Abstract
Purpose
Despite the rapid proliferation of empirical studies on generative artificial intelligence (GenAI) in second language (L2) writing, a systematic review of how GenAI supports L2 writing learning remains scarce. This study aims to fill this gap by providing a comprehensive synthesis of existing research.
Design/Approach/Method
We conducted a systematic review of 55 pertinent articles published in SSCI journals between January 2023 and July 2025, employing Activity Theory as our analytical framework. The review systematically examined six key components within the activity system: subjects (participant characteristics), tools (research instruments), community (research contexts), division of labor (social roles and power distribution), rules (procedural norms), and objects (learning goals).
Findings
Our analysis revealed that tools, division of labor, and rules serve as mediating components that exert both positive and negative effects on learning objects (i.e., the goals of learning activities). The framework effectively captured the complex interactions among different elements within GenAI-assisted L2 writing research.
Originality/Value
This study contributes to the literature by providing the first systematic review of GenAI-assisted L2 writing research within a theoretical framework. The Activity Theory perspective offers novel insights into understanding how different components interact in this emerging field.
Introduction
Equipped with generative models that can improvise contextually relevant responses through natural language interactions, generative artificial intelligence (GenAI) technology can understand natural language and address spontaneous queries (Chen, Wei et al., 2025; Lin et al., 2025; Mi et al., 2025; Rong et al., 2025). While researchers have cautioned that GenAI may bring negative impacts on L2 writing, such as over-reliance on the technology and instances of plagiarism (Shi et al., 2025), this technology has been found to facilitate the learning, teaching, and assessment of second language (L2) writing. Specifically, GenAI can serve as a learning partner, supporting the writing process, such as brainstorming ideas, preparing outlines, and proofreading (Guo & Li, 2024; Huang & Wang, 2025). Moreover, GenAI can assist human teachers by providing complementary assessments and delivering individualized feedback on student writing (Yao et al., 2025).
Currently, the number of empirical studies on GenAI-assisted L2 writing instruction is increasing; however, comprehensive reviews of pertinent research remain limited. To our knowledge, there have been three review studies in this regard. Ibrahim and Kirkpatrick (2024) synthesized 42 studies to explore how ChatGPT supported L2 English writing teaching and learning, suggesting that ChatGPT can strengthen students’ writing motivation, enhance teachers’ instructional efficiency, and offer immediate and personalized feedback to learners. Meanwhile, Teng (2024) examined both the affordances and the challenges of using ChatGPT in L2 English writing in a systematic review of 20 pertinent articles. The author highlighted ChatGPT's capacity to cultivate writing skills through human–AI interaction. However, challenges persist concerning ethical issues and the potential erosion of human creativity and critical thinking skills. Likewise, Xiao et al. (2025) conducted a review of 16 studies, exploring ChatGPT's supportive roles as a writing assistant, performing tasks such as generating ideas, paraphrasing, or summarizing across different writing phases. It also serves as an assessment tool by offering error detection and grammar checking.
While the three aforementioned studies provide valuable insights, they focused on the role of ChatGPT in L2 writing, and thus, the wider landscape of GenAI tools, ranging from self-made chatbots to emerging large-language-model variants, remains unmapped. Besides, the three studies primarily review users’ perceptions toward the tool, while a comprehensive synthesis of how GenAI in general shapes L2 writing learning is still lacking. Finally, there is a paucity of a theoretical framework in the review of empirical research on GenAI-assisted L2 writing. This deficiency impedes the systematic evaluation and integration of findings across studies. A systematic review of existing literature from a theoretical perspective is needed.
Activity theory (AT) provides a perspective for examining how humans and technology interact within collective activities (Lin et al., 2019). This framework serves as a valuable lens not only for guiding empirical studies on technology-supported teaching and learning, but also for reviewing existing empirical research in this domain, as illustrated in the following section.
Activity Theory and Its Application in Literature Review
Originated from Vygotsky's (1978) tripartite model, AT comprised subjects, objects, and tools as the core components of psychological development, positing that desired outcomes (i.e., objects) are achieved by individuals (i.e., subjects) through the mediation of tangible or intangible tools (Cho, 2017). The subsequent generation of AT developed by Leont’ev (1981) embraced the collective nature of human activity that is driven by shared motives, thereby highlighting the interplay between subjects and their community. Engeström (2015) enriched the notion of collectivity by integrating additional components, that is, the rules and division of labor, which concurrently facilitate the achievement of desired outcomes within a social context. Simply put, the subjects, in conjunction with the community, are motivated by objects through the use of tools. This process is constrained by rules and accompanied by a task distribution between community members (i.e., division of labor) (Engeström, 2001).
