Generative artificial intelligence (GAI) is increasingly used as a pedagogical assistant in online learning; yet how young students, such as secondary school students, enact epistemic agency (EA) when learning with GAI remains underexplored. In this study, we examined the types and levels of EA demonstrated by secondary school students with scaffolded GAI and investigated the relation among students' prior academic proficiency, self-efficacy in learning, and EA demonstrated. Participants were 122 seventh graders who engaged in an online bilingual inquiry unit on photosynthesis. The unit supports students in Taiwan in learning science inquiry and English as a foreign language, with GAI prompted and integrated to support students' scientific explanation writing in English. Data collected and analyzed included students’ prior academic proficiency through assessments, self-efficacy in learning through questionnaires, and student–GAI conversation turns through textual conversation logs. Results identified seven types of student–GAI shared epistemic agency (SEA), with at least 86% of students demonstrating such agency at least once. Correlation results indicated that the frequency of EA demonstrated by students was associated with their prior language proficiency. Cluster analysis with the retained four-profile solution further suggested that student–GAI SEA was likely to be activity-specific. The identified SEA categories provide a framework for examining student–GAI SEA, and the instructional design illustrates how GAI can be prompted to preserve opportunities for student contribution. Design implications for supporting students with different needs are discussed.
Computational thinking (CT) is a core problem-solving competence that originated in computer science and now applies across disciplines and to everyday life. It has been widely integrated into K–12 curricula, which creates an urgent need for valid assessments of students’ performance. However, instruments targeting CT practices remain scarce for lower primary learners, and many tools depend on specific programming platforms, capture limited process evidence, or lack adequate psychometric evaluation. This study addresses this gap by developing and validating a game-based assessment tool for CT practices in lower primary students, named CT Practices Quest for Lower Primary (CTPQ-LP). The tool was developed using an Evidence-Centered Game Design (ECGD) framework, with coherent student, task, and evidence models designed to operationalize five practices: pattern recognition, abstraction, problem decomposition, algorithmic thinking, and debugging. The content was validated through expert review and cognitive interviews, followed by a pilot study (n = 74) and a main study (n = 510) administered to students in Grades 1–3 at a Chinese primary school. We examined reliability, item characteristics, and factor structure using classical test theory (CTT) and confirmatory factor analysis (CFA). We further investigated performance differences across background characteristics and analyzed gameplay records to identify challenges encountered in solving CT problems. The study contributes an assessment tool and a transparent development procedure that can inform CT classroom practice, guide future assessment design, and support evidence-informed improvement of early primary instruction.
Emotions play a fundamental role in learning by shaping learners' cognitive engagement and influencing learning outcomes. In generative AI (GAI)-supported programming, learners interact with AI in different ways. However, limited research has examined how different interaction modes are associated with learners' emotional dynamics. This study investigated students' emotions across three human–AI interaction modes (i.e., AI-reliant, AI-independent, and human–AI hybrid) during AI-assisted programming. A total of 150 undergraduate students completed programming tasks with access to ChatGPT, during which video recordings of their facial expressions and electrodermal activity (EDA) were collected. Facial expressions were analyzed to examine both emotion distributions and temporal transitions using lag sequential analysis, while EDA was used to measure affective arousal. Results reveal that none of the seven facial-expression categories differed significantly across the three modes. Lag sequential analysis uncovers that the three modes shared reciprocal anger ↔ disgust, fear ↔ surprise, and neutral ↔ sadness transitions. Descriptively, the AI-reliant mode showed transitions between positive and negative emotions, the AI-independent mode showed a relatively neutral-centered transition network, and the human-AI hybrid mode demonstrated a relatively surprise-centered transition network. However, none of the transition probabilities differed significantly across modes. In terms of EDA analysis, the hybrid mode exhibited significantly more skin conductance responses (SCRs) in total and significantly more SCRs per minute than the AI-reliant mode. These findings suggest that emotion states can serve as signals for understanding how learners engage with GAI during programming. The study highlights the importance of designing emotion-aware AI-supported programming environments that balance technological assistance and learners’ emotional engagement in problem-solving.
