When an algorithm does the screening, can AI-written text win just by looking polished and authoritative?

This explores whether AI-styled text gets an unfair edge when an algorithm does the screening, and whether surface style is what wins. The corpus has no direct studies of screening résumés, admissions essays or grant applications. It does have strong evidence on the mechanism such an advantage would run through: automated evaluators reward how text looks, sometimes more than what it says.

The clearest evidence is about LLM judges. Two of their biases are 'semantics-agnostic': an authority bias (trusting text that cites references, even fake ones) and a beauty bias (preferring rich formatting like headers, bullets and bold). Adding fake citations or extra formatting reliably raises a judge's score, with no access to the model and no clever optimization needed (Can LLM judges be fooled by fake credentials and formatting?). Model-generated text tends to come with exactly these features already built in: tidy structure, confident framing, reference-like phrasing. So the advantage doesn't have to be designed. It can show up by default whenever one model's habits match another model's preferences.

Why would model text carry those habits so consistently? One reason: generation follows a smooth path toward the most typical continuation rather than exploring competing positions (Does LLM generation explore competing claims while producing text?). The result is polished, middle-of-the-distribution prose. That is also the kind of text a model-based judge, trained on similar data, finds most 'normal' and credible. A related warning applies: model output reflects the model's learned patterns, not evidence about the world, so treating it as a real signal of quality is a category error (Should we treat LLM outputs as real empirical data?).

The less obvious part is that style is the easiest layer to fake and the easiest to detect, but it isn't the most telling one. Telling authors apart by stylistic patterns is nearly solved, even for small models (Can language models truly understand literary style?). Yet AI fiction stays detectable with style removed entirely, from deeper choices like how characters act and how time is ordered (Can AI stories be detected without analyzing writing style?). For screening, that cuts both ways. A screener that scores surface polish can be gamed. A screener built on deeper structure is sturdier, and it may also flag AI-assisted writing that surface 'humanizing' edits don't hide. There is a related risk: frontier models damage documents through subtle corruption that leaves the surface intact (Does model capability change how documents degrade?). A polished document is not necessarily a sound one, and a style-sensitive screener will struggle to tell the two apart.

In short, the corpus supports the mechanism: automated evaluators can be swayed by formatting and authority cues that AI text produces naturally. It does not yet measure how large that advantage is in real hiring, admissions or peer-review pipelines. That gap is worth noticing in itself.