Currently, there have been a few review studies about technology-assisted language learning from the lens of AT (Lin et al., 2019; Zhang et al., 2023). For instance, Zhang et al. (2023), in reviewing 40 studies, utilized AT to clarify detailed components of technology-enhanced peer feedback and revealed how various factors interact with each other and mediate the activity effectiveness. This study showed that the dynamic process of interactions between learner factors (subjects), learners’ roles in peer interactions (division of labor), the environment of peer interactions (community), technical issues (tools), and mechanisms (rules) may mediate peer feedback activity outcomes (objects).
The conceptualization of AT components in the existing review (Lin et al., 2019; Zhang et al., 2023) establishes foundations for the employment of AT in reviewing empirical studies on technology-assisted language learning (Figure 1). Specifically, subjects denote the participants in the learning activity; objects indicate the goals of the learning activity; tools are the instruments supporting language learning; community refers to the condition in which the learning activity is implemented; rules indicate the norms of the learning activity; and division of labor pertains to the duties of the participants in the learning activity.
Activity theory framework.
The analysis of existing literature within the framework of AT contributes to a systematic review of the various aspects of research, as well as their mediated relationships, on the application of GenAI in L2 writing education. In this review, we specifically focus on how various contextual dynamics within the GenAI-mediated L2 writing activity system shape learners’ L2 writing learning experiences and outcomes. Two research questions (RQs) guide the whole review:
Method
Data Collection
The review adopted a systematic content analysis to identify and evaluate relevant literature (Lin et al., 2019). The search period spanned from January 2023 to July 2025, when GenAI was experiencing significant growth and increasingly influencing higher education. Given the large number of pertinent research articles on the application of GenAI in L2 writing, we exclusively selected Social Sciences Citation Index (SSCI) articles from the Web of Science Core Collection, which typically presented high-quality research. Our search was filtered for “article” entries written in “English.” We developed three sets of keywords based on the previous literature reviews (Teng, 2024; Xiao et al., 2025) and combined them with the AND operators. These keyword groups included technology terms (e.g., GenAI, GAI, ChatGPT, or AI chatbot), writing terms (e.g., writing, essay, or composition), and language learning terms (e.g., second language, foreign language, EFL, or ESL).
The data screening criteria are as follows. First, the selected articles were restricted to empirical studies, since these studies report original findings on GenAI use in L2 writing. Second, only studies of L2 writing were considered, excluding those on L1 writing or other educational settings. Third, articles that use automated writing evaluation or conventional chatbots (rather than GenAI) were excluded. Fourth, included studies should examine the enhancement of L2 writing learning and specify the process of how students utilize GenAI.
The data collection procedure (Figure 2) shows that the initial search yielded 255 papers. In the second round, after reviewing the titles and the abstracts, we excluded articles based on research domain, instruments, and methodology, leaving 114 articles. The third round involved a full-text review to further assess eligibility based on the inclusion criteria. Additionally, two more articles were added by cross-referencing. A final selection of 55 papers was selected for the review.
Data collection procedure.
To maintain consistency for inclusion and exclusion of papers, the two authors organized three rounds of discussions concerning article identification and screening. In cases of uncertainty during the exclusion phase, both authors examined the full texts of relevant articles and engaged in a dialogue to determine whether to exclude or include them.
Data Analysis
Data analysis was conducted both deductively within the framework of AT and inductively, taking the following steps. We first thoroughly read the collected literature for a comprehensive understanding of the data, followed by the generation of initial codes. Subsequently, we reviewed all initial codes and organized them into categories that corresponded to the six components of AT (Table 1). For example, an AI chatbot, Reflectium, was first coded as the sub-theme of “fine-tuned” and then the theme of “usage techniques,” which fell into the category of “tools.”
Coding Scheme for the Review From the Lens of AT.
The two authors collaboratively examined ten articles to establish a consensus on the coding scheme. Subsequently, each researcher individually analyzed the remaining articles, adhering to the agreed-upon coding method. Upon identifying additional sub-themes that fell outside the existing scheme, we expanded the coding scheme by incorporating these new sub-themes. The coding outcome revealed satisfactory reliability between the authors (inter-coder reliability coefficient = .91). Any remaining discrepancies were resolved through discussion.
Based on the coding results, we conceptualize the six components of the activity system in GenAI-assisted L2 writing research (Figure 3). Subjects refer to the characteristics of participants in the studies (e.g., educational level); objects denote the research goals (e.g., examination of the impacts of GenAI on L2 writing performance); tools refer to the instruments used (e.g., various types of GenAI instruments and usage methods); community is the research condition, including the duration of the research and the research environment; division of labor pertains to distribution of the social roles and power among participants (e.g., role allocation); and rules refer to the norms that regulate the research procedure such as training on GenAI usage.
