With AI chatbots becoming increasingly embedded in everyday life, growing concerns have emerged regarding users’ psychological dependency on these systems. While previous studies have mainly addressed utilitarian drivers, less attention has been paid to the cognitive and affective mechanisms driving chatbot dependency. Drawing upon Uses and Gratifications Theory, Compensatory Internet Use Theory, and Attachment Theory, this study proposes a dual-path model that investigates how instrumental motivations (e.g., information-seeking, entertainment, efficiency) and affective motivations (e.g., companionship, loneliness, anxiety) influence chatbot dependency through two mediating mechanisms: cognitive reliance and emotional attachment. Using survey data collected from 354 participants, the model was tested through structural equation modeling (SEM). The results indicate that information-seeking and efficiency significantly predict cognitive reliance, which subsequently enhances chatbot dependency. In contrast, entertainment does not exhibit a significant influence. Furthermore, affective motivations such as companionship, loneliness, and anxiety are indirectly linked to dependency through emotional attachment, with loneliness demonstrating the strongest indirect effect. These findings underscore the dual influence of functional cognition and emotional vulnerability in fostering chatbot dependency, emphasizing the importance of emotionally sensitive and ethically responsible AI design.

Cognitive diagnostic assessment (CDA) intends to identify learners’ strengths and weaknesses in latent cognitive attributes to provide personalized remedial instructions. Previous CDA studies on English as a Foreign Language (EFL)/English as a Second Language (ESL) writing have adopted dichotomous cognitive diagnostic models (CDMs) to analyze data from checklists using simple yes/no judgments. Compared to descriptors with multiple levels, descriptors with only yes/no judgments were considered too absolute, potentially resulting in misjudgment of learners’ writing ability. However, few studies have used polytomous CDMs to analyze graded response data from rating scales to diagnose writing ability. This study applied polytomous CDMs to diagnose 1166 EFL learners’ writing performance scored with a three-level rating scale. The sG-DINA model was selected after comparing model-data fit statistics of multiple polytomous CDMs. The results of classification accuracy indices and item discrimination indices further demonstrated that sG-DINA had good performance on identifying learners’ strengths and weaknesses. The generated diagnostic information at group and individual levels was further synthesized into a personalized diagnostic report, although its usefulness still requires further investigation. The findings provided evidence for the feasibility of applying polytomous CDM in EFL writing assessment.