How does revealing AI authorship change reader trust?

Understanding Reader Perception Shifts upon Disclosure of AI Authorship reports a controlled study of 261 participants who judged a fictitious author across six acts of writing, producing 990 evaluations. Disclosure "generally erodes perceived trustworthiness, caring, competence, and likability," the paper finds, and "the most precipitous declines" fall in "social and interpersonal writing." The sharpest case is the Interact act, where AI use drew "strong negative reactions" and participants called the text "cold" and "impersonal." Object-oriented acts (Convince, Imagine, Explore) were viewed more favorably. The title leaves out competence on purpose: the discussion also reports that "in the Convince and Interact act, AI use boosted perceived competence," which sits uneasily with the abstract's general claim.

The design is deception-based. All 18 texts were entirely AI-generated, with minor human proofreading, but participants were told that a random share of sentences, from 0% to 100% in 10% steps, was generated or edited by AI. The paper attributes the negative shifts to three themes from participant feedback: "a perceived loss of human sincerity, diminished authorial effort, and the contextual inappropriateness of AI." Its regression analysis, whose specification the excerpt does not give, finds that a higher disclosed AI ratio "consistently led to more negative perceptions of the author." For the Interact act, the explanation is that AI "is perceived as incapable of genuine empathetic engagement," so its use "can be interpreted as a violation of social expectations, making human attribution critical."

The closest library note, on audience awareness of AI involvement, measures a different outcome. Its Thin Line evidence reports that awareness raised critical scrutiny while persuasive sway stayed between 34% and 62%, so disclosure modulates rather than blocks influence. This excerpt measures how the author is judged, not whether the reader is moved. Together the two suggest disclosure can cost an author standing even where the content still persuades. The author-side note on experienced and attributed authorship describes users claiming authorship they do not feel; this excerpt shows readers discounting the effort an author claims once a share is disclosed. The persona-distortion study measures perceived traits across 29 dimensions, but this excerpt holds fully AI-generated text fixed and varies only what readers are told, so its penalty concerns disclosure rather than the text.

The excerpt does not report effect sizes, the regression specification, the Table 3 values, or any manipulation check showing that participants believed the disclosed shares. It covers one Japanese participant sample, one fictitious author, one repeated-measures session, and texts written in English by GPT-4o and then translated; the conclusion itself calls for cross-cultural and longitudinal work. The authors also note that counting sentences ignores their semantic weight, so a disclosed core thesis may matter more than a disclosed detail. The implication is narrow: the penalty is a measured reaction to a stated share of AI involvement in short texts, not a fixed cost of disclosure in real collaboration, where AI may touch core arguments or only stylistic polish, a split this design cannot separate.