Activity theory framework in this review.
Components and Mediating Effects in GenAI-Assisted L2 Writing Research
This section initially examines the six components within the activity system of GenAI-assisted L2 writing research (RQ1). Subsequently, it explores the mediating effects that emerge in this activity system, aiming to elucidate how different components mediate the objects of L2 writing learning (RQ2).
Subjects
Subjects refer to characteristics of participants in GenAI-assisted L2 writing studies, including the regions, educational level, and number of participants. As shown in Figure 4, the majority of studies focus on undergraduate-level education (42 studies), while a smaller portion examines elementary- (one study), secondary- (eight studies), postgraduate-level education (two studies), and mixed levels (two studies). Regarding geographic distribution, 49 studies were conducted in Asia, with China representing the largest share (34 studies). Meanwhile, six studies were conducted in America, Europe, and Africa. Overall, existing research has predominantly focused on undergraduate students in Asian contexts. Notably, significant differences exist in socio-cultural norms, educational practices, and institutional environments across regions, which may influence technology acceptance and integration. Future studies could incorporate more diverse geographical and cultural settings to provide additional, complementary empirical evidence for the application of GenAI in L2 writing instruction.
Subjects: participant characteristics.
Concerning the number of participants, 12 studies involved fewer than 30 participants, 32 studies included 30 to 100 participants, while 12 studies encompassed over 100 participants. Studies with fewer than 30 participants used qualitative methods with multiple data types, such as interviews, log data, and reflections (e.g., Chen, Huang, Lai et al., 2025; Fathi & Rahimi, 2026). Quantitative methods were predominantly conducted in studies involving more than 30 participants. It is crucial to recognize that the number of participants does not solely determine the quality of a study. The key to robust research findings lies in selecting methodologies and data types that are appropriate for the sample size and align with research norms (Creswell & Creswell, 2023). Future research should be mindful of these considerations to ensure methodological rigor and validity.
Tools
In this review, tools refer to the instruments used in the studies, including various types of GenAI instruments and their usage techniques (Figure 5). ChatGPT was the most frequently investigated tool (34 studies; e.g., Lee, 2024). Meanwhile, 16 studies involved GenAI tools developed by the researchers, such as Reflectium (Neshaei et al., 2025), and five studies employed other commercial GenAI tools, such as Gemini and Poe (Wu & Xu, 2025). The rapid development of GenAI presents exciting opportunities for future research. The latest version of ChatGPT, for instance, could be utilized in upcoming studies to explore its enhanced capabilities. Other GenAI tools and self-made GenAI tools that are tailored to specific writing purposes could also be considered for future research. The self-made tools were developed for certain purposes, such as offering visualized GenAI feedback (Zou et al., 2025), supporting argumentations (Guo & Li, 2024), online collaborative writing (Hu et al., 2025), or peer assessment (Lin et al., 2025). For example, LogicalHamster, a self-made tool, assisted learners with logic learning (Zhang, Zou, & Cheng, 2025). Incorporating these diverse tools could enrich the existing literature and provide a more comprehensive understanding of their applications in L2 writing studies (Rong et al., 2025).
Tools: instrument characteristics.
Regarding the usage technique, methods such as fine-tuning or zero-shot prompting were employed (Wang & Gayed, 2026). Fine-tuning refers to the process of further training a pre-trained model on a specific dataset or for a particular task. Zero-shot prompting allows the model to generate responses without any prior examples or contexts that are specific to the task in the prompt (Wang & Gayed, 2026). In the existing research, zero-shot prompting was the most common method (39 studies), followed by the fine-tuning method (16 studies).
Several key design and functional principles have been identified for the effective fine-tuning of GenAI writing tools. First, the interface should adopt a supportive style by offering praise and encouragement (Zou et al., 2024), include space for reflection (Kim et al., 2025), and be integrated with other supportive tools (Hu et al., 2025). Second, the tuning of GenAI should align with specific writing goals and be grounded in sound theoretical foundations to ensure pedagogical effectiveness (Lin et al., 2025). Functionally, it must support self-regulated or collaborative learning by providing explicit instruction on writing concepts, accompanied by examples and explanations, as well as customized exercises (Neshaei et al., 2025). It should also evaluate learners’ performance and offer adaptive feedback based on individual progress (Zhang, Zou, & Cheng, 2025).
