Методика преподавания языка | Филологический аспект: Методика преподавания языка и литературы Методика преподавания языка и литературы №03 (38) Май 2026 - Июнь 2026

Генеративный ИИ как партнёр по обратной связи при письме: исследование эффективности ChatGPT в обучении письму на втором языке

Сун Чэнчэнь
аспирант первого года обучения, Российский университет дружбы народов (РУДН), г. Москва, Россия; e-mail: 1042254243@rudn.ru

Аннотация: Данное исследование рассматривает возможность использования генеративного искусственного интеллекта, в частности ChatGPT, в качестве вспомогательного инструмента обратной связи в методике обучения письму на втором языке (L2). Опираясь на современные эмпирические данные (2023–2026 гг.), в статье оценивается сравнительная эффективность обратной связи, генерируемой ИИ, по сравнению с традиционной обратной связью преподавателя по различным параметрам качества письма, включая грамматическую точность, лексическое богатство, связность и тематическое развитие. В обзоре обобщены данные квазиэкспериментальных исследований и систематических обзоров, показывающие, что генеративный ИИ обладает существенными преимуществами в предоставлении обратной связи на семантическом уровне, особенно в отношении структурирования дискурса и тематической проработки, в то время как специализированные инструменты автоматизированной оценки письма, как правило, лучше справляются с обнаружением поверхностных синтаксических ошибок. В исследовании также рассматривается модерирующая роль сложности запроса (промпта) в формировании качества предлагаемых ИИ исправлений, а также отношение студентов к процедурам пересмотра текста с поддержкой ИИ. Результаты показывают, что совместная «человеко-машинная» модель обратной связи, в которой ИИ устраняет стандартные языковые ошибки, а преподаватели делают акцент на когнитивных процессах высокого порядка и персонализированной поддержке, представляет собой наиболее продуктивную в педагогическом плане стратегию. В заключении обсуждаются практические рекомендации для подготовки преподавателей языков и предлагается теоретическая рамка включения генеративного ИИ в программы обучения письму на втором языке.
Ключевые слова: генеративный ИИ, ChatGPT, методика обучения письму на втором языке (L2), систематизированная критика, написание текстов, человеко-машинная синергия, языковые навыки

Sun Chengchen
1st year PhD student Peoples’ Friendship University of Russia, Moscow, Russia

Abstract: This research investigates the feasibility of generative artificial intelligence, specifically ChatGPT, as a complementary feedback instrument within second language (L2) composition pedagogy. Utilizing current empirical data (2023–2026), the article evaluates the comparative efficacy of AI-generated feedback compared to conventional instructor feedback across various dimensions of writing excellence, encompassing grammatical precision, lexical richness, coherence, and thematic expansion. The review integrates evidence from quasi-experimental research and systematic reviews showing that generative AI exhibits significant advantages in offering semantic-level feedback, especially regarding discourse structuring and thematic elaboration, while specialized automated writing assessment tools typically perform better in surface-level syntactic error detection. The investigation additionally examines the moderating function of prompt complexity in influencing the quality of AI-generated corrective suggestions, alongside student perspectives on AI-supported revision procedures. Results indicate that a collaborative "human-AI integrated" feedback framework, where AI manages standard linguistic errors while instructors emphasize higher-order cognitive processing and personalized support, constitutes the most instructional productive strategy. The paper concludes with practical implications for language teacher training and suggests a theoretical framework for incorporating generative AI into L2 composition syllabi.
Keywords: AI generation, ChatGPT, L2 composition pedagogy, systemized critique, script crafting, human-machine synergy, linguistic skill.

Generative artificial intelligence (AI) has developed rapidly in recent years, and educational technology has been changed by this development since the early 2020s. After OpenAI in the United States announced the large language model ChatGPT in November 2022, scholars and practitioners around the world have begun to study both the new potential and the problems that arise from high-quality text generation by LLMs. In the field of second language (L2) teaching, the above modifications have been quite successful. Historically, writing pedagogy has been considered one of the most resource-intensive areas in language teaching due to the large amount of feedback that teachers need to provide; now, it is also an important subject of research for integrating generative AI systems into teaching strategies [1, p. 152].

Rising interest in ChatGPT and other generative artificial intelligence systems for language education stems from multiple overlapping reasons. At first, the international spread of English-speaking higher education has raised the number of L2 scholars requiring commentary on scholarly composition to a high level, putting significant pressure on writing labs and individual educators. In previous versions of Automated Writing Evaluation (AWE) instruments, such as Grammarly and Criterion, while they can identify surface-level errors, they have also received criticism for their inability to provide in-depth recommendations on broad writing problems, such as logic, organization, and development of content [2, p. 55]. Finally, the interactive and context-aware character of ChatGPT and its follow-up models, such as GPT-4, suggests that more advanced and customised guidance that resembles the specific features of tutor-learner dialogue is possible [3, p. 4].

