Enterprise AI Analysis

A randomized control trial was conducted with students from a research-oriented university in China, assigned to control, process-only (Po), and process-with-condition (PwC) groups. Data collection involved pre-task surveys (demographics, ISDIMU, prior knowledge) and real-time trace data (clicks, mouse movements, keystrokes) during a 2-hour academic English writing task. Trace data was coded into learning actions and then into seven SRL processes. Scaffolds were triggered at specific times (5th, 56th, 93rd minutes) for Po (SRL process-based) and PwC (SRL process and learning condition-based) groups, with messages generated by GPT-40 using structured prompts.

Ordered Network Analysis (ONA) revealed that the PwC group (analytics + GenAI with conditions) demonstrated significantly more effective metacognitive learning patterns compared to the control group, engaging more in oriented writing by utilizing task instructions and rubrics. The Po group (analytics + GenAI for process only) showed some improvements but did not produce significant differences in SRL patterns compared to the control group. This indicates that incorporating dynamic learning conditions into GenAI-powered scaffolds is crucial for enhancing SRL.

Dynamic Time Warping (DTW) and k-means clustering identified two groups: compliers and non-compliers. Compliers exhibited intensive or sustained engagement with prompted SRL processes, while non-compliers largely disregarded suggestions. Differences in SRL patterns between compliers and non-compliers became significant after the second scaffold, where compliers revisited previous notes and information, demonstrating effective writing, while non-compliers followed a more linear process (checking instructions, continuous reading, then writing).

The study confirms that analytics-based GenAI scaffolding can effectively support SRL by adaptively responding to real-time SRL processes and learning conditions. It highlights the importance of considering learners' conditions for effective scaffolding and demonstrates that GenAI can capture dynamic changes at scale. The findings contribute to the literature by designing, implementing, and evaluating the impact of adaptive scaffolds on learners' SRL processes using real-time analytics with GenAI.

Limitations include the contextual nature of the prompt design, which may limit generalizability. Future research should replicate the study across different domains and age groups. Non-compliance by some students warrants further investigation (e.g., through interviews) to refine scaffolding. Small sample size for the third scaffold's complier/non-complier analysis also limited findings. Future work will leverage GenAI for scalable, adaptive learning support and AI-driven personalization.