Label or not: AI-generated news labels and perceived credibility among Chinese readers

Introduction Artificial intelligence increasingly writes the news, and Chinese regulation has required explicit labels on AI-generated content since September 2025, yet what such labels do to readers’ trust remains an open question. This experiment tested whether an AI-generated news label changes perceived credibility, whether perceived human involvement explains the change, and whether readers’ AI literacy conditions it. Methods A sample of 1,024 Chinese adults (51.7% female; mean age 30.8 years) recruited through an online panel was randomly assigned to read one of two news stories that either carried a prominent AI-generated label or carried no authorship label, in a 2 (label: AI-generated vs. no label) × 2 (news topic: health technology vs. urban culture) between-subjects design with 256 per cell. Participants rated perceived human involvement, message credibility, and their own AI literacy. Results The label lowered perceived credibility (β = −0.31, p < 0.001) and perceived human involvement ( F (1, 1,020) = 218.60, p < 0.001). Involvement mediated the label effect on credibility: the indirect effect was β = −0.19, 95% CI [−0.23, −0.16], about three fifths of the total effect. AI literacy moderated the label-to-involvement path (β = 0.10, p < 0.001): the label penalty was β = −0.52 at low literacy and β = −0.32 at high literacy, and the index of moderated mediation was β = 0.05, 95% CI [0.02, 0.08]. Discussion The credibility cost of AI news labels travels through perceived human involvement and concentrates among readers with little AI literacy. Labeling regimes can keep AI-generated news transparent without flatlining trust, but they work best alongside AI literacy education.