Although the main reason for this study is relatively simple, there are also many educational problems: How effective is generative AI as a writing response aid compared with traditional methods, and in what circumstances can it be used to assist or replace human guidance reasonably well? The four elements of this study are: the accuracy and comprehensiveness of the comments produced by AI; how often students read and apply these comments in their drafts; the impact of different prompt types on the quality of feedback; and the broader educational implications of integrating AI into writing classes [4, p. 8].

This paper will assess how well the generative AI tool ChatGPT can be applied to improve the teaching of composition in the second language (L2) through a full-scale study based on current empirical data. The novelty of this work is that it combines research results from 2023 to 2026, and during this period, academic attention to this field has been expanding significantly. Based on the collected data of prompt design, relative feedback utility, students’  perceptions, and teaching methods, this paper has compiled an overview of existing studies to offer language teachers and scholars a single source of information, and has also put forward some directions for further research. The organization of this paper is as follows: Section 2 reviews the existing research on generative artificial intelligence in the field of linguistics education; Section 3 conducts a comparative analysis of outputs from artificial intelligence and those provided by humans; Section 4 explores the role of prompt engineering; Section 5 proposes a human-AI collaborative model; and finally, Section 6 offers some overall conclusions and suggestions for future studies.

Automated Writing Evaluation (AWE) in language education has a history spanning more than half a century; in the 1960s, rudimentary rule-based systems such as Project Essay Grade (PEG) were first developed, and since then, natural language processing techniques have continuously improved. By the early 2000s, proprietary AWE instruments had been widely used in second language writing environments, and systems such as Grammarly, Criterion (developed by the Educational Testing Service), and WriteToLearn provided computer-assisted feedback on syntax, punctuation, vocabulary selection, and to a lesser extent, organization [2, p. 52]. The above applications offered some real benefits by relieving teachers of an additional burden; as a result, teachers were free from such small tasks and could focus on other aspects of teaching. Nevertheless, extensive research has identified several key limitations, including a tendency towards formulaic expression, an inability to make rational judgments, and advice that, although accurate, was too limited in scope for comprehensive restructuring [5, p. 1024].

With the rise of generative AI, most notably represented by ChatGPT in late 2022 and GPT-4 in March 2023, a change has taken place in the tools that can help students with their writing. AWE is a rule-based system that performs pattern recognition and syntactically rules, whereas a generative AI model trained on a large scale of data with deep neural networks can generate contextually aware and semantically rich responses to a user’ s query. The new structure of the model can process and understand language at the level of meaning, providing more reasonable suggestions for adding content, structuring it logically, maintaining consistency with existing works, etc., unlike earlier automated tools [1, p. 158]. This innovation in language education has achieved remarkable results at the level of the revolution, and scholars have described it as both a promising outlook and an impediment to progress [3, p. 6].

The two main theoretical systems are used to study the educational integration of generative artificial intelligence in L2 writing. The first is Vygotsky’ s theory of social learning, and his idea of the Zone of Proximal Development (ZPD) suggests that students are able to achieve higher levels of proficiency with the help of more skilled teachers [6, p. 86]. Within the scope of AI-assisted writing, the generative AI platform can be viewed as a "more knowledgeable other" that provides support within a student’ s Zone of Proximal Development (ZPD) by offering suggestions, exemplary works, and error corrections according to the student’ s current skill level. The following framework uses an interactionist view of feedback and stresses the impact of meaning construction and student participation in the correction process. Therefore, the effectiveness of AI-generated feedback is not only about its accuracy but also how much it inspires learners to think about the grammatical structure and connotations in their own work during revision [7, p. 120].

Recently, organised studies have begun to incorporate the new empirical data on generative artificial intelligence in language teaching. Fang and Han (2025) have conducted an in-depth study of the new research in foreign language teaching with ChatGPT, listing applications such as reading, writing, speaking and translation, and addressing persistent problems of accuracy, bias and academic integrity [1, p. 160]. Karagoz (2025) specifically studies AI-generated criticism in English composition classes, systematically conducts applied research, and finds that although generative AI can provide many answers, its effectiveness is highly dependent on how the query is set and students’  own knowledge [2, p. 60]. Nguyen and Nguyen (2025) have listed a number of applications and problems of ChatGPT in language education, suggesting that instructional blueprints need to treat AI as a support tool, not as a replacement for teachers’  guidance [3, p. 10].

