1　Introduction

This section elaborates on the background and significance of Generative Artificial Intelligence (GenAI)-assisted second language (L2) writing research. As a core productive skill, L2 writing faces challenges such as delayed feedback and insufficient personalization. GenAI offers new pathways to address these difficulties through its multiple functions, including immediate feedback and content ideation. Although existing studies have confirmed its positive effects, most focus on efficacy validation, with limited systematic discussion on embedding mechanisms and pedagogical structures. International reviews have formed relatively mature knowledge maps, but they are primarily based on English-language journals, making it difficult to capture local practices in the Chinese academic community. Domestic existing reviews either cover overly broad scopes or rely mainly on international journal samples, lacking systematic analyses focused on CSSCI achievements with L2 writing at the core. In view of this, this study takes CSSCI journal papers as samples and conducts analysis from five dimensions: topic distribution, research design, AI roles, theoretical application, and Chinese context characteristics, aiming to reveal the topic structure and research orientation in the Chinese context, and to answer four research questions regarding publication trends, role positioning, theoretical features, and localization performance, thereby providing structural reflection for subsequent research.

2　Literature Collection and Coding

This section describes the literature collection and coding methods for GenAI-assisted L2 writing research. Literature retrieval used CSSCI source journals from CNKI as the data source, with the time range set from November 30, 2022 to December 31, 2025. A keyword cross-combination strategy was adopted, and after manual verification, 34 valid papers were finally included. The coding framework comprises five layers: topic focus layer, research design and context layer, tool and interaction characteristics layer, theoretical application layer, and Chinese context layer, covering 12 coding dimensions and 21 coding variables. Among these, the topic focus layer and Chinese context layer were supplementary additions based on local research characteristics. Coding was completed jointly by two researchers, with pilot coding consistency reaching 92%. During formal coding, inconsistent items were discussed one by one until consensus was reached.

3　Research Results

This section presents the systematic literature analysis results of domestic GenAI-assisted L2 writing research. Publication trends show that from 2023 to 2025, the number of publications in this field rose rapidly, shifting from sporadic exploration to concentrated output, with results distributed across 17 CSSCI source journals, though no systematic review has yet emerged. Research foci can be clustered into three categories: human-AI interaction processes, writing text quality and teaching effects, and learners' individual psychological experiences. Among these, research on text quality and teaching effects accounts for over half, while human-AI interaction research comprises less than 20%. In terms of research design, empirical research accounts for 76.5%, with the paradigm having shifted from value analysis to empirical validation. Research subjects are highly concentrated in college English writing classrooms, with scarce evidence from other educational stages and language varieties. Method selection presents problem-driven characteristics, with text analysis, quasi-experiments, and case studies as mainstream approaches. Data levels are shifting from outcome data to process data, though interactive data collection rates remain relatively low. Regarding tool usage, ChatGPT dominates, with domestic large language models having low application proportions and mostly serving as substitutes. AI roles are primarily graders/feedback providers, with intervention concentrated in revision and evaluation stages, while insufficient support exists for front-end ideation and generation. Theoretical application levels show that nearly 60% of studies cite theoretical frameworks, mainly writing pedagogy, sociocultural theory, and learning engagement theory, but lacking systematic integration between theories, with research on revising existing theories still absent. In terms of localization attention, studies generally respond to national policy discourse, confront local teaching challenges, and reflect traditional educational concepts such as teacher leadership. Some studies form unique contributions by utilizing characteristics of Chinese as a second language, but deep development of domestic tools and research agenda construction still await advancement.

4　Discussion

This section systematically analyzes the core characteristics and deep limitations of domestic GenAI-assisted L2 writing research. The study finds that this field has completed the paradigm transformation from value analysis to empirical validation, but exhibits significant deficiencies in sample structure, evidence form, theoretical application, and localization pathways. At the sample level, research is highly concentrated in college English classrooms, with limited generalizability across educational stages and language varieties. This pattern stems from the accessibility advantages of higher education settings combined with publication mechanisms, forming common characteristics with international research. Regarding process evidence, interactive data is relatively scarce, with research focusing more on "whether it is effective" rather than "how it works," constraining deep understanding of human-AI collaboration mechanisms. The shift from outcome validation to mechanism explanation becomes an important future direction. In role construction, GenAI is mostly positioned as a feedback tool in post-processing stages, continuing the research tradition of automated writing evaluation systems, with insufficient exploration of its potential as a cognitive partner throughout the entire process. Dynamic role switching and strategy regulation mechanisms await exploration. Theoretical application presents a situation of simultaneous progress and bottlenecks: although certain theoretical consciousness has formed, it mostly remains at the stage of introduction and application, without achieving the leap toward developing theories. The applicable boundaries of sociocultural theory and other human-AI interaction explanations still require clarification. Localization characteristics are reflected in the multiple embeddings of policy context, educational concepts, Chinese language context, and domestic tool ecology, but these factors currently serve mostly as background descriptions without fully transforming into mechanism variables and theoretical resources. Future efforts need to promote local experience from situational description toward mechanism analysis and theoretical construction, to enhance the explanatory power of the Chinese context in academic dialogue.

5　Conclusion

This section summarizes the development trajectory and core characteristics of domestic GenAI-assisted L2 writing research: research has shifted from value discussion to empirical testing, with subjects concentrated on college English learners, focusing on feedback comparison, writing quality, and teaching effects, while human-AI interaction process research remains insufficient. GenAI mostly functions in back-end revision and evaluation of writing, with limited intervention in front-end ideation and middle-stage generation. Theoretical application has strengthened but with little innovative integration. Research simultaneously shows localization attention, but factors such as policy context, educational concepts, language characteristics, and domestic tool ecology have not yet fully entered mechanism analysis. The authors point out that the field is moving from efficacy validation toward mechanism explanation, recommending expansion of basic education and non-English language samples, deepening examination of human-AI collaboration mechanisms through process data, and strengthening theoretical integration of local experience to enhance research explanatory power.