It seems that AI is sefl-preferncing and each LLM model prefers its own content.
What happens when a candidate uses AI to write a CV but it’s not the same platform as the ATS? Find out..
AI is often marketed as the solution to managing bias in the hiring process so that everyone has an equal shot on a level playing field. But now, according to recent research, that story is getting more nuanced. A growing body of research suggests AI screening doesn’t eliminate the old distortions, so much as introduce new ones, and some of them are stranger than anything we’ve seen before. Thanks to Abu Dhabi based, Global Head of Talent, Matt Buckland for flagging this up to me.
Researchers carried out an experiment that was a little unsettling. They took a number of real CVs, written by real people, before ChatGPT existed, so no one could accuse them of sneakily running them through AI. The owners all had the same basic skills and experience and team sizes. Then they processed those CVs through seven AI models (GPT-4o, LLaMA 3.3-70B, Qwen 2.5-72B, DeepSeek-V3, Mistral-7B, and others) and asked each one to rewrite them.
The follow up was asking every platform to evaluate each CV.
Every model picked the one they had created themselves.
Bias against human-written CVs
This happened with almost 100% consistency so that there was no way the findings could be presented as discrepancies. The result showed that the bias against human-written CVs was substantial, with self-preference ranging from 67% to 82% across major commercial and open-source models. GPT-4o chose its own rewrite 97.6% of the time. The original, written by a human, with judgment, skills and lived experience, almost never won.
You might wonder: maybe the AI versions really are better or is this is just progress? The researchers tested that too. Real humans graded the CVs for clarity, coherence, usefulness and also controlled for quality, but the bias not only didn’t shrink, it worsened. Even when human reviewers said plainly that the original was the stronger document, the models still picked their own work.
More than AI prefers AI
This goes deeper than “AI prefers AI.” AI prefers its own AI. This bias is produced within AI-AI interactions when the model’s own evaluative behaviour systematically favours outputs that align with its own generative patterns. DeepSeek-V3 leaned toward its own output far more than it did toward LLaMA’s. GPT-4o did the same. The contention is that there is something akin to an accent or dialect at work, identifiable markers, perhaps a phrasing pattern which each model recognises as its own, and marks that marker accordingly.
The labour market consequences are evident. Simulations across 24 occupations show that candidates using the same LLM as the evaluator are 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written CVs, with the largest disadvantages in business-related fields such as sales and accounting.
That will be a key difference between getting an interview and getting nothing.
An additional hurdle
Most large companies are already running some form of AI processing inbound applications, which means the system doing the screening and the system that rewrote your CV might be the same model, or close enough to recognise the family resemblance. Or maybe the LLM will be a different version, in which case your CV will not be rated as highly.
As the researchers put it, this bias rewards access to very specific generative technologies and penalises those without it, even when applicants are otherwise equally qualified.
You could spend years building real experience and learning how to tell your story well, as well as write a CV that actually reflects your personality and achievements. Or you could have the same model that’s screening your CV, rewrite it in its own image first. The real issue is how can you know which ATS system is using which particular LLM platform.
One of those LLM modelled CVs is more likely to get you shortlisted, rather than the work you do. The researchers maintain that this bias can be reduced by more than 50% through simple interventions targeting LLMs’ self-recognition capabilities
When someone gets rejected now, there’s a new question sitting alongside the old ones. Not just was there a better candidate? But: did they sound more like the machine reading them?
At 3Plus International, we work one-to-one with professionals to cut through the noise and help you show up as the distinctive, high-value candidate you actually are so AI works for you, not replaces you.
Let’s make sure your search strategy is yours. Discover our individual services – contact us