One of the first results of this study is that the feedback from generative AI can be considered a "split personality". First, generative AI can provide more interpretative-level responses; it can find flaws in the reason, suggest narrative expansion strategies, recommend structural revisions to the text, and show language improvements in expression that traditional rule-based AWE tools cannot [4, p. 15]. Conversely, generative AI is still limited in deep contextual understanding; it may generate advice that is syntactically correct but functionally inappropriate, propose modifications in conflict with academic norms, or fail to consider the author’ s specific readers and purposes [2, p. 65]. The two characters have significant consequences for how generative AI will be incorporated into L2 writing teaching; this will be the subject of the following sections in this paper.

Quasi-experimental studies have been conducted to compare the effectiveness of AI-generated comments with traditional teaching corrections and existing AWE tools more frequently. Nam, Lee and Park (2026) conducted one of the more extensive direct evaluations to determine how AIDT (Artificial Intelligence Discussion Tool) and ChatGPT differ in their feedback on L2 composition across multiple indicators of writing quality [4, p. 8]. The results showed that the synergy of the advantages was not random; ChatGPT frequently led in providing ideas for expanding the subject matter, introducing new concepts, and organising text; moreover, the proposals it put forward were regarded by the expert judges as more profound and intellectually stimulating. AIDT, however, was relatively weaker in detecting syntactical errors and problems with text flow, as seen in the focused development of superficial language analysis [4, p. 22]. Therefore, it is reasonable to assume that various AI platforms will be applied in different parts of the feedback process, and a combination of these platforms may be necessary.

Yu and Xie (2025) used a different analytical method to study the effect of generative AI commentary on teachers’  instruction rather than comparing two independent AI platforms [7, p. 115]. Quasi-experimental studies have been conducted by taking Chinese university-level EFL students as the sample, and their quality of corrective input and extent of integration into subsequent drafts (feedback incorporation) have been assessed. Although the generative AI generated feedback for language accuracy and comprehensiveness was similar to that given by teachers in writing, investigations found that the teachers’  recommendations had a significantly higher application rate. Based on this mismatch, the investigators believe that human teachers can consider a child’ s particular level of learning and growth, as well as other factors in their lives to provide tailored feedback that AI cannot offer [7, p. 128]. Students were also more willing to act on the recommendations when they knew that a reliable person had made the recommendations rather than a computer.

Jean and Li (2025) investigated an essential mediating factor of the differential effectiveness of AI feedback: prompt complexity [8, p. 248]. The three levels of prompting techniques and their applications in automated written correction by ChatGPT have been empirically studied. The basic prompt merely asked the AI to "correct the errors in this paper" and was therefore unable to surpass Grammarly’ s accuracy. Zero-shot prompts provided explicit directions, such as "locate and correct all syntactic, semantic and stylistic errors in this academic paper", and achieved results close to those of Grammarly. One-shot prompts are more effective; by including an example of the desired correction style and assignment instructions, they have significantly increased the error-detection rate over Grammarly [8, p. 256]. The above results indicate that, at present, the actual application of generative AI in teaching composition lacks effective feedback and that this deficiency can be improved by modifying the prompt engineering strategy.

In short, the above comparative studies show that AI-generated and human-provided feedback are not in opposition but rather work well together; similarly, different AI systems are also mutually compatible. Generative AI platforms, such as ChatGPT, are suitable for offering initial comments on both the surface-level and semantic errors in written works, and can provide students with timely, all-encompassing and personalised feedback on their papers. Educators will be freed from doing the more laborious work of proofreading and, as a result, have more time for high-level educational activities such as fostering deep thinking, logical reasoning and inquiry, improving students’  writing ability, and providing tailored support for individual learning needs [7, p. 132]. This integrated model is consistent with the sociocultural theory mentioned in Section 2; thus, AI is regarded as a support tool within the student’ s ZPD, but the essential function of human instructors in providing interpersonal, situational and culturally sensitive productive feedback should not be ignored [5, p. 1030].

