This paper investigates how ideological bias and classification errors vary across AI-mediated information environments. We conduct an incentivized laboratory experiment in which participants classify news items as true or false across three withinsubject phases: a human-written baseline, an AI-generated misinformation phase in which the role of AI is not made explicit, and an AI-detection phase in which participants receive an explicit AI probability signal. Average classification accuracy does not differ significantly across phases. The central result is instead that errors are strongly structured by ideological congruence: participants are more likely to accept congruent false news as true and to reject incongruent true news as false. The AI-detection phase suggests greater skepticism toward information, reflected in a higher false-negative rate. This skepticism seems to be selectively applied: the relation between congruence and classification errors becomes stronger when AI is introduced explicitly. The analysis of detector responsiveness points to non-selective partial mistrust of AI signals. Overall, AI-based verification does not mechanically correct ideologically structured classification errors.
