In 2025, LLMs favored their own generated CVs over human ones 65–80% of the time.
If you're still submitting a handwritten resume, here's the hard truth: you might already be at a disadvantage. A groundbreaking empirical study—covering 2,245 human-written CVs across 24 occupations—reveals that AI hiring tools systematically prefer resumes generated by their own kind. And the bias isn't small. It's overwhelming.
This isn't just a technical quirk. It's a fundamental shift in how hiring works—and if you're not adapting, your application could be getting silently filtered out before a human ever sees it.
Why This Matters for Job Seekers
The rise of AI in hiring was supposed to eliminate human bias—but what if it's just introducing a new kind? One that doesn't care about your experience, skills, or potential, but rather whether your resume sounds like it was written by a machine.
Here's the reality:
- Your resume isn't just competing with other humans—it's competing with AI-optimized versions designed to trigger the right keywords, structure, and tone.
- The system is rigged: if an LLM scores resumes, and that LLM prefers its own outputs, human-written CVs start at a deficit.
- This affects real jobs: the bias is most severe in high-stakes fields like accounting, sales, and finance—where structured, data-heavy resumes align closely with LLM preferences.
If you've ever applied to a job and wondered why you didn't even get a callback—this could be why.
The Study That Exposed the Bias
This isn't theory. It's empirical evidence from a 2025 study (published in 2026) that put AI hiring tools to the test.
The experiment: 2,245 human-written CVs across 24 occupations, run through two key tests—(1) Human vs. LLM: how often do LLMs prefer AI-generated resumes over human ones? and (2) LLM vs. LLM: do LLMs favor resumes written by the same model?
The models tested:
- GPT-4o
- GPT-4-turbo
- DeepSeek-V3
- Qwen-2.5-72B
- LLaMA 3.3-70B
The shocking results:
- Bias vs. human CVs: GPT-4o preferred AI-generated resumes 80% of the time; the others (GPT-4-turbo, DeepSeek-V3, etc.) around 65%.
- Self-preferencing: only DeepSeek-V3 showed significant self-bias (>28%).
- Most biased fields: accounting, sales, finance.
- Least biased fields: agriculture, arts, automotive.
What does this mean? If you're applying for a finance role with a human-written resume, your application is 65–80% less likely to be favored by an LLM screener than an AI-generated one.
Counterarguments & Nuance: Is This Really Unfair?
Before we declare the system broken, let's consider the other side—the AI's defense:
- "LLMs prefer their own outputs because they're better." AI-generated resumes often have consistent formatting, optimized keywords, and structured bullet points—things that also help human recruiters. Maybe the bias is just a byproduct of better resume design.
- "Self-preference isn't malicious—it's a training artifact." LLMs are fine-tuned on their own outputs. If a model sees a resume that looks like its own work, it might score it higher simply because it's familiar.
- "Not all models are equally biased." Only DeepSeek-V3 showed strong self-preference (>28%). Others had minimal self-bias, suggesting this isn't a universal rule—yet.
The rebuttal: even if these explanations are true, they don't change the practical reality for job seekers. You're still competing against AI-optimized resumes. The system still advantages those who use AI tools. And fairness isn't about intent—it's about outcomes.
How to Optimize Your Resume for Both Humans and AI
You don't have to choose between human appeal and AI compatibility. The best resumes do both. Here's how to combine them:
- Keyword optimization + natural language: use industry terms naturally in your bullet points.
- Structured formatting + scannable layout: clear headings, consistent bullet points, no walls of text.
- Standardized sections + storytelling: include Skills, Experience, and Education—but add a short summary at the top.
- ATS-friendly design + visual appeal: simple fonts, no tables or graphics, but use bold and italics for emphasis.
Practical tips:
- Reverse-engineer LLM preferences: use tools like Jobscan or ResumeWorded to see how your resume scores against job descriptions, and mirror the language in the posting—LLMs (and recruiters) look for keyword matches.
- The 60/40 rule: 60% AI-optimized (keywords, structure, bullet points) and 40% human touch (a compelling summary, achievements with context, and personality).
- Test your resume against AI: paste it into ChatGPT or Claude and ask, "Score this resume for a [Job Title] role. What's missing?" If it suggests more keywords or structure changes, consider them.
- Don't over-optimize: avoid stuffing keywords unnaturally, robotic language, or sacrificing readability. Write for humans first, then tweak for AI.
The Big Questions We Need to Answer
This study raises critical questions about the future of hiring—and your place in it.
Technical: How does self-preferencing emerge in transformer architectures? Is it a training-data artifact (LLMs see their own outputs more often), or a structural bias in how attention mechanisms score similarity? Can we debias LLMs to treat human and AI resumes equally?
Ethical: Should AI hiring tools be allowed if they disadvantage humans? Should regulators mandate bias audits for AI hiring tools (as the EU AI Act requires)? Do job seekers have a right to know if their application was screened by AI?
Practical: How do you "beat the AI" as a job seeker? Should you always use AI to write your resume, even if it feels inauthentic? What's the minimum viable AI optimization to stay competitive?
Futuristic: Will hiring become a battle between AI-generated and human resumes? Could we see a two-tier hiring system—one for AI-optimized candidates, one for humans? Or will human-written resumes become a premium signal of authenticity?
Your Turn
The AI hiring revolution is here—and it's not waiting for permission. The question is: are you adapting, or are you getting left behind? Have you noticed your human resume getting fewer callbacks? Are you already using AI to optimize your applications? Should we push back against AI hiring tools—or embrace them?
Study reference: AI Self-preferencing in Algorithmic Hiring (arXiv:2509.00462).