Generative AI and Second Language Writing

1 Introduction

With the launch of ChatGPT in November 2022, generative AI (GenAI) has ushered in a new era in human history. GenAI, represented by ChatGPT, DeepSeek, and other programs with similar functions, is a technology that leverages patterns extracted from training materials to generate responses to human inquiries by predicting the most likely next words. GenAI has the following characteristics. First, it is a large language model based on linguistic probabilities; therefore, its outputs are determined by linguistic contingencies and the strengths of bonds between linguistic units. Because it is a language model and is not based on contextual, rhetorical, pragmatic, or semantic cues, it may generate responses that are linguistically coherent but semantically illogical and unreasonable. Second, it is generative in that it can produce new information based on existing input materials. The input materials are publicly available materials selected and fed by IT specialists, engineers, or personnel whose qualifications are unknown. The selection criteria for training materials are also unknown. The latest update to the training materials for ChatGPT, probably the most popular GenAI tool, occurred in October 2023. It is important to clarify that GenAI is not connected with any external source of data, nor the Internet, so there is no ongoing learning. It is also necessary to clarify that according to ChatGPT, user responses are not used for training purposes, which may help dispel qualms and concerns over privacy and confidentiality. Third, it is interactive and can engage in human-like conversations with users. It may generate multiple turns, keep previous turns in memory, and retrieve and process previously stored information when requested. However, the kind of interaction it engages in with users lacks key features of natural conversations (Voss and Waring 2024). For example, it is not sensitive to user characteristics and does not use appropriate communication strategies in response to user differences; it also does not make constant adjustments based on the flow of the conversation and ongoing dynamics. Fourth, it simulates human intelligence and has the ability to perform various higher-order tasks that require or involve information processing such as analyzing, categorizing, comparing, inferencing, reasoning, summarizing, and so on.

GenAI’s versatility is evident in second language (L2) writing. Writers can use GenAI to brainstorm ideas, create outlines, summarize source information, translate information in their first language into the second language, search for L2 forms that match planned messages, proofread their essays for language, content, and rhetoric, etc. Teachers can utilize GenAI to evaluate and provide feedback on students’ writing, which is logistically challenging given the large number of assignments teachers have to evaluate. Other issues such as teachers’ failure to recognize errors and provide accurate, consistent feedback, as well as struggles teachers experience in providing feedback on the organizational aspects of writing (Lee 2008; Li and Vuono 2019; Truscott 1996), can be easily addressed with the assistance of GenAI. In addition to benefiting students and teachers, GenAI has brought opportunities to researchers, whose research territory and terrain have been expanded and reshaped. Despite GenAI’s affordances and opportunities for L2 writing, there is a lack of guidance on its effective use, there are pitfalls that have been identified by research, and many pressing issues remain unexplored in the research. This article synthesizes the various strands of research investigating the role of GenAI in second language writing, including GenAI literacy, feedback on L2 writing, prompt engineering, discourse comparison between GenAI and human writing, and ethics. Integrated into the synthesis of research findings is a discussion of new concepts and the implications of the research for theory construction, further research, and L2 writing practice and pedagogy.

2 GenAI Literacy

GenAI literacy refers to users’ knowledge of the fundamentals of GenAI’s mechanism, ethics, affordances, and limitations, and users’ ability to effectively use GenAI. GenAI literacy is an important concept due to GenAI’s ubiquity and limitations and users’ lack of knowledge about how to maximize the benefits of this new technological tool. It is important to clarify that GenAI literacy involves not only users’ awareness of GenAI’s limitations (bias, inaccurate output, etc.), but also users’ understanding of its functions and strategies to leverage it effectively. GenAI literacy entails both declarative knowledge – knowledge about its operational principles, affordances, etc. – and also procedural knowledge – knowledge about how to actually apply GenAI to solve problems. Whereas declarative knowledge takes the form of awareness and mental representations, procedural knowledge is demonstrated through performance, behavior, and use. Thus, literacy goes beyond mere awareness, although it is often interpreted and operationalized as awareness; in other words, it includes both “knowledge about” and “knowledge how”. It is necessary to further clarify that declarative knowledge is not restricted to metacognitive knowledge, such as acknowledging the importance of AI literacy (e.g., “It’s important to learn how to use GenAI”), or confirming knowledge (e.g., “I am aware of the limitations of GenAI”). Measures of GenAI literacy must include items that target specific aspects, such as strategies for effective prompt engineering, as well as activities that require users to demonstrate their GenAI competence through behaviors or task performance.

