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: What task to offload to GenAI in the writing feedback process? – Effects of task-offloading approaches on EFL learners’ writing skill development
(What tasks should be offloaded to Generative AI in the writing feedback process?
—Effects of task-offloading approaches on EFL learners’ writing skill development)
Authors: Chun Lai, Mengru Pan, Kai Guo, Yaqiong Cui
Journal: Computers & Education, 252, 105675 (2026)

1. Background
• Potential and Challenges of Generative AI Feedback
Generative AI has the potential to provide immediate and personalized feedback in English writing learning. On the other hand, the authors point out that AI output may contain inaccuracies or advice that does not fit the context. Furthermore, they note that if learners accept AI suggestions as they are, they may lose opportunities to evaluate their own writing and think about problems and ways to improve.

• What is Task Offloading?
The “task offloading” focused on in this study refers to learners delegating part of the cognitive work they would normally perform to external tools or environments. Writing involves tasks such as drafting, evaluating, identifying problems, considering improvements, and revising. The authors state that the content and depth of the thinking performed by learners can change depending on which tasks are offloaded to Generative AI. Previous studies have often examined methods where learners write a text and Generative AI provides evaluation and revision suggestions. However, this method results in delegating even the important cognitive and metacognitive task of evaluating the text to the AI. Therefore, this study focuses on “which tasks to offload to Generative AI” during the writing process and examines the impact on writing skills and other factors.

2. Research Objectives
The research objective is to clarify which tasks in the writing feedback process are effective for the development of learners’ writing skills when offloaded to Generative AI. It also examines how different task-offloading approaches lead to differences in learners’ cognitive engagement, metacognition, and self-efficacy.
There are two research questions in this study:
RQ1: How do differences in the tasks offloaded to Generative AI during the writing feedback process affect the development of EFL learners’ writing skills?
RQ2: How do differences in the tasks offloaded to Generative AI lead to differences in learners’ cognitive engagement, metacognition, and self-efficacy?

3. Research Methods
This study was conducted in a writing class for argumentative essays with 101 undergraduate English majors at a university in China. The participants’ English proficiency was at an intermediate level. Participants were divided into the following three conditions:
• Condition 1: Generative AI creates a draft, and learners evaluate and revise those texts.
• Condition 2: Learners write a draft, Generative AI provides feedback, and learners revise.
• Condition 3: Without using Generative AI, learners perform drafting, self-evaluation, and revision.
Data used included pre- and post-writing tests, questionnaires measuring self-efficacy, peer review assignments, conversation logs with Generative AI, and semi-structured interviews. The analysis examined changes in writing skills, as well as interactions with Generative AI, the quality of feedback, and learners’ perceptions.

4. Results and Discussion
4.1 Writing Skills and Cognitive Engagement
The analysis showed that Condition 1 and Condition 2, which used Generative AI, showed higher gains in writing skills than Condition 3, which did not. The highest gains were seen in Condition 1, where learners offloaded drafting to the AI and performed evaluation and revision themselves. Learners in Condition 1 interacted more with the AI than those in Condition 2, engaging in more learning-oriented exchanges to seek explanations, advice, and concrete examples. Interviews also indicated that learners in Condition 1 viewed the AI-generated text not as a finished product, but as learning material to be analyzed and evaluated.
On the other hand, in Condition 2, there was a tendency for learners to ask the AI to revise directly after receiving feedback, rather than revising it themselves. The authors point out the possibility that even if learners could understand the problems, they did not know how to revise them with their own vocabulary and expressive abilities.

4.2 Metacognition and Self-Efficacy
In the peer review assignment, learners in Condition 1 showed particularly high results in their ability to identify problems in texts. The authors interpret this as being due to the experience gained in discovering problems themselves through the activity of repeatedly evaluating AI-generated texts. In the conditions using AI, effects were also seen in the ability to explain the reasons for problems and suggest improvements.
Regarding self-efficacy, Condition 1 also showed the highest results. Learners in Condition 1 recognized that they had the ability to improve their writing because of the experience of improving AI-generated texts based on their own judgment. On the other hand, in Condition 2, interviews revealed anxiety that even if the text was improved, they did not know whether it was due to their own ability or the AI’s support.

4.3 Importance of Tasks Offloaded to Generative AI
The authors state that the reasons for the high effectiveness seen in Condition 1 may be that reading AI-generated texts in detail against a rubric promoted cognitive engagement, that evaluation and revision deepened metacognition regarding writing and the limitations of AI, and that the experience of improving AI texts led to self-efficacy. This paper points out that while Generative AI supports learning, it may replace learners’ thinking depending on how it is used. The authors state that it is important to position Generative AI not as a tool that simply performs revisions, but as a role that learners evaluate, question, and improve.

5. Conclusion and Limitations
This study showed that incorporating Generative AI into the writing process has the potential to promote the development of writing skills. In particular, having Generative AI create a draft and having learners evaluate and revise it was shown to be highly effective in terms of cognitive engagement, metacognition, and self-efficacy. The authors also suggest incorporating activities such as evaluating AI-generated texts using a rubric and examining the validity of AI revision suggestions into classes.
Limitations include that the subjects were limited to English majors at one university in China, that metacognition was measured indirectly through tasks, and that the use of external Generative AI could not be completely controlled.

Reflections
This paper demonstrates the necessity of designing Generative AI as a role that promotes evaluation and metacognition from the perspective of “task offloading.” I think the point that delegating feedback to Generative AI may reduce opportunities for learners to evaluate their own writing and think about problems and solutions is important. On the other hand, activities where teachers present English texts containing errors or peer reviews have been conducted for a long time, so I felt it was somewhat ambiguous why it is necessary to use Generative AI. The necessity of Generative AI taking on this activity was also discussed in the seminar. Also, since this study targets English majors, it is not clear whether similar effects can be obtained for learners with lower proficiency. When implementing tasks like Condition 1, I think it is necessary to prepare clear and easy-to-understand rubrics in addition to generating texts that match the learners’ level.