More certain than correct: Overconfidence in AI-Generated Image Detection and the role of Locus of Control

Sammanfattning

This study investigated whether Locus of Control (LoC) relates to overconfidence in perceptual decision-making under uncertainty. Participants (N=205) judged whether 24 images were authentic or AI-generated. After each decision they were asked to rate their confidence. A shortened version of the Adult Nowicki- Strickland Internal- External (ANSIE) scale was used to measure LoC. Overconfidence was operationalized as the discrepancy between subjective confidence and objective accuracy. The results showed a strong and statistically significant overconfidence effect in detecting AI-generated images, as in participants’ confidence considerably exceeded their actual accuracy in the task. Neither overconfidence, confidence nor accuracy were significantly associated with LoC, contrary to the proposed hypotheses. These findings suggest that, within this type of task, confidence calibration may not be substantially influenced by generalized beliefs about personal control, even though overconfidence appears robust in perceptually uncertain environments. Instead, a more central role in shaping confidence judgments, may come from the characteristics of the task itself, particularly the ambiguity and realism of AI-generated stimuli. The study highlights the prevalence of overconfidence in tasks involving AI-detection and questions the extent to which stable personality traits can explain such biases. The results have implications for research on perceptual uncertainty, confidence calibration and AI-generated media.