Why do people penalize AI disclosure even when they request it?

Siavosh Sahebi, Paul Formosa, and Sarah Bankins, cited in Taylor Dykes's Nielsen Norman Group article on AI disclosure, name a "disclosure paradox": the same people who say they want AI use disclosed go on to react negatively once it is. In their study, Dykes reports, "participants rated AI-attributed text as less trustworthy, less authentic, and less useful, even though they said that disclosing AI use was important."

The paradox separates a stated preference (disclosure should happen) from a revealed reaction (disclosed content gets penalized), and the article treats this gap as one reason disclosure guidance cannot follow a single rule: giving audiences what they ask for does not protect the content from the rating drop that follows. The article pairs this with a related but distinct asymmetry: Jingchao Fang, Victoria Xiaohan Wen, and Mina Lee found that writers were "far less likely than readers to consider an AI disclosure necessary" when the two groups judged the same writing scenarios — people want to be told about others' AI use more than they want to disclose their own.

The paradox adds a stated-versus-revealed-preference framing that the library's nearest disclosure notes mostly test behaviorally rather than name directly. Do readers trust unlabeled AI-written messages as much as human ones? shows a comparable gap from the disclosure side — unlabeled content rated fine, disclosed content penalized — without asking participants what they wanted; Sahebi and colleagues' contribution is to show the gap holds even when the same participants explicitly endorse disclosure as important. It sits alongside How does revealing AI authorship change reader trust? and Does disclosing AI use damage how trustworthy you seem? as further evidence that disclosure carries a measurable trust cost independent of whether audiences say they want it. It is cited within When should you disclose AI use in written content? as one reason no single disclosure rule applies.

The excerpt gives no sample size, method detail, or publication venue for Sahebi, Formosa, and Bankins' study beyond Dykes's secondhand summary, so the paradox should be read as a reported finding relayed through a design article rather than one checked against the original paper. If the pattern holds beyond this one study, it implies that satisfying a stated demand for transparency is not sufficient to protect perceived trust or quality — disclosure policy built only on what audiences say they want may still produce the rating penalty that disclosure itself causes.