A survey of 486 undergraduates in China found that students who experienced their AI learning tools as highly adaptive reported less enjoyment, more worry while studying and more boredom during the semester. By the end of the semester, they also reported weaker self-regulated learning—the ability to plan, manage and adjust their own study.
The findings do not show that adaptive AI caused these changes. The study was observational, relied on self-reports and did not measure grades, usage logs or observed study behavior. It also combined several types of AI tools, meaning the results cannot be attributed to one particular platform.
What the survey found
Jing Li and colleagues surveyed students in 28 classes at four Chinese universities across one semester. The students completed measures during weeks 2–3, 9–10 and 16–17. The survey assessed how adaptive, personalized and transparent their AI tools felt, along with AI literacy, study emotions and self-regulated learning.
Students who gave their AI environment a higher adaptability rating at the start of the semester tended to report lower self-regulated learning at the end. The overall standardized link was −0.42. In simpler correlations, perceived adaptability was associated with lower enjoyment, more anxiety related to studying and more boredom.
Lower enjoyment and higher anxiety and boredom at mid-semester accounted for 54.8% of the link between adaptive-feeling AI and weaker self-regulated learning. The study’s model explained 28.4% of the differences in self-regulated learning between students.
Students averaged 3.8 hours per week using AI tools. Reported tools included general chatbots such as ChatGPT and Wenxin Yiyan, university tutoring platforms, course-specific recommendation systems and language-learning apps. Only the tutoring platforms and course recommendation systems were described as truly adaptive; the main measure captured how tailored the overall AI experience felt to students.
Why personalization might matter
Adaptive learning platforms use algorithms to change what a student sees next. They may adjust difficulty, order, pacing or feedback based on answers, speed and errors.
This can be useful, but researchers have also raised a “personalization paradox.” When software makes more decisions about planning and pacing, students may have fewer opportunities to practise making those decisions themselves. Self-regulated learning includes setting goals, organizing the study environment, choosing strategies, managing time, seeking help, checking progress and adjusting when needed.
The researchers discussed the results using control-value theory, which considers how perceived control and the value of a task may relate to study emotions. In the survey, enjoyment was associated with more self-regulated learning, while anxiety and boredom were associated with less.
AI literacy appeared to change the pattern
The associations were weaker among students with higher AI literacy. For students with the most AI know-how, the links between adaptive-feeling AI and anxiety and boredom were less than half as strong. The link with lower enjoyment was no longer statistically significant.
For boredom, the emotional route to weaker self-regulated learning was reported as −0.14 at lower AI literacy and −0.04 at higher AI literacy. At higher literacy, none of the three emotional routes was clearly different from zero in the reported analysis.
However, the study measured AI literacy rather than teaching it. It therefore cannot show that an AI-literacy course would create the same buffering effect.
What students and educators can take from it
The results suggest that convenience and personalization may not be the only things to consider when choosing or designing AI learning tools. Students may also benefit from understanding why a system recommends a task, checking whether its suggestions fit their goals and retaining some responsibility for planning and pacing their work.
The researchers recommended AI-literacy courses and platforms that explain their recommendations and allow students to adjust them. Those approaches still need to be tested in trials, so they should be viewed as proposals rather than proven solutions.
Other research cited in the source has found benefits from intelligent tutoring systems, while separate work has reported that unrestricted access to a chat assistant may create problems after access is removed. These findings point to a more complicated picture: AI support may help learning in some settings, while the way support is delivered could affect students’ independence and emotions.
Important limits of the evidence
- The survey established time order but did not prove cause and effect.
- Emotions and study habits were not measured at the start, so students who were already anxious or disorganized may have rated their tools differently.
- All measures were self-reported, and each measure was taken only once at each wave.
- The survey did not include grades, usage logs or observed study behavior.
- The adaptive-environment scale was developed for this study.
- The anxiety measure referred to worry while studying, not an anxiety disorder.
- The participants came from four universities in China, and attitudes toward AI and study norms may differ across cultures.
For now, the study is best read as a warning against assuming that more personalization is automatically better. Adaptive AI may be most useful when it supports students’ decisions without quietly replacing the practice of making them.
AI tools were used to assist with the preparation of this article.
Leave a Reply