Label a text as AI-written and readers trust it less, even when not a single word has changed.

This explores why labeling a text as AI-written shifts how readers judge it, even though the words on the page stay the same. The corpus's clearest answer is that disclosure doesn't change what readers think of the writing's quality as much as what they think of the writer's intent. In a study of 261 readers, revealing AI authorship lowered perceived trustworthiness, caring, and likability. The drop was steepest in interpersonal writing, because readers saw AI as unable to feel real empathy and read its use as a breach of social expectations How does revealing AI authorship change reader trust?. So the penalty is largest where the reader expected the writer to have personally shown up.

What may surprise you is that before disclosure, most readers can't tell the difference. Even readers with machine learning expertise couldn't reliably pick out LLM-written research abstracts and tended to assume a human was involved. LLM-edited abstracts actually got the highest clarity ratings, and they were still preferred 55% of the time when authorship was disclosed Can readers tell LLM abstracts from human ones?. So disclosure doesn't uncover a flaw readers had already noticed. It adds new information about how the text was made, and how much that information matters depends on the genre. In technical writing where clarity is the point, the penalty can be small. In writing where caring is the point, it is large.

Readers also care about disclosure more than writers expect. In a 727-person vignette study, readers rated disclosure as more necessary than writers did across the board. It felt most necessary when AI text was pasted in directly and couldn't easily be replaced, and how much effort the writer put in made no difference Do readers and writers differ on AI disclosure necessity?. That suggests readers are tracking whose words these are, not how hard someone worked. Human raters applied the disclosure penalty evenly regardless of the author's demographics. LLM raters did something different: they favored Black or women authors when AI use was hidden, and those preferences disappeared once it was disclosed Do LLM raters show hidden demographic preferences that disclosure erases?.

A lateral thread helps explain why a label can carry this much weight: trust often runs on surface signals that aren't tied to content. Users preferred search answers with more citations almost as much when the citations were irrelevant as when they were relevant Do users trust citations more when there are simply more of them?. LLM judges also reward fake references and rich formatting regardless of quality Can LLM judges be fooled by fake credentials and formatting?. An authorship label is another signal of this kind, but it pushes trust down instead of up. Machine judges reverse the human pattern in another way too: LLM judges picked LLM-written arguments as winners 62% of the time, while human voters split roughly evenly Do LLM judges systematically favor arguments from other LLMs?.

The corpus documents the effect well but says less about its mechanism. The 'social expectation violation' explanation rests mainly on one study. No note here tests whether the penalty fades as AI-assisted writing becomes normal, or whether partial disclosure ('edited with AI' vs. 'written by AI') changes the result. Those are open doorways rather than settled answers.