The framework of prompt engineering refers to the system of generating and improving the input requests given to large language models to obtain desired outcomes with high accuracy, relevance and functions. Prompt Engineering in the field of educational applications for generative AI has gradually taken shape as a required capability that significantly affects both the quality and pedagogical value of the output generated by these systems [8, p. 247]. Unlike rule-based AWE systems, the system’ s computation logic does not work the same way for all kinds of user-input types; therefore, interaction with generative AI systems is extremely sensitive to the accuracy, structure and depth of instructions. Responsiveness offers both prospects and obstacles for the use of education: it provides the possibility of personalised feedback in line with different educational goals, but at the same time, a factor that can greatly affect the results varies based on the user’ s ability to write a good prompt.

Scholars have begun to divide the various levels of prompt enhancement and their effects on response quality. Generally speaking, broad questions such as "evaluate my draft" or "improve this section" are too vague, and thus the results generated by the AI system are often inconsistent or not optimal because there are no particular instructions about which parts of the work to focus on for evaluation. Zero-shot queries are explicitly presented in the medium-range section, without providing examples, and direct the AI to address specific aspects of prose quality such as syntactic accuracy, logical flow, etc. [8, p. 250] At the high-end level, a few-shot query will include one or more examples of the desired output structure to show the AI how to write a particular kind of critique in terms of genre, intensity, etc. Jean and Li (2025) studied how the development of general-to-few-shot prompting has established a dose-response relationship in the accuracy of feedback; that is, each increase in prompt complexity results in an improvement in the output at a measurable rate [8, p. 258].

As a result, the concept of "prompt literacy" is emerging: the skills and capabilities required for interactive communication with AI platforms through carefully constructed cues to achieve desired academic outcomes. Under the system of L2 composition teaching, prompt literacy refers to the ability to recognize correction purposes (such as syntactic errors, awkward phrasing, etc.), identify the intended audience and style requirements, obtain references at an appropriate level of accuracy and professional expertise, and repeatedly engage in different phases of AI cooperation to refine both the initial query and the subsequent replies [3, p. 14]. As a component of electronic literacy in language education, prompt literacy is a new set of skills that goes beyond traditional ideas of technological ability and data literacy to include the unique relationship mechanics of human-computer interaction [9, p. 1002].

However, at present, there is a significant gap between the level of prompt engineering skills students have attained and the ideal methods promoted in academic research. Research shows that many L2 learners, particularly those at an early stage of proficiency, use ChatGPT with only simple prompt methods and do not know how to use the various functions of the platform [2, p. 68]. Therefore, a new area for improving teaching is required: dedicated training in prompt engineering should be added to the composition syllabus. Teachers can teach students how to construct good prompts to improve the quality of feedback provided by AI, help students gain a deeper understanding of their own writing process through metacognition, and change their attitude towards using AI from one of dependence to one of initiative and strategy in the writing process [9, p. 1008].

5. Discussion: Toward a Human-AI Collaborative Framework

Based on the conclusions of the above discussion, a single view has been formed: generative AI is a powerful but not autonomous tool for revising L2 writing. Its function still depends on all sorts of factors, such as what kind of feedback is needed, how complicated the guiding instructions are, how proficient in skills learners are, and what kind of learning environment it is used in. Rather than viewing AI as a substitute for teachers or an untrustworthy trend, the data suggest that it should be considered a capable ally in a cooperative criticism system that employs the combined strengths of human insight and digital intelligence [7, p. 135].

Based on the above synthesis, this paper proposes a three-level human-AI cooperative structure for L2 writing correction. AI will make the basic modification and correction suggestions; only simple errors in grammar, such as syntax, orthography and punctuation, will be handled, and some initial recommendations for structure, logic and word choice will be provided. At this time, the speed, consistency and stability of the AI platform will be used to provide students with prompt responses to common writing problems. At the secondary level, teachers provide more advanced guidance and situation-based help, focusing on elements of composition that require deep subject knowledge, recognizing specific student needs, understanding cultural and stylistic norms, and having the potential for relational teaching to build students’  self-confidence and inspiration [5, p. 1035]. At the third level, students develop promptness literacy and analytical evaluation abilities; they are no longer passive recipients of correction but actively communicate with AI tools by issuing tactical cues, meticulously assess the strengths and applicability of proposals offered by the AI, and independently decide which advice to adopt and which to ignore [9, p. 1010].

This model has several teaching purposes. First of all, it proposes a new role for teachers in L2 composition teaching and is moving away from correction (which is becoming more and more automated) to guidance, support and cultivating reflective awareness of AI. The language pedagogy curriculum needs to add teaching on how to use artificial intelligence tools in writing training; that is, what technical skills are needed to judge and apply AI systems, as well as how to design teaching plans that incorporate AI in a meaningful way at all stages of teaching [1, p. 172]. Educators will more frequently take on the role of AI literacy teachers to help students develop essential analytical reasoning abilities in a scholarly setting full of AI-generated materials and feedback [3, p. 12].

