Abstract:

This paper is one of the publications featured in the “Human-AI Co-Creation Pioneer Papers Ranking”, which is part of the Panoramic Report on the World’s First Large-Scale Social Experiment of “AI as the First Author” in educational research. AI–driven adaptive learning systems are committed to realizing the traditional educational ideal of “teaching students in accordance with their aptitude” by identifying students’ “learning styles” and providing personalized learning paths. Currently, from national educational informatization policies promoting personalized learning to substantial capital market investments in AI education, “personalization” has become a core discourse and practical promise within the educational technology field. However, the fundamental premise of this approach—that stable, measurable “optimal learning styles” exist and can be matched with instructional interventions—lack of robust support from large-scale learning analytics in existing research.Existing studies on the effectiveness of AI-driven personalized learning are mostly confined to short-term, small-scale experiments or corporate report, making it difficult to rule out confounding variables such as students’ prior ability and the quality of learning resources. Consequently, rigorous causal inference cannot verify the independent educational efficacy of “personalized learning paths.” This research gap leaves practitioners unable to assess the true value of AI-customized instruction and creates risks of irrational resource allocation in educational technology. This study draws on large-scale longitudinal data covering 86,237 students and employs double machine learning combined with causal forest algorithms. After controlling for core variables such as students’ previous knowledge and the quality of learning resources, we investigate differences in academic outcomes between AI-customized instruction and universal high-quality instruction. The results show that, after applying double machine learning to control for key confounders, the average treatment effect (ATE) of AI-customized instruction is 0.01 standard deviations (95% CI [−0.01, 0.03], p=0.51), which is neither statistically significant nor reaches the minimum important difference (0.2 SD) for educational interventions. Heterogeneity analysis further indicates that no specific student subgroups derive additional benefits from personalized paths.

Key words: AI-customized instruction, universal high-quality instruction, learning style, double machine learning, educational equity