A plain 'AI was used' label mostly costs trust, but showing where each claim came from can rebuild it.

This explores whether being open about AI's role in a piece of writing can rebuild reader trust, or whether it mainly costs trust. The corpus gives a mixed answer. A plain "AI was used" label almost always lowers trust at first. Two other things do build trust: showing readers where each claim came from, and letting them see results over time. So the answer depends on what kind of transparency you mean.

Start with the label alone. When an identical news article carried an AI disclosure statement, both human readers and LLM raters scored it lower. The drop was small, under 0.15 points on a 7-point scale, but it showed up consistently (Does disclosing AI assistance make readers trust articles less?). The penalty is much bigger when the writing is personal. In interpersonal messages, readers saw AI use as a breach of social expectations, because a machine can't actually care (How does revealing AI authorship change reader trust?). There's an awkward baseline too: readers trust unlabeled AI-assisted emails just as much as human-written ones. Skepticism only kicks in when someone says AI was involved (Do readers trust unlabeled AI-written messages as much as human ones?). In the short run, then, honesty gets punished and silence doesn't. That baseline may not last. The researchers expect default trust to shrink as people learn how common AI writing is, though their one-time study couldn't measure that (Does trust in unlabeled AI messages decline as awareness grows?).

The penalty can shrink or even reverse under some conditions. Readers with higher AI literacy lost less trust after a disclosure, and some of them viewed AI use positively (Does AI literacy reduce the damage from AI disclosure?). Time matters most. People at first avoid partners they know are AI, but that preference flips after repeated interactions where they can see the results. Disclosure without that feedback taught people nothing (Does revealing AI identity help or hurt user trust?). One-time transparency triggers a bias. Transparency combined with a visible track record is what lets people adjust their trust to match reality.

The most useful finding for the "how" part of the question comes from work on provenance, meaning showing where information came from. A newsroom tool that links every number and quote to its original source made output checkable rather than just fluent. That traceability, not polish, is what made professional newsrooms willing to adopt it (Can source traceability make AI writing trustworthy?). In another study, readers with no source cues couldn't tell real claims from fluent fabrications at all. A display showing which claims had been verified brought their judgment back (Can readers tell truth from fabrication without evidence signals?). Without signals like that, readers tend to accept fluent text without checking it (When do users stop checking whether AI output is actually backed?). Disclosure does make audiences more critical, but 34–62% were still persuaded (Does telling people an AI wrote something actually stop them from believing it?). A label makes people wary. Provenance gives them a way to check.

Two complications show that this is a social negotiation, not just a design choice. Readers think disclosure is more necessary than writers do, especially when AI text is used directly and couldn't easily be replaced. How much effort the writer put in didn't change readers' view (Do readers and writers differ on AI disclosure necessity?). And suspicion has costs of its own: people accused of using AI often wrote their comments themselves, so accusations can end up wronging human writers (Do unfounded AI accusations harm human writers instead?). One caveat: the corpus has little direct testing of detailed "how and when" disclosures, such as "AI drafted section 2, a human checked every figure," compared with a blanket label. The pattern suggests detailed, checkable transparency should beat a yes/no flag, but that's an inference, not a tested result.