Large Language Models (LLMs) are transforming computer science (CS) education by providing immediate problem-solving support and on-demand knowledge. However, commercial LLMs frequently deliver complete solutions, fostering overreliance and diminishing the productive struggle essential for deep learning. Neurological study further suggests that LLM dependence can significantly reduce brain activity, threatening long-term retention. In response, educators have begun customizing LLM-based AI assistants to deliver structured scaffolding. Despite these efforts, how different AI interaction patterns shape learning outcomes in CS contexts remains poorly understood. To address this gap, we propose Chat-Bloom (CB), a taxonomy that classifies the cognitive offloading level of both student prompts and AI responses. We validated CB through inter-rater reliability testing (Gwet’s AC2 = 0.81) and trained ML models achieving 81% classification accuracy for large-scale analysis. Applying CB to CS1 chat logs, we identify distinct interaction sequences, compare cognitive offloading patterns across AI assistants, and examine their relationship to student outcomes. These findings position CB as a practical framework for evaluating and improving AI-assisted learning in CS education.