Causal attributions for ChatGPT errors: what is the role of AI literacy?

This exploratory study examined whether users’ AI literacy is associated with the causal explanations they provide for a specific type of incorrect output generated by the ChatGPT chatbot. The aim is to offer an exploratory framework that may contribute to filling a research gap, particularly in the Italian context, concerning the causal dimensions of the attributions made by AI users. An online sample of 348 adults recruited via Prolific read a scenario describing frequent ChatGPT errors in generating APA-formatted bibliographic citations. Participants selected the primary cause of the errors and then rated its perceived causal properties using the Italian Causal Dimension Scale (SDCit), capturing locus (internal vs. external to the chatbot), stability (stable vs. variable), and controllability (by the chatbot vs. by others). AI literacy was assessed with the Artificial Intelligence Literacy Scale (AILS) and categorised into four levels. Multivariate GLM analyses (MANCOVA) tested associations between AI literacy and causal dimensions, with and without adjustment for age, gender, and frequency of use. Higher AI literacy appears to be associated with more technical and potentially remediable explanations for incorrect outputs generated by ChatGPT, thereby supporting educational and communication interventions aimed at fostering calibrated trust and critical engagement with AI systems.