Several severe constraints of the current research system need to be explicitly acknowledged. At first, the majority of research on AI feedback for L2 writing has been short-term in scope, typically involving interventions that last for a few weeks, and thus fails to examine how AI-driven feedback affects the progress of writing over extended periods. In addition, most of these studies have been conducted in university-level English as a Foreign Language (EFL) environments and have paid little attention to K-12 settings, heritage language learners or non-English language learners. Ethical problems such as academic dishonesty, data confidentiality and algorithmic bias have also not been addressed empirically in the existing studies [9, p. 1015]. Prospective inquiries should use long-term research methods, various groups of participants, and continuous ethical considerations of the consequences of adding AI to teaching for equity and student autonomy to address the above limitations.

Based on all empirical research published between 2023 and 2026, this paper investigates the effectiveness of generative artificial intelligence, such as ChatGPT, in L2 teaching as writing coaches. According to the above research results, generative artificial intelligence will help improve the quality of the feedback system for L2 learners and add various new functions to the previous AWE platforms. ChatGPT and analog tools have specific advantages in providing content-level suggestions for thematic expansion, structural coherence and conceptual synthesis, whereas specialized domain AWE systems still outperform others in superficial flaw correction [4, p. 30]. The quality of the AI-provided feedback is directly related to the complexity of the prompt design, and few-shot learning has achieved significantly better results than conventional approaches [8, p. 260].

However, according to the above data, generative artificial intelligence will not replace teachers at present. Instructor guidance consistently produces better adoption due to the interpersonal, situational and personalised factors that AI cannot replicate [7, p. 130]. The optimal teaching model recommended by the tripartite cooperative system in this paper is to integrate AI with manual instruction systematically: have AI handle repetitive grammatical corrections and initial comments, allow teachers to focus on higher-order thinking, individualised support, and cultivating AI awareness, and encourage students to actively use both forms of input through prompt proficiency and analytical evaluation skills [9, p. 1012]. The collaboration mode retains the special character of composition teaching and uses the efficiency and scope of AI innovation.

Several paths for the following studies are presented here. Longitudinal research is necessary to assess the long-term effects of AI-driven criticism on the growth of L2 composition over an extended period, moving beyond the brief interventions that have characterised most academic work recently. Research in K-12 schools will provide necessary references for understanding how younger students, with varying levels of digital literacy and metacognitive development, interact with and learn from AI-generated comments. Cross-cultural studies can explore how different societies view innovation, traditions in pedagogical authority, and standards for scholarly works to affect the effectiveness of artificial intelligence-assisted writing training. In short, to effectively promote the AI feedback ability of both students and teachers, well-structured programmes have been necessary at the level of education, and the good use of generative AI in writing courses cannot be achieved without the preparedness of the people who operate it [1, p. 175].

1. Fang S., Han Z. On the nascency of ChatGPT in foreign language teaching and learning // Annual Review of Applied Linguistics. 2025. Vol. 45. P. 148–178.
2. Karagoz I. AI-generated feedback in English writing instruction for language learners: A systematic review // The Reading Matrix: An International Online Journal. 2025. Vol. 25. No. 1. P. 51–75.
3. Nguyen H. T., Nguyen P. T. ChatGPT in language education: Applications and implications for teaching and learning // Teaching English as a Second Language Electronic Journal (TESL-EJ). 2025. Vol. 29. No. 1.
4. Nam Y., Lee J. M., Park H. Improving L2 writing through artificial intelligence: Comparing AIDT and ChatGPT feedback // English Teaching. 2026. Vol. 81. No. 1. P. 3–36.
5. Lee I. Teacher feedback in second language writing: A cognitive-affective-situated perspective // Journal of Second Language Writing. 2020. Vol. 49. P. 100737.
6. Vygotsky L. S. Mind in Society: The Development of Higher Psychological Processes. Cambridge, MA: Harvard University Press, 1978.
7. Yu H., Xie Q. Generative AI vs. teachers: Feedback quality, feedback uptake, and revision // Language Teaching Research Quarterly. 2025. Vol. 47. P. 113–137.
8. Jean G., Li M. Impact of prompt sophistication on ChatGPT’ s output for automated written corrective feedback // ReCALL. 2025. Vol. 37. No. 2. P. 245–264.
9. Zhai X. [et al.]. Generative AI in language education: Ethical considerations and pedagogical practices // Computers & Education. 2024. Vol. 211. P. 104998.