When AI Agrees with Me: Disclosure Labels, Confirmation Bias, and the Sharing of AI-Generated Misinformation Among Jordanian Young Adults

This study examines whether AI disclosure labels alter the evaluation and sharing of misinformation when messages align with recipients’ prior beliefs. A randomized mixed-factorial online experiment involved 900 Jordanian young adults who evaluated eight Arabic-language social-media posts each (7200 post-level observations). The disclosure condition (no label, “AI-generated,” or “AI-generated: This information has not been independently verified”) varied between participants; veracity and belief congruence varied within participants. Crossed mixed-effects models showed that a generic AI label modestly reduced message credibility but did not significantly reduce sharing intention, whereas the explicit verification warning reduced both outcomes relative to the generic label. Belief-congruent misinformation received higher credibility and sharing ratings, with a stronger credibility advantage for false than true posts. Message credibility partially mediated the congruence–sharing association. Actively open-minded thinking weakened congruence effects and improved true–false credibility discernment. The findings extend the Heuristic-Systematic Model by showing that disclosure wording, motivated processing, and individual thinking dispositions shape distinct stages of misinformation response. They support labels that communicate the verification status and media-literacy interventions that help users distinguish agreement from accuracy. The baseline-adjusted sensitivity analysis did not support the attitude-reinforcement interaction (b = −0.030, p = 0.628).