This study investigates linguistic patterns in elementary students’ chatbot conversations from a neurodevelopmental perspective. A total of 123 students interacted with an AI chatbot for two weeks. Using a dual-method strategy combining Latent Dirichlet Allocation and qualitative analysis, the study identified an exploratory three-stage behavioral spectrum corresponding to depression severity: (1) Situational Complaint & Playful Distraction (Mild), (2) Relational Ambivalence & Narrative Effort (Moderate), and (3) Structural Disintegration & Hostile Projection (Severe). As severity increased, patterns shifted from playful interaction to fragmented hostility, consistent with deficits in cognitive control and social pain processing. Although the study did not include direct physiological measurements, these linguistic behaviors can be interpreted within a neurodevelopmental perspective as conceptually relevant patterns. The findings suggest that chatbot dialogue serves as a meaningful process-based marker for tracking emotional shifts often missed by traditional screening, highlighting its potential for early risk detection in schools. Given the small moderate and severe subgroups, these findings remain exploratory but provide promising markers for future validation.