Community
In this review, community refers to the research conditions, including the duration of research and research environments of GenAI-assisted L2 writing (Figure 6). All studies adopted a longitudinal design, with pre-, post-, and/or delayed post-tests. Among these studies, 24 studies spanned from one to four weeks, 16 studies lasted between five and nine weeks, and only twelve studies extended beyond 10 weeks. Generally, studies with a longer time span could allow for a deeper exploration of how GenAI tools facilitate the development of students’ writing skills over time. For example, when students are given abundant time to document the changes, analyze GenAI feedback, and resort to other available sources for clarification, they may more effectively incorporate GenAI feedback into their revisions (Asadi et al., 2025). Conversely, time constraints can lead to students’ heavy cognitive load (Woo, Wang et al., 2024). Considering the small number of studies with time spans longer than 10 weeks, future research could extend the duration of studies to provide more valuable insights into the sustained impact of GenAI tools on students’ learning.
Community: research conditions.
The research environments wherein GenAI tools were employed serve as a key contextual factor. A total of 46 studies were conducted in naturalistic classroom environments, while 9 studies were conducted under controlled conditions. The predominance of class-based studies highlights the practical application of GenAI tools in real educational settings, providing insights into their effectiveness in everyday teaching and learning scenarios. However, the smaller number of controlled studies suggests a need for more rigorous investigations to isolate specific variables and better understand the mechanisms through which GenAI tools impact L2 writing.
Division of Labor
Division of labor in this review pertains to the distribution of social roles and power among participants in existing research. This includes two main aspects: whether the participants used AI tools individually or collaboratively, and who was responsible for creating the prompts. As depicted in Figure 7, 43 studies involved individual tool use by students, while only 12 studies explored collaborative use. This stark imbalance suggests that the field has yet to fully explore how collaborative GenAI use reconfigures teacher–student and peer–peer interactions. Future work could therefore examine such joint engagement with GenAI tools in L2 writing learning.
Division of labor: distribution of social roles and power.
Regarding prompt making, students were primarily responsible for making prompts in 39 studies examining their inherent ability to utilize GenAI and their activities in this process (e.g., Fathi & Rahimi, 2026). Meanwhile, in 16 studies, researchers and instructors provided the designated prompts for the participants, particularly in studies where GenAI tools were used as feedback providers (e.g., Alsofyani & Barzanji, 2025) or as learning tools (e.g., Woo, Guo et al., 2024; Woo, Wang et al., 2024). Clearly, allowing students to generate their own prompts mirrors real-world usage and nurtures their independent engagement with GenAI, which explains the predominance of this design. Meanwhile, when researchers or instructors supply the prompts, it is often a deliberate methodological choice to ensure that the GenAI's feedback meets the specific quality and focus required by the study. Each approach thus carries distinct merits, and the choice between them should align with the researcher's overarching purpose.
Rules
Rules refer to the norms that regulate the GenAI-assisted L2 writing research, including writing genres, submission requirements, and training on the usage of GenAI (Figure 8). Concerning writing genres, the majority of writing tasks focused on essays, including expository or argumentative writing (41 studies). Only three studies concentrated on academic paper writing (i.e., Ou et al., 2025), while eleven studies explored stories or other genres. The focus on expository and argumentative essays is consistent with the academic demands faced by college students, who constituted the most predominant sample in the literature. However, the limited exploration of other writing genres suggests an opportunity for further investigation. Since GenAI has increasingly influenced academic writing (Nguyen et al., 2025) and may have great potential in supporting creative writing (Woo, Guo et al., 2024), expanding research to include a broader range of writing genres could provide valuable insights into the diverse applications of GenAI tools in L2 writing. This broader focus would better reflect the varied writing tasks students encounter in different levels of education and could enhance the applicability of GenAI tools across diverse academic contexts.
Rules: norms that regulate the GenAI-assisted L2 writing studies.
Submission requirements pertain to the number of writing tasks assigned and the number of submitted drafts in the reviewed studies. Among them, 26 studies required only one writing task, while 29 required between two and nine tasks. Meanwhile, all studies limited submissions to a maximum of two drafts: an initial draft and a revised one. Given that multi-draft composing has become a well-established practice in L2 writing learning (Zhang & Hyland, 2022), the potential influence of multi-draft practices that involve three or more rounds of submission on GenAI-assisted writing warrants further investigation.
Training on the usage of GenAI involves providing instructions on how to effectively utilize these tools in L2 writing. In 45 studies, the researchers provided explicit training concerning three aspects: GenAI technology use (e.g., Huang & Mizumoto, 2024), integration of GenAI with other pedagogical elements, such as writing genres and feedback approaches (e.g., Woo, Guo et al., 2024), and implementation of innovative teaching modes that integrate GenAI tools into the writing learning process (e.g., Li et al., 2025; Liu et al., 2024).
