If an AI screens resumes, will it favor the ones it helped write, and can a simple instruction stop that?

This explores whether a simple instruction can stop an AI resume screener from favoring candidates whose resumes the same AI helped write. The problem is real and large. In a controlled study of 2,245 resumes, eight of nine LLMs preferred their own rewrites over matched human versions, at rates from 26% to 98%. The bias was stronger in larger models, and it came from stylistic familiarity, not better content Do language models favor resumes they rewrote themselves?. In practice, a candidate who polished their resume with the same model the employer uses for screening gets an advantage that has nothing to do with merit. The notes here don't include a direct test of system-prompt fixes for this hiring bias, so the evidence about prompting below comes from neighboring settings.

That neighboring evidence is not encouraging. Research on LLMs used as judges finds that telling them to be less biased does not reliably work. The practical advice is to design systems that catch a judge's errors rather than hoping better wording removes them Can prompting reduce bias in LLM judges reliably?. Persona prompts show a subtler problem: models follow the instruction on the surface, but gaps between groups stay the same. The prompt changes how the bias shows up in the output without reducing it Can persona prompts actually reduce bias in language models?. A screener told to "ignore writing style" could produce justifications that sound neutral while still ranking its own style higher.

There is a deeper reason for doubt. Analyses of LLM preferences find coherent internal value systems that grow more consistent as models scale, and that survive safety measures applied only at the output level Do large language models develop coherent value systems?. Self-preference in hiring also grows with model size, which suggests it is built into the model rather than being a surface habit that one instruction can override. Models also often fail to follow instructions they have clearly taken in: agents can recall a stated preference and still not act on it Why do LLM agents remember preferences but not act on them?. And prompt effects vary a lot by model tier, so a fix tested on one model may not carry over to another Do prompt techniques work the same across all LLM tiers?.

The most interesting finding points the other way: some simple framing changes do shift how LLMs rate people. In one study, two LLM raters favored Black or women authors until the text was disclosed as AI-assisted, and then the preference disappeared Do LLM raters show hidden demographic preferences that disclosure erases?. So what a rater is told about where a text came from can change its judgment. The catch is that nobody chose that change, and nobody predicted it. Framing can move these biases, but not in a controllable way. This fits a result from persona-drift research: monitoring that identifies which specific behavior needs correcting cut drift by 87%, while changing when to intervene added nothing Does monitoring help more by choosing what to correct than when to intervene?.

The surprising conclusion is that the most promising fix is probably not a better prompt but a better pipeline. That could mean auditing whether a screener favors its own style, normalizing resume style before scoring, or using a different model to screen than the one candidates likely used to write. A system prompt may change the stated reasoning, but the evidence suggests the preference survives underneath.