GenAI literacy is a relatively new concept derived from the notion of AI literacy, so its conceptualization and measurement are similar to general AI literacy. Based on a thorough review of the research, Ng et al. (2021) identified four components for the construct of AI literacy: (1) know and understand, (2) use, (3) evaluate and create, and (4) ethical issues. Ng et al. mapped the first three components onto Bloom’s taxonomy of cognitive skills, which stand in a hierarchy with six levels defined in terms of the information processing demands required at each level. Ranked from the lowest to the highest cognitive demands, these skills include know, understand, apply, analyze, evaluate, and create. Applied to Ng et al.’s (2021) framework of AI literacy, the stage of “know and understand” involves learning the basics of AI including its mechanism and functions; “use” refers to how to apply or use an AI tool to solve problems; and “evaluate and create” are two high-order skills users draw on to assess the quality of AI output and create new applications. The fourth component, ethics, permeates all stages of AI use and does not fall into the hierarchy of cognitive thinking. Drawing on Ng et al.’s framework, Warschauer et al. (2023) proposed a similar framework for GenAI literacy for second language writing, which also applies to L2 learning in general. In this framework, GenAI literacy has five components: understand, access, prompt, corroborate, and incorporate. “Understand” is similar to “know and understand” in Ng et al.’s framework, “access and prompt” corresponds to “use”, “corroborate” is equivalent to “evaluate”, and “incorporate” has no counterpart in Ng et al., but it can be considered part of ethics as it mainly concerns proper use of GenAI output. Wang and Wang (2025) conducted a small-scale exploratory, observational study investigating 10 L2 English writers’ GenAI use and literacy in a writing class at a liberal arts college in the U.S. The L2 writers worked on a writing assignment using ChatGPT. They made reflections on their use of ChatGPT and participated in semi-structured interviews, and the process of their using ChatGPT in writing was captured via a screen recorder. Based on the multimodal data they collected, the researchers proposed a model of critical GenAI literacy, emphasizing the critical dimensions of GenAI literacy. The model consists of four components: critical awareness (akin to “understand” and “know and understand” in the above mentioned models), critical strategies for human-AI interaction (resembling “access and prompt” and “use”), critical evaluation of affordances (similar to “evaluate” and “corroborate”), and critical positionality. Critical positionality, which is missing from other models, refers to maintaining user agency, autonomy, and authorial voice during GenAI use instead of overreliance on it, which may render the author’s identity invisible.

In the three models that were discussed above, “know basics of GenAI”, “use”, and “evaluate” are common components across all the models, although they are labeled slightly differently. Ethics is an independent component in Ng et al.’s model, is included in the “incorporate” component in Warschauer et al.’s model, and is part of the critical awareness component (e.g., bias) in Wang and Wang’s model. Critical positionality is a major component of GenAI literacy in Wang and Wang’s conceptualization but missing from the other two models. Based on the analysis of the three models and my own understanding, I propose the following components for the construct of GenAI literacy.

3 Ethics

One of the most pressing issues accompanying the advent of GenAI is ethics, which refers to common beliefs of the members of a community or society about the appropriacy of human conduct. In current conceptualization, ethics concerns diversity, equity, and inclusion (DEI), the wellbeing of individuals and the humanity, and social progress. Common principles of ethics may manifest differently within a profession or community. Regarding the use of GenAI, ethics can be understood from the perspectives of bias, ethical use, equity, and confidentiality, which are discussed in the following paragraphs.