The training on technology use in the reviewed studies followed a structured, step-by-step approach. Teachers began by clearly defining the objectives of the writing assignment to determine when and how ChatGPT could be appropriately used (Yin & Dou, 2025). Instruction was delivered through instructional videos or live demonstrations, allowing students to observe effective GenAI interactions in real time. Following the demonstration, students formulated prompts, asked follow-up questions, and critically evaluated and integrated GenAI-generated text into their own writing (Huang & Wang, 2025). They were taught to use GenAI for specific writing tasks such as brainstorming, generating ideas, and improving argument structure (AL Fraidan, 2025), as well as for receiving feedback on structure, grammar, and vocabulary (Deygers et al., 2025).
Additionally, training included instruction on multiple writing genres and feedback approaches. For instance, when students received explicit instruction on argumentation (Zhang, Zou, & Cheng, 2025), the focus was on key elements such as claims, evidence, backing, and rebuttals, as well as how these elements can be addressed using GenAI (Tai et al., 2025; Zare, Al-Issa et al., 2025). In problem-solution tasks, students drew on personal experiences (e.g., public transportation) and used GenAI to generate or refine solutions (Meşe et al., 2025). Concerning the training on GenAI feedback, key features of effective peer feedback (Lin & Hwang, 2025) or self-feedback (Mahapatra, 2024) were commonly included. Students were trained to provide feedback on writing based on the targeted rubric or analyze the genre features of a written text by using GenAI.
Implementation of innovative teaching modes was grounded in established or newly developed theoretical frameworks and pedagogies. For example, based on a conceptualization of Critical GAI Literacy, Ou et al. (2025) developed a self-regulated learning-based micro-curriculum to foster L2 doctoral students’ competencies to use GenAI for academic writing. In another study, Liu et al. (2024) combined GenAI with the Cognitive Academic Language Learning Approach to enhance elementary students’ L2 writing performance, self-regulated learning strategies, and writing motivation.
Notably, only six addressed the issue of plagiarism, despite it being a major concern in GenAI-assisted learning (Cotton et al., 2024; Yan, 2023). This oversight may be attributed to the inconsistent views of plagiarism in the new era, which blur the boundaries between original work and AI-generated content (Yan, 2023). Given the significant implications of GenAI for academic integrity, it is essential to equip students with the skills to navigate these tools responsibly and effectively.
For studies that included training on avoiding plagiarism, the strategies were designed to promote academic integrity and critical awareness. Students were clearly informed that GenAI could not be used to produce the final essay, and any misuse would result in penalties, including an automatic zero (AL Fraidan, 2025). This rule was enforced through in-session observations and debriefs by the teacher (Zare, Al-Issa et al., 2025). Students discussed examples highlighting the potential biases inherent in GenAI tools and were required to rewrite GenAI-suggested revisions in their own words (Tai et al., 2025). To further foster critical thinking and prevent uncritical reliance on the tool, students submitted their responses and revisions to GenAI along with revision reports to document and justify the changes they made (Tsai et al., 2024) and participated in collaborative interpretation of GenAI feedback (Shi et al., 2025).
Objects
In this review, objects refer to the research goals of the studies. Our analysis identified three principal research goals in GenAI-assisted L2 writing research that align with the objects of the activity system. As shown in Figure 9, 44 empirical studies examined the impacts of GenAI tools on L2 writing performance, while 38 studies focused on the impacts on students’ psychological responses. Only nine studies investigated the impacts on behavioral responses.
Objects: Research goals of the reviewed studies.
Impacts of GenAI on L2 Writing Performance
The impacts of GenAI on writing performance were predominantly positive, as evidenced by 23 studies that performed within- and/or between-group analysis. Generally, students engaged in GenAI-supported L2 writing learning demonstrated better writing quality than those who did not (e.g., Liu et al., 2024; Song & Song, 2023). However, the literature also presents inconsistent findings. Two studies found no significant difference between the experiment and control groups (Alsofyani & Barzanji, 2025; Escalante et al., 2023), while fifteen reported mixed effects (e.g., Hwang et al., 2024; Niloy et al., 2023). For example, Hwang et al. (2024) found that ChatGPT primarily enhances students’ writing in surface-level aspects, with minimal enhancement in higher-order elements, such as content, organization, and cohesion.
