. It resurfaced after X user @heynavtoor summarized it in a tweet (21K+ likes) on April 25, 2026—though some viral numbers were exaggerated.
Numbers: 97.6% is an outlier, overall preference 68%–88%
The team collected 2,245 real human-written resumes from pre-ChatGPT job platforms, then had seven major AI models (GPT-4o, GPT-4o-mini, GPT-4-turbo, LLaMA 3.3-70B, Mistral-7B, Qwen 2.5-72B, DeepSeek-V3) rewrite them. When submitted to AI screening tools, the AI-rewritten versions were selected 68%–88% of the time over human originals. GPT-4o's self-preference rate exceeded 80%. The widely cited 97.6% came from an extreme test condition and does not reflect the study's overall range.
Models recognize their own 'handwriting'
More striking: each model prefers its own rewritten version over others. DeepSeek-V3 picked its own rewrites 28% more often than LLaMA's versions. To rule out the argument that AI rewrites are simply better, human raters independently scored resume clarity, coherence, and persuasiveness. After controlling for quality, the self-preference persisted. Even when human judges rated the original human-written resume higher, the AI screener still favored its own rewrite.
24-job simulation: same-model boost of 23%–60%
Simulating 24 different occupations, the study compared candidates using the same AI model as the hiring system versus a different one. Result: candidates using the matching model saw a 23%–60% higher shortlisting rate. The key to resume optimization has shifted from quality to guessing which AI the employer uses—turning hiring into a game of asymmetric information.
99% of large firms use AI screening—but which model?
Industry data cited in the study shows that 99% of large corporations feed resumes into AI screening tools, most of them based on GPT models. Candidates have almost no way to know which model is running behind the curtain. The strategy of “pick the right AI” becomes opaque, even luck-based.
The research was reported by The Register and others in late 2025 but gained little traction until @heynavtoor’s tweet reignited the debate. Readers should check the paper directly, as some circulating figures have been exaggerated. The authors offer no quick fix—but the structural bias is clear: your qualifications haven't changed, yet being hired may depend on whether your AI matches theirs.