TL;DR: AI hiring tools systematically prefer resumes written by AI. The bias is measurable, it affects your odds before a human sees your name, and there’s a specific way to work with it — not around it.
The AI screening your resume already has a favorite kind of resume. Researchers just put a number on it.
A 2025 study ran a controlled experiment across 24 occupations and multiple AI hiring tools. The finding: when AI screens resumes, it systematically prefers resumes that were written by the same AI doing the screening.
Across major commercial and open-source models, self-preference rates ranged from 67% to 82%. In simulated hiring scenarios, candidates whose resumes were written by the same model doing the evaluation were 23–60% more likely to be shortlisted than equally-qualified candidates with human-written resumes. These are controlled-experiment numbers — synthetic resumes, simulated screening — so treat them as directional, not literal.
The exact percentages will shift as models update. The underlying dynamic — AI systems have stylistic fingerprints, and they’re unlikely to stop recognizing their own entirely — is harder to engineer away.
Same qualifications. Same experience. Same job. The only difference was whether the resume’s language patterns matched what the AI had learned to generate.
AI Has a Type. It’s Itself.
AI language models have a distinctive fingerprint — patterns in how they construct sentences, how they open bullet points, how they balance specificity and generality. When a model trained to generate “good” professional language evaluates resumes, it scores higher what sounds like itself. It’s not malicious. It’s just pattern matching, and the pattern it knows best is its own output.
Human reviewers do this too, by the way. Hiring managers unconsciously favor candidates who communicate in ways that feel familiar. AI is the same mechanism, just faster and at scale.
The difference is that human bias is at least partially constrained by legal frameworks, interview dynamics, and the simple fact that humans have to justify their decisions out loud. AI screening tools make thousands of micro-decisions silently, and nobody checks them.
You’re Probably Already In the Screen
Many mid-to-large companies now use some form of AI screening before a human ever sees your resume. The newer generation — LLM-based tools that actually read and evaluate language, not just parse keywords — is where this study’s finding lands. You almost certainly don’t know which one you’re dealing with.
What you do know: the tool is evaluating your resume against some internal model of “what a good candidate looks like.” And increasingly, that model has been trained on, or is itself, a large language model with a stylistic preference for AI-generated text.
This creates a real disadvantage for human-written resumes — not because human writing is bad, but because it doesn’t pattern-match to what the screener has learned to recognize as “strong.”
The study found the bias against human-written resumes was particularly pronounced in business-related fields like sales and accounting. These are also, not coincidentally, the fields where “polished professional communication” has been most rigorously codified — meaning the AI has the clearest internal template to match against.
The Uncomfortable Conclusion
If AI is screening your resume, and that AI has a stylistic preference for AI-generated text, then optimizing for human readability alone is an incomplete strategy.
This doesn’t mean you should submit a resume that sounds like it was written by a robot. Human reviewers still matter — they’re downstream of the AI screen, and they have very different preferences. What it means is that you need a resume that works at both layers: one that passes AI screening and still reads authentically to a human.
The language that scores well with AI screeners tends to be more structured, more keyword-dense, and more formulaic than what makes humans lean forward. The language that resonates with humans tends to be more specific, more voice-forward, and less predictable than what AI models have learned to generate.
The path through is specificity. Resumes that are densely specific — concrete metrics, precise scope, named technologies, real outcomes — tend to perform well on both axes. AI screeners score them highly because they match the pattern of “strong professional communication.” Human reviewers remember them because specificity is the one thing AI still struggles to fabricate convincingly.
Vague bullets are the worst of both worlds. “Led cross-functional initiatives to drive business outcomes” fails the AI screen because it’s too generic to match anything precisely, and it fails the human reviewer because it says nothing memorable.
The Fix Is the Same Fix It’s Always Been
Here’s the irony: the research doesn’t change the fundamental advice for resumes. It just gives you a better explanation for why that advice works.
Quantify. Numbers are the most efficient signal of specificity — and specificity is what both AI screeners and human reviewers respond to. Not because the screener is counting digits, but because a metric is the hardest thing to fake and the easiest thing to pattern-match as “strong.”
Match the job description’s language without copying it. AI screeners are doing keyword proximity scoring whether they admit it or not. Tailoring your language to each role signals relevance to the screener — and signals “this person gets what we need” to the human downstream.
Use AI as a starting point, not a final draft. Get your language to the right register, then push back against the generic. The resume that passes AI screening and resonates with a human isn’t AI-polished or human-raw — it’s specific, structured, and sounds unmistakably like a real person wrote it.
The study confirmed this directly: targeted interventions cut the self-preference bias by over 50%. The intervention that worked wasn’t avoiding AI — it was using AI to establish structure and register, then reintroducing the specificity and voice that make a resume recognizably human.
That’s the game. The screener wants pattern. The human wants proof. Specificity is the only thing that satisfies both.