One possible reason for the inconsistent findings is that most studies evaluate writing progress based on changes in scores between the pre- and the post-test. Such evaluation may lead to disproportionate improvement, with the significant score increase observed only among students who had lower original scores (Tsai et al., 2024). This highlights the necessity for developing new methods to evaluate students’ writing performance. One possible solution is to examine the impact on specific dimensions, such as critical thinking in argumentation (Chen, Huang, Lai et al., 2025) or creativity in story writing (Niloy et al., 2023). Another solution is to assess students’ progress on multiple writing tasks, which may provide a more reliable evaluation of students’ writing competence (Yan, 2024).
Notably, among these 44 studies, only 15 investigated the new writing gains, which means students wrote a new essay without GenAI support during the post-test or the delayed post-test. These studies explored how GenAI can facilitate the durable acquisition of writing skills by assessing whether any improvements observed during GenAI-assisted drafting persist once the scaffolding is removed. Of the 15 studies, three reported no significant improvement (e.g., Alsofyani & Barzanji, 2025). Furthermore, among the five studies that included a delayed post-test, two found that the initial gains had vanished by the later assessment, even though the immediate post-test had been encouraging (Deygers et al., 2025; Yan, 2024). Since GenAI should be used to cultivate durable writing competence (Yan, 2024), we need to determine the specific conditions under which it actually promotes writing development and how long those gains persist once the support is withdrawn.
Impacts of GenAI on Psychological Responses
Concerning the impacts of GenAI on participants’ psychological responses, existing studies have focused on students’ motivational factors (e.g., ideal L2 self) (14 studies), cognition (e.g., cognitive load and metacognitive awareness) (six studies), and perceptions (18 studies). Regarding motivation, ten studies unanimously reported that GenAI can enhance students’ motivation to study L2 writing. However, three studies reported that GenAI-assisted L2 writing may have mixed or insignificant effects on writing motivation. For instance, Zare, Al-Issa et al. (2025) found that interaction with GenAI initially enhanced students’ motivation to study writing; however, this increased motivation declined during the delayed post-test. Moreover, students’ ideal L2 writing self in the ChatGPT group was significantly lower than that in the automated evaluation writing group (Shi et al., 2025).
Only five studies have investigated students’ cognition when using GenAI from various perspectives, including cognitive load (i.e., Woo, Wang et al., 2024) and metacognitive awareness (Lee, 2024). Mixed findings existed as well. While three studies reported reduced cognitive load even at the one-month follow-up (i.e., Chen, Huang, Ye et al., 2025), one study reported high cognitive load (Woo, Wang et al., 2024). Moreover, metacognitive awareness improved in groups that received teacher and peer scaffolding after using GenAI (Lee, 2024).
Finally, 19 studies reported that GenAI use leads to mixed perceptions among students. On one hand, students held positive attitudes toward GenAI since the tool assisted them in improving their writing (Fathi & Rahimi, 2026; Huang & Wang, 2025). On the other hand, they encountered challenges in creating effective GenAI prompts and comprehending its responses (Duong & Chen, 2025) and worried about the loss of critical thinking and creativity after using GenAI (Tai et al., 2025). These findings highlight that GenAI should only serve as a “writing collaborator” and students must retain executive control over every decision (Huang & Wang, 2025, p. 1).
Impacts of GenAI on Behavioral Responses
The impacts of GenAI on behavioral responses (e.g., peer feedback provision and student engagement) have been examined in eight studies. Three studies examined peer feedback activity. As feedback receivers, students who used GenAI to write essays tended to receive less useful peer feedback (Misiejuk et al., 2025). As feedback providers, students provided high-quality feedback on their peers’ writings with the assistance of GenAI (Guo et al., 2024).
The remaining five studies tracked shifts in students’ behavioral, cognitive, and emotional engagement after GenAI usage. Although one study found that behavioral engagement with ChatGPT feedback was lower than with teacher feedback (Zou et al., 2025), the other four demonstrated heightened engagement when students actively employed various GenAI tools to assist their writing (e.g., Hu et al., 2025; Zare et al., Zare, Ranjbaran Madiseh et al., 2025). Although these studies have captured notable changes in engagement in GenAI-assisted writing learning, they have overlooked interaction, that is, the dialogic exchanges between students and the GenAI system, as a distinct facet of engagement. A recent investigation has shown that such student–GenAI interaction constitutes a unique dimension of engagement in L2 writing (Rong et al., 2025). Future research should therefore trace how behavioral, cognitive, and emotional engagement evolve throughout students’ interactions with GenAI.
