Hello everyone. This is Tanaka from the doctoral program. In this article, I would like to introduce a paper we read at our recent English seminar.
Paper Title: How does AI match students’ proficiency levels to provide support? Student-driven AI-mediated scaffolding in EFL writing instruction
(How does AI match students’ proficiency levels to provide support? Student-driven AI-mediated scaffolding in EFL writing instruction)
Authors: Zhiwei Qi, Yuqing Liu, Wenlin Liu, Wei Xu
Journal: System, Volume 142 (2026)
1. Background and Previous Research
In EFL (English as a Foreign Language) writing, while generative AI enables personalized support, it has been pointed out that using it to directly present completed sentences may hinder learners’ independent thinking and lead to over-reliance on AI. To address these challenges, “scaffolding,” which provides appropriate support based on the learner’s level of understanding, is crucial. This study is based on Vygotsky’s sociocultural theory. In this theory, the “Zone of Proximal Development (ZPD)” refers to the gap between what a learner can achieve independently and what they can achieve with the support of others. Furthermore, Aljaafreh and Lantolf (1994) identified the following three principles of scaffolding:
• Graduated assistance: Moving gradually from implicit hints to explicit explanations.
• Contingent assistance: Providing support when the learner needs it and reducing it once they can solve the problem independently.
• Dialogic activity: Grasping the learner’s level of understanding through dialogue and providing appropriate support.
Based on these theoretical frameworks, the authors developed an AI-mediated scaffolding model (AI-MSM) that allows learners to judge the support they need and adjust the level of AI assistance. This study examines the effects of this model on writing proficiency, learner engagement, and motivation.
Research Questions
This study sets the following four research questions (RQs):
• RQ1: Does AI-MSM improve writing performance?
• RQ2: Does AI-MSM improve learner engagement?
• RQ3: Does AI-MSM influence learning motivation?
• RQ4: How do learners perceive the effectiveness of AI-MSM?
2. Research Methodology
Development of AI-MSM:
The AI-MSM developed in this study is a model that analyzes instructional goals, the learner’s past compositions, and error patterns to provide personalized support tailored to writing activities. During the drafting stage, it supports idea generation and outline creation, while in the revision stage, it supports improvements in vocabulary, sentence structure, and grammar. A feature of this model is that learners can transition from implicit to explicit support as needed. For example, if a learner struggles with revising an English sentence, they first receive hints or questions from the AI, and if they still cannot solve it, they can request specific grammatical explanations or revision examples.
Additionally, the study adopts a “Humans-in-the-Loop” mechanism where instructors monitor dialogue logs between learners and ChatGPT in real-time, intervening if the AI provides inappropriate responses or if the learner becomes stuck.
Experimental Procedure:
A 5-week (10-session) experiment was conducted with 80 undergraduate students majoring in English at a foreign language university in China. Participants were randomly assigned to an experimental group (40 students: using AI-MSM) and a control group (40 students: traditional instruction).
In the first week, only the experimental group learned how to operate ChatGPT, and in the second week, both groups learned the basic structure of writing. From the third to the fifth week, they wrote four English essays; the experimental group received AI-MSM, while the control group received traditional instruction (individual writing and teacher feedback).
Measurement Methods:
• Writing proficiency: Pre- and post-tests using a modified version of IELTS Writing Task 2 (200 words, 30 minutes)
• Engagement: Behavioral, emotional, and cognitive engagement (9 items)
• Motivation: Intrinsic motivation, self-efficacy, and test anxiety (7 items)
• Interviews: Semi-structured interviews with 13 students from the experimental group
Analysis Methods:
For quantitative analysis, Multivariate Analysis of Covariance (MANCOVA) was used, controlling for pre-test scores, and FDR correction using the Benjamini–Hochberg method was applied to account for multiple testing. Qualitative analysis was also conducted on the interview data.
3. Results and Discussion
• Writing performance (RQ1):
There was no significant difference between the groups in the pre-test, but a significant difference in overall writing proficiency was confirmed in the post-test (p < 0.001). In particular, the experimental group showed significantly higher results in three areas: task achievement, coherence and cohesion, and grammatical range and accuracy. On the other hand, no significant difference was found in vocabulary (p = 0.084).
The authors suggest that while sentence structure and grammar have relatively clear rules, vocabulary acquisition requires a deep understanding of context and nuance, making complementary instruction by teachers important.
• Learner engagement (RQ2):
The experimental group showed significantly higher results in all aspects of behavioral, emotional, and cognitive engagement. In particular, the largest effect was confirmed in cognitive engagement (partial η² = 0.139). The authors suggest that the process of learners critically examining AI suggestions and requesting necessary support promoted metacognition and Self-Regulated Learning.
• Learning motivation (RQ3):
Although a significant difference between groups was confirmed in overall motivation (p = 0.042), after FDR correction, no significant differences were found in intrinsic motivation, self-efficacy, or test anxiety. The authors cite learners’ anxiety regarding the accuracy and reliability of AI responses as one of the factors.
• Learner interviews (RQ4):
Five categories were extracted from the interviews with 13 students:
1. Promotion of creativity and cognition: AI presented new perspectives, encouraging idea development and critical thinking.
2. Adaptive learning and graduated support: Hints and explanations provided as needed supported independent problem-solving.
3. Conflict between skill independence and AI reliance: Learners were concerned about AI reliance and were seen adjusting their usage methods themselves.
4. Requests for tool improvement: Improvements were requested regarding the accuracy of detecting unnatural English expressions or expressions unsuitable for the context.
5. Technical limitations and reliability risks: Reports were made regarding misunderstandings or the presentation of uncertain information by the AI.
Role of Humans-in-the-Loop:
In this study, instructors monitored dialogue logs with the AI in real-time and intervened as necessary. The authors state that this teacher support compensated for the technical limitations of the AI and supported the learners’ appropriate problem-solving.
4. Conclusion and Limitations
• Conclusion: It was shown that AI-MSM is effective in improving the writing proficiency (especially sentence structure and grammar) and learner engagement of EFL learners. In particular, it is suggested that the mechanism where learners themselves adjust the level of support may have promoted cognitive engagement. On the other hand, the effects on vocabulary and motivation were limited.
• Limitations: The subjects were limited to 80 English majors in China, the intervention period was short at 5 weeks, and only the teachers of the experimental group received AI training. The authors also state that the reliability of some measurement scales was relatively low, and the results should be interpreted with caution.
Reflections
I chose this paper for the seminar because I was interested in the educational approach proposed, where learners judge and adjust the level of support they seek from generative AI based on their own proficiency, and the fact that the study verifies its effectiveness. The system design of this study, where learners recognize difficulties and seek support as needed, is also relevant to my own research theme, “Self-Regulated Learning.”
This study emphasizes the importance of “Humans-in-the-Loop,” with teachers monitoring the system usage status of learners in real-time and intervening when necessary. However, I wondered whether it is possible to apply this to more autonomous learning, such as study outside of class, and whether the learners’ autonomous learning will continue even when teacher intervention is removed.
Also, the results showed that there was no significant improvement in “vocabulary” scores in learning outcomes using AI-supported assistance, which is also a challenge in my own research. I would like to consider how to improve writing learning support using generative AI for vocabulary improvement in future classroom practice and research design.