Bias refers to incomplete or unequal representation of the integral parts, components, members, or perspectives of a unit or community. Bias can take the form of over- or under-representation of certain stake-holders (groups as well as individuals) or entities in a phenomenon, process, or event. GenAI bias takes the form of the over-representation of Western ideology, which underlies the training materials, all of which are written in English and contributed by English speakers. It can also be argued that GenAI outputs are imbued with U.S. ideology and culture because the training materials are in American English; accordingly, other ideologies and cultures are underrepresented. From the standpoint of second language writing and second language learning in general, second language perspectives are missing not only in training materials, but also in policy-making – L2 users are not considered in policies – as well as access – GenAI outputs are not accessible to L2 users whose English proficiency is often limited. Bias also exists in AI detection, a topic to be revisited below.

Ethical use is probably the greatest concern for GenAI use in writing in general and second language writing in particular. Ethical use concerns whether, when, how, and for what purposes GenAI should and should not be used. These questions should and can be examined empirically. Casal and Kessler (2023) surveyed 27 journal editors about their opinions on acceptable GenAI use for publishing purposes. 16 of them considered it acceptable to use GenAI to edit texts, 14 endorsed using it to generate computer code, 11 felt it Ok to write a summary of an article for public use, and 10 thought it ethical to write the abstract of an article. Few or no editors supported using GenAI to write parts of or the whole main document, and six agreed with the statement that under no circumstance is it acceptable to use GenAI. In general, it would seem that editors considered GenAI a tool that can be used to proofread to improve the clarity and readability of manuscripts, and they thought it unethical to use it to generate new content or perform tasks for which originality is valued.

Evidently, one major topic related to ethical use is plagiarism, which refers to using GenAI-generated output as one’s own work without acknowledging the source or using GenAI output in replacement of one’s own effort. Plagiarism can be approached and studied from various perspectives, such as plagiarism detection and prevention of plagiarism. Plagiarism detection has received much attention since the launch of ChatGPT. Liang et al. (2023) submitted TOEFL essays written by ESL speakers and essays written by eighth-grade US students to multiple detectors and found that while some detectors correctly recognized eighth-graders’ writing as human-written at a high accuracy rate, the detectors misclassified an average of over 60 % of ESL essays as GenAI-generated. Therefore, they concluded that GenAI detectors are biased against ESL writing. They suspected that the bias is due to the mechanism of GenAI detection, which is based on perplexity scores. Perplexity is an index that represents the extent to which language use is predictable, so lower perplexity or higher predictability stands for higher likelihood of GenAI outputs. This method of GenAI detection, which is based on perplexity or predictability, is the same as the mechanism of GenAI per se, which is also based on predictability. Liang et al. (2023) further found that ESL essays’ perplexity scores were low, which made the researchers suspect that the high false positive rate for ESL writing was likely due to the low perplexity scores. The low perplexity scores were likely attributable to the lack of linguistic sophistication of ESL writing, which made the writers’ language use highly predictable. To verify this hypothesis, they increased the ESL essays’ lexical sophistication by using ChatGPT and found that the inaccurate classification rate was significantly reduced. In another study, conducted by Jiang et al. (2024), a GenAI detector was used to classify human- and GenAI-written GRE essays. The human writers included both native speakers and non-native speakers of English. The researchers found a near-perfect classification rate and did not find any bias against nonnative speakers. There are two explanations for the disparate findings of Jiang et al. (2024) and Liang et al. (2023). One is the higher proficiency of the GRE essay writers in Jiang et al. than the TOEFL essay writers in Liang et al.’s study, and greater closeness to native speakers’ writing may have reduced the false positive rate. Another explanation is the training that was conducted in Jiang et al. that enabled the GenAI detector to learn the linguistic features of the genre of GRE writing and give a better detection performance. In this regard, this study suggests the importance of domain-specific training so that GenAI can have sufficient input to extract regularities related to the particular discourse and linguistic features. In general, GenAI detection is still in its infancy because of its high false positive rates, and in fact, based on Dang and Wang’s (2024) study on U.S. universities’ policies for GenAI use, 61 out of 100 universities officially banned the use of GenAI detection software.