Mediating Effects on Writing Learning
Tools as Meditators
In the reviewed studies, tools (i.e., the types and usage techniques) affect the objects of GenAI-assisted L2 writing research (i.e., the impacts on writing performance, psychological responses, and behavioral responses). GenAI has been found to both positively and negatively influence student learning. On the positive side, GenAI offers several advantages, such as promoting active human-AI interactions and helping students focus on linguistic features. These benefits have contributed to students’ favorable perceptions of the GenAI feedback, as well as improved writing skills and motivation (e.g., Nguyen et al., 2025; Song & Song, 2023), particularly in logic learning (Zhang, Zou, & Cheng, 2025; Zhang, Zou, Cheng et al., 2025). The tool's instant feedback enabled students to independently manage their progress, reducing anxiety by decreasing reliance on direct teacher intervention (Wu & Xu, 2025) and enhancing students’ engagement with English writing (Zare, Ranjbaran Madiseh et al., 2025). Besides, GenAI can improve the quality of peer feedback and develop the writing abilities of those who provide such feedback (Guo et al., 2024).
However, GenAI's tendency to foster over-reliance among users may diminish students’ critical thinking and creative writing ability (Niloy et al., 2023). This over-reliance can result in marginal improvements in students’ writing skills (Alsofyani & Barzanji, 2025). Additionally, the tools’ potential inaccuracy and ethical challenges may raise concerns among students and teachers (Asadi et al., 2025; Huang & Mizumoto, 2024). Lastly, feedback overload, characterized by excessive, general, and repetitive comments, may lead to students’ rejection of GenAI feedback (Duong & Chen, 2025).
Notably, self-made tools with fine-tuning technique played a crucial role in students’ writing performance and student engagement. These tools have shown great potential in enhancing writing motivation and facilitating peer feedback (Guo & Li, 2024; Guo et al., 2024). They are also beneficial for reflective writing (Neshaei et al., 2025), increasing student engagement (Hu et al., 2025), and improving writing logic, flow, and self-efficacy in argumentative writing (Zhang, Zou, & Cheng, 2025; Zhang, Zou, Cheng et al., 2025). Furthermore, self-made tools have been found to contribute to academic self-efficacy and writing development (Kim et al., 2025; Wu & Xu, 2025).
In sum, the existing research underscores the dual nature of GenAI in L2 writing contexts. On one hand, these tools enhance writing skills, motivation, and independent learning. On the other hand, they risk fostering over-reliance, diminishing critical thinking, and causing feedback overload. Meanwhile, fine-tuned tools, tailored for specific writing tasks, demonstrate considerable promise in addressing these challenges; however, issues of accuracy, consistency, and ethical considerations still require further exploration to maximize their effectiveness and reliability in educational settings.
Division of Labor as Meditators
Division of labor (i.e., individual or collaborative use of GenAI tools and prompt making) influences writing learning outcomes. The collaborative use of GenAI encouraged teachers to provide helpful feedback, thereby encouraging students to incorporate more feedback into their revisions (Yan, 2024). Likewise, when GenAI was used alongside teacher and peer scaffolding, it enhanced students’ metacognitive awareness and academic writing skills (Lee, 2024). Hu et al. (2025) also emphasize the importance of integrating collaborative learning to foster deep cognitive engagement.
Prompt making also impacts writing learning. While student-developed prompts produced varying or even negative outcomes (Hwang et al., 2024; Niloy et al., 2023), students supported by pre-developed prompts generally demonstrated in-depth perceptions, enhanced writing performance, positive attitudes, and increased metacognitive awareness (Lee, 2024; Li et al., 2025).
It is noteworthy that the teacher-led prompting method may also fail to ensure positive results (Alsofyani & Barzanji, 2025; Escalante et al., 2023). This approach may hinder the development of students’ critical thinking when interacting with the tool and utilizing the tool-generated feedback for revisions, subsequently leading to the loss of creativity and agency (Shi et al., 2025). For instance, in a study where teachers designed the prompts to provide personalized feedback via emails, the GenAI feedback groups did not show better writing improvements because students lacked the opportunity to ask follow-up questions (Escalante et al., 2023). In contrast, prompts created by students and teachers together resulted in more favorable outcomes (Yan, 2024).
In brief, the collaborative and integrated use of GenAI tools is more likely to yield positive writing outcomes. This analysis highlights the concept of “distributed agency” (Godwin-Jones, 2024, p. 16), which advocates for a partnership that distributes roles among teachers, students, and GenAI. Future research could investigate how distributed agency, particularly students’ agency, can be more effectively implemented to facilitate the independent development of various prompt engineering techniques and enhance the utilization of GenAI-generated prompts.
