Trust the machine or the human? A dynamic neuro-cognitive model of AIGC credibility assessment

With the proliferation of Artificial Intelligence Generated Content (AIGC) in digital education，understanding how students assess the credibility of AI-generated content has become particularly crucial. To investigate college students' credibility assessments of AIGC, the present study draws on predictive coding theory and dual-process theory to construct and provide initial empirical support for a "Dynamic Neuro-Cognitive Model of AIGC Credibility Assessment". This study employed a 2 (information source: AI vs. human) × 2 (content type: opinion-based vs. factual) within-subjects design. By integrating behavioral measures with event-related potentials (ERPs), we examined both behavioral performance and cognitive-neural characteristics as students read texts from different sources and of different types. Behavioral results indicated that AIGC elicited lower credibility ratings and longer reaction times than human-generated content, reflecting algorithm aversion and heightened cognitive vigilance. In contrast, opinion-based content elicited higher credibility ratings and shorter reaction times than factual content. At the neural level, both AIGC and opinion-based content elicited larger N400 and Late Positive Potential (LPP) amplitudes, consistent with greater semantic prediction error at early stages and deeper emotional engagement and motivated reasoning at later stages. These findings suggest that cognitive trust and affective trust jointly contribute to human-machine trust. This study provides preliminary ERP-based neural evidence for understanding human-machine trust and offering critical implications for the future optimization of AIGC design and fostering public trust in AI-generated information.