This study proposes an AI framework for automatically recognizing students' personality traits by combining Moodle learning logs, speech recordings, and handwriting data collected during regular classroom activities, without requiring specialized sensors or questionnaires. In an experiment involving 76 students at Sophia University, the framework achieved particularly high prediction accuracy for the Big Five personality traits, including Conscientiousness (97%) and Agreeableness (95%). The results further revealed that the choice and combination of data modalities have a greater impact on prediction performance than the choice of machine learning algorithm itself. The study highlights the potential of multimodal educational data for supporting personalized learning and adaptive educational interventions.

This exploratory study investigated how dependence on generative AI affects cognitive performance and cerebral blood flow, measured using functional near-infrared spectroscopy (fNIRS), among 82 young adults. The findings suggest that, when completing tasks without AI assistance, participants with higher levels of AI dependence achieved similar task performance but showed increased brain activity, indicating a compensatory cognitive effort. In contrast, when AI assistance was available, participants who relied heavily on AI completed tasks more quickly but exhibited reduced brain activation, whereas those who used AI more selectively maintained levels of brain activity comparable to those observed during independent problem solving. These results highlight the importance of balanced AI use and provide initial evidence regarding the cognitive implications of generative AI in learning environments.

This study proposes a lightweight multitask Transformer model that automatically selects only the most informative video frames (the top 30% keyframes) from classroom surveillance footage and simultaneously detects ten types of student behavior as well as the presence of abnormal behaviors. By leveraging the pretrained vision-language model CLIP, the proposed approach achieved approximately 91% accuracy in behavior recognition and 98.5% accuracy in abnormal behavior detection. In addition, it reduced training time by 31% and computational cost by 70% compared with conventional methods. The study also developed a web-based application that allows users to upload classroom videos and instantly receive analysis results, demonstrating both the practical applicability and efficiency of AI-assisted classroom monitoring systems in educational settings.