In addition to GenAI detection, Dang and Wang’s (2024) also examined university policies on the prevention of plagiarism in writing. Their study showed that the strategies the universities suggested included using personalized assignments such as incorporating students’ personal experiences, the local context, etc.; using process-based writing instruction where students are asked to engage in the learning process instead of focusing on the product (e.g., getting a good grade); adopting alternative assignment modalities such as podcasts, presentations, etc. instead of relying on written assignments only; incorporating assignments requiring students to apply skills and knowledge to solve problems. In addition to the above strategies, one way to prevent plagiarism is to require students to disclose whether they used GenAI and for what purposes. However, Tan et al. (2025) found that disclosure of GenAI use in writing led to lower ratings than non-disclosure for the same essays.

Next, I address the last two aspects of GenAI ethics: equity and confidentiality. Equity refers to equal access to GenAI by all individuals regardless of their age, gender, ethnicity, location, nationality, profession, religion, and socioeconomic status. However, universal, equitable access to GenAI is not reality, and many individuals and communities do not have access to GenAI or all features and content because of the lack of infrastructure or because of economic and political reasons. Confidentiality refers to the possible public access to personal information and user responses; however, according to ChatGPT, user responses are not used as training materials, so the concern about confidentiality seems to be assumed.

4 Feedback on Writing

Feedback on writing refers to comments on L2 writers’ writing performance or quality. In L2 writing research, feedback has been extensively investigated, mainly from the perspective of whether corrective feedback can improve the linguistic accuracy of L2 writing and whether feedback’s effectiveness is affected by other factors such as learners’ individual differences in anxiety, working memory, etc. (An and Li 2024; Kim and Li 2024; Li and Roshan 2019; Vuogan and Li 2023). Other topics that have been examined include teachers’ feedback-providing practices, teacher and student beliefs about corrective feedback, and the congruence and incongruence of teachers’ beliefs and their feedback-providing practices (Li 2017). The large amount of experimental and observational research has been synthesized in meta-analyses (e.g., Kang and Han 2015) and narrative reviews (e.g., Li and Vuono 2019). The advent of GenAI marks the beginning of revived and intensified interest in written feedback because providing feedback on writing is a major affordance of GenAI. Besides the replication of traditional research in GenAI contexts, GenAI opens up new research territories, topics, and perspectives, such as comparisons between GenAI and teacher feedback-providing practices, students’ engagement with GenAI feedback, etc. GenAI also makes it possible to examine topics that received little attention in previous research such as content- and discourse-related feedback, reformulation of students’ essays, etc. In the following sections, I provide a taxonomy of written feedback, synthesize the limited research that has been conducted to date, and discuss the importance and methods of feedback evaluation – a unique topic related to GenAI.

5 Prompt Engineering

A prompt refers to instructions given to GenAI to perform a task or fulfil user requirements. Prompt engineering is the process of optimizing the instructions for GenAI to achieve a desired outcome. Prompt engineering is a unique topic for GenAI, and it is a crucial step because the quality of GenAI output is determined by the prompt it receives. A good prompt requires a certain level of GenAI literacy such as its affordances and principles of prompt engineering as well as domain-specific knowledge about the task, topic, and discipline. A good prompt should be clear, specific, and contextualized. Clarity is achievable by using unambiguous language, avoiding expressions with multiple interpretations, clarifying concepts that are not straightforward, and exemplifying expected output. Specificity requires the user to provide all the details necessary for the successful performance of the task GenAI is entrusted with, including input materials, action to be taken by GenAI, and output specifications. Contextualization pertains to providing information about the setting for the task GenAI is asked to perform, stakeholders’ characteristics (such as the writer’s L2 proficiency, age, grade level), and the audience or consumers of GenAI output.

An ideal prompt for GenAI use in L2 writing should consist of the following three components, illustrated by using the example of written feedback.