Rules as Meditators
Rules (e.g., training on GenAI usage) influence writing learning both positively and negatively. As depicted in “Rules” section, training exhibited three forms: training on technology use, integration of technology with other teaching elements, and implementation of innovative teaching modes. On the positive side, with training on GenAI technology use, students demonstrated a more positive L2 motivational self-system, particularly in terms of their ideal L2 self and learning experience (Huang & Mizumoto, 2024). Besides, researchers also provided training on how to integrate GenAI with various pedagogical elements, including writing genres and feedback approaches. This guidance promotes idea generation, enhances connection and coherence, and improves grammatical accuracy in writing (AL Fraidan, 2025; Woo, Guo et al., 2024). Additionally, some researchers provided training on innovative teaching modes that incorporated corpus-based instruction or provided visualized feedback (e.g., Zou et al., 2025). These interventions not only enhanced students’ writing performance, self-regulated learning strategies, writing attitude, and motivation in the post-test (Li et al., 2025; Liu et al., 2024), but also cultivated their GenAI literacy in L2 writing (Ou et al., 2025).
On the negative side, during training sessions, merely providing students with exemplar prompts may be insufficient. This approach may hinder the development of students’ critical thinking when interacting with the tool and utilizing the tool-generated feedback for revisions, subsequently leading to the loss of creativity and agency (Shi et al., 2025). Meanwhile, detailed training on technology, alongside insufficient emphasis on writing skills, may lead to a heavy cognitive load for students. For instance, Woo, Wang et al. (2024) introduced secondary students to a series of elements, such as GenAI chatbots, prompt engineering, and the think-aloud protocol within one hour and 45 min. Consequently, the students reported a relatively high-level cognitive load, possibly due to their struggles with managing essential writing phases such as planning, drafting, and reviewing, which were not adequately addressed during the training sessions.
The analysis above highlights the crucial role of teachers in providing effective training for students. Teachers should not only introduce students to GenAI but also guide their effective use of this technology. Moreover, teachers should balance training to equally prioritize technological literacy and writing proficiency. Providing such training may effectively improve students’ usage experiences of GenAI and learning performance (Al Fraidan, 2025). Notably, there is a lack of comprehensive studies examining the impact of training on plagiarism in the context of students’ writing development. Given the widely-acknowledged concerns of academic integrity in GenAI-assisted L2 writing (Asadi et al., 2025), future research may explore this area by investigating how different training approaches influence students’ understanding of academic integrity and their ability to produce original work.
Conclusion and Pedagogical Implications
This paper reviews the existing empirical studies on GenAI-assisted L2 writing through the lens of AT. The review identified the six components of the activity system, namely, objects, subjects, tools, community, division of labor, and rules, in the literature on GenAI-supported L2 writing. The analysis also highlights the mediating roles of tools, division of labor, and rules in shaping the objects within this context. The employment of the AT framework of literature review offers a comprehensive lens for analyzing the complex interactions between human and technology, allowing for a nuanced understanding of the roles and dynamics within GenAI-assisted L2 writing research. Currently, research on the application of GenAI in L2 writing instruction is rapidly evolving, with new empirical studies emerging monthly. This dynamic growth underscores the vibrant development of this area. As such, staying abreast of the latest developments is crucial for researchers and teachers alike. The review may serve as a reference point for future empirical studies on the application of GenAI in L2 writing education.
Based on the findings in this review, we also propose the following pedagogical implications. First, with the assistance of professionals, teachers may consider designing task-specific, fine-tuned GenAI tools, because these tools have been found to produce larger and more durable gains in writing development (Neshaei et al., 2025). Second, it is essential to embed GenAI within a staged, dialogic protocol, as evidence reveals that collaborative prompting produces stronger and longer-lasting gains (Hu et al., 2025; Yan, 2024). A workable sequence may begin with the teacher modeling criteria-rich prompts, move to paired students who jointly interpret AI feedback, and end with a whole-class debrief that reconnects suggestions to disciplinary norms, thereby cultivating metacognitive awareness and curbing passive over-reliance. Third, teachers can introduce the concept of AI biases and train students’ paraphrasing skills with illustrative examples, interactive activities, discussions, and reflective exercises (AL Fraidan, 2025; Song & Song, 2023). This not only helps students avoid the academic-integrity risks triggered by GenAI but also drives them to exercise creativity and critical thinking through rewriting, questioning, and reflection.
Footnotes
Author Contributions
Mi Rong contributed to the conceptualization of the study, acquisition of data, the analysis and interpretation of data, and was instrumental in drafting and editing the manuscript. Yuan Yao was responsible for the conceptualization of the study and played a pivotal role in revising the manuscript for intellectual content.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Foreign Language Research Joint Project of Hunan Provincial Social Science Foundation (Grant No. 25WLH02).
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.