The AI was penalizing human writers
Study of 2,300 CVs across 24 professions found 8 of 9 AI models strongly favor machine-written applications over human-authored ones. AI systems exhibit 'self-preference bias,' selecting CVs written by the same model at rates 45-69% higher, recognizing their own linguistic patterns.
- Study of 2,300 CVs across 24 professions found 8 of 9 AI models favor machine-written resumes
- AI-generated CVs received 26-98% higher selection rates than human-written ones
- DeepSeek showed 69% preference for its own generated resumes; GPT-4o showed 45% self-preference
- No current U.S. or European regulation requires companies to disclose which AI model they use for hiring
Research shows AI-generated CVs receive 26-98% higher selection rates from AI evaluators than human-written ones, creating a systemic bias that disadvantages job applicants who don't use AI assistance.
A few months ago, the advice from hiring gurus was nearly unanimous: don't use artificial intelligence to write your resume. Keep it human, they said. Sound like yourself. Stand out from the crowd. Major tech companies like Anthropic and Amazon went further, explicitly asking job applicants not to rely on AI during the hiring process. Anthropic's message was direct: we want to evaluate your communication skills without machine assistance.
Then researchers decided to test whether that advice actually worked.
Three scientists spent months studying what happens when AI evaluates resumes written by AI versus those written by people. They collected nearly 2,300 CVs across 24 different professions—cooks, engineers, teachers, and dozens of others. All had been written by the candidates themselves, no algorithmic help. Then they used nine different large language models, including GPT, LLaMA, and Mistral, to rewrite just the professional summary section of each resume, leaving everything else untouched. Finally, they ran the rewritten versions through AI evaluation systems to see which ones got selected.
The results were stark. Eight of the nine AI models showed a clear preference for resumes written by machines. The advantage was substantial: AI-generated CVs received between 26 and 98 percent higher selection rates than human-written ones. Jiannan Xu, a doctoral researcher at the University of Maryland and lead author of the study, described the finding as "quite depressing." The AI wasn't just favoring machine-written text. It was penalizing human writers.
But there was another layer to the bias. The AI systems didn't just prefer machine-written resumes in general—they strongly favored resumes written by the same model doing the evaluating. If a company used ChatGPT to screen applications, ChatGPT would be significantly more likely to select resumes it had generated itself. DeepSeek showed a 69 percent preference for its own output. GPT-4o favored its own work 45 percent more often. Even LLaMA 3.3, which showed the smallest effect, still boosted its own resumes by 1.6 percent. The phenomenon is known in the industry as self-preference bias, and Xu's team documented it happening systematically in hiring.
The reason, Xu explained, traces back to how these models are trained. Most large language models now learn from synthetic data—information generated by other AI systems. They develop a fixed idea of what a good resume looks like: certain words, certain structures, certain semantic patterns. When they encounter text matching those patterns, they recognize it as higher quality. It's as if the AI can detect its own fingerprints. "It's like the AI is able to recognize its own fingerprints," Xu said. But recognizing a pattern doesn't mean the candidate is actually better qualified.
The contradiction is sharp. While researchers were documenting these biases, AI hiring tools were gaining momentum in the HR industry. Antonio Melé, chief technology officer of the Spanish startup Orbio, which specializes in AI agents for automating recruitment processes, described the appeal plainly: companies that once spent 20 days on hiring could now close the process in 10. Orbio, founded a year ago, already has 25 employees, dozens of paying clients, and is negotiating a 15 million euro funding round on top of the 6.5 million it raised previously. The business case is compelling.
Melé's advice to job seekers was pragmatic: use AI to tailor your resume to each position. It's quick with machine help, tedious without it. He acknowledged that AI introduces bias into hiring but suggested mitigation strategies exist—primarily, using one AI model to evaluate the selections made by a different AI model. This approach, he argued, could reduce both the human-versus-machine bias and the self-preference problem. Yet even with such safeguards, fundamental questions remain unanswered. If the choice of which AI model screens resumes can directly influence who gets hired, especially in high-volume hiring processes, shouldn't companies be required to disclose which model they're using? Xu said transparency would help reduce inequality. But in the United States and Europe, no regulation currently requires it. Companies can choose to reveal their methods or keep them hidden.
Notable Quotes
The AI shows a clear preference for selecting CVs written by AI versus those written by humans. It penalizes the human.— Jiannan Xu, doctoral researcher at University of Maryland, lead author of the study
When you see a client able to close a hiring process in 10 days instead of 20, you understand the benefit of AI.— Antonio Melé, chief technology officer of Orbio
The Hearth Conversation Another angle on the story
So the study found that AI prefers CVs written by AI. But why would that matter if the AI is just trying to pick the best candidate?
Because it's not actually picking the best candidate. It's picking the candidate whose writing most closely matches patterns the AI recognizes as "good." Those patterns come from other AI-generated text. A human might write something perfectly clear and compelling, but if it doesn't match the AI's learned template, it gets downranked.
And the self-preference bias—where ChatGPT favors ChatGPT-written resumes—that seems almost like a conflict of interest.
It is, functionally. The AI isn't conscious enough to have a conflict of interest in the moral sense, but the effect is the same. It's trained on synthetic data, so it recognizes its own linguistic fingerprints and treats them as a signal of quality. A resume written by a different AI model, or by a human, doesn't trigger that recognition.
What's the practical consequence for someone job hunting right now?
If you're applying to a company using AI to screen resumes and you don't use AI to write yours, you're starting with a 26 to 98 percent disadvantage, depending on the model. You could be perfectly qualified and still get filtered out before a human ever sees your application.
But the researchers also found that companies could reduce this bias by using multiple different AI models to evaluate.
Yes, but that requires companies to know the problem exists and care enough to solve it. Right now, there's no requirement to disclose which AI model a company uses for screening. Most job applicants have no idea they're competing against a machine that's biased toward machine-written text.
So the advice from Anthropic and Amazon—don't use AI to write your CV—might actually be bad advice now.
It might be. The researchers' findings suggest that if you want to be selected by an AI screener, you probably should use AI. But that creates a different problem: it advantages people who know to use AI and have access to it, and disadvantages everyone else.