People say AI disclosure labels matter, then trust the labeled content less anyway — should policy plan for that gap?

This asks whether AI disclosure rules should plan for the trust drop that follows a disclosure as its own problem, apart from what audiences say they want. The corpus suggests yes, because stated preferences and actual reactions come apart. Sahebi, Formosa, and Bankins found that people who said disclosure was important still rated AI-labeled text as less trustworthy, less authentic, and less useful once it was labeled Why do people penalize AI disclosure even when they request it?. A policy built only on surveys of what audiences want would require the label and then be caught off guard by the penalty that comes with it.

The tension sharpens because the audience asking loudest for labels isn't the one who pays for them. In a 727-person study, readers rated disclosure as more necessary than writers did, especially when AI text was pasted in directly and couldn't easily be replaced Do readers and writers differ on AI disclosure necessity?. Writers carry the trust cost and readers set the demand, so a preference-only rule settles the question on the side that faces no downside.

That doesn't make hiding AI use the safe choice. Schilke and Reimann found that AI use kept quiet and later uncovered produces the steepest trust decline of all, worse than disclosing upfront Does hidden AI use cost more trust when exposed?. So the penalty is real, but it's the smaller of two costs. A good policy treats disclosure as accepting a known, limited hit in order to avoid a larger hit that might come later.

The less obvious point is that the penalty may not be permanent, and policy can shape how long it lasts. Research on revealing that a partner is an AI found that people first avoid AI partners, but the bias reverses once they repeatedly see how things actually turn out. Disclosure without that outcome feedback produced no adjustment at all Does revealing AI identity help or hurt user trust?. That changes the design question from "should we label?" to "does the label come with a way for people to see whether the AI-assisted work held up?" A one-time label may simply lock in the first-impression penalty.

A caveat from nearby work: trust signals are easy to game in both directions. Users favor answers with more citations even when the citations are irrelevant Do users trust citations more when there are simply more of them?. That suggests audiences react to the surface of a label more than to what it actually says. The corpus is strong on the gap between what people say they want and how they react, but thin on tested ways to word or format a disclosure, so how to shrink the penalty in practice is still an open question here.