1. Input: materials GenAI works with. In the case of feedback on writing, input materials include (1) the writing task, including the exact writing prompt and its requirements such as time limit and word limit as well as the nature and goal of the writing task, and (2) the writers, such as students’ background (e.g., 8th-grade ESL learners at a Korean middle school), age, and English proficiency.

2. Task: the duty GenAI is asked to perform. This component also entails assigning a role for GenAI, for example, “You are an ESL teacher, and your job is to give feedback on students’ writing”.

3. Output: the outcome to be generated by GenAI. If GenAI is asked to provide feedback, this component should include information about the type of feedback (global or local; direct, indirect, or metalinguistic; focused or unfocused); the visual display of feedback (integrated, interspersed, or detached); the language (L1 or L2) in which feedback is given; the audience or consumer of the feedback (e.g., ESL writers); the criteria or rubric for evaluating students’ writing performance so that GenAI feedback is aligned with curricular goals.

Several strategies for prompt engineering should be highlighted. One is training, where GenAI is provided with a model performance or some vicarious experience that enables it to learner from examples. Prompts that involve exemplification and demonstration are called few-shot prompts, which are distinguished from zero-shot prompts where no examples or demonstrations are given. Of course, zero-versus few-shot prompts should not be taken as a dichotomy but a continuum instead. Complex tasks such as giving feedback on an essay or evaluating writing performance require few-shot prompts. Simple tasks can be completed with zero-shot prompts, such as “Please correct the grammar in the following paragraph” (Wang and Wang 2025, p. 6). Another strategy is to divide a complex task into several iterations, each focusing on one subtask, such as targeting different aspects of writing in the case of written feedback.

To conclude this section, I present the prompts used in some studies on L2 written feedback to exemplify the variation of prompts across studies, which may have potentially caused disparate results regarding GenAI’s feedback providing practices. In Guo and Wang (2024), GenAI was asked to provide feedback on content, organization, and language separately in three rounds, which allowed the researchers to explore the different foci of GenAI and teacher feedback. However, the teacher participants did not seem to be required to provide feedback on the three aspects in three rounds. It is unclear how this minor variation may have influenced the results of the study. Koltovskaia et al. (2024) provided example prompts to students for use in different stages of writing such as brainstorming, outlining, drafting, etc. To receive feedback, GenAI was asked to give “some constructive criticism” (p. 1, Supplementary Information), identify grammar and punctuation errors and suggest corrections, etc. Lin and Crosthwaite (2024) prompted GenAI to provide feedback on grammar, vocabulary, organization, and ideas, but GenAI was given the freedom to provide feedback in “any form” and “use one or numerous strategies” (p. 6). The prompt of Zou et al.’s (2025) study specified content, organization, and language as the aspects for GenAI feedback and required feedback to be given in Chinese, the students’ L1. As can be seen, these prompts have commonalities and disparities, but slight disparities may cause a major impact on GenAI’s feedback-providing practice.

6 Writing Assessment

GenAI can be used to assess writing quality or grade students’ writing assignments. To be clear, assessing is different from giving feedback because the former aims primarily to evaluate writing quality and the latter seeks to enhance writing quality and facilitate learning. A major topic for research is whether GenAI assessments are valid and reliable, and the predominant method used in the research is to compare GenAI and human ratings on the same essays following the same rubric to see whether the two ratings are consistent or correlated. A strong correlation suggests that GenAI ratings are comparable with human ratings and are therefore valid. Cast in assessment terms, this type of evidence is collected to verify a test’s concurrent validity, namely whether the test is comparable with an alternative test that measures the same construct, especially a test that has proven valid. It must be clarified that concurrent validity rests on the assumption that the criterion or existing test is valid; thus, if the test being validated is not correlated with the existing test, it means the test under validation is invalid. However, in the case of GenAI assessment, we cannot conclude that GenAI assessment is not valid if it is not aligned with teacher assessment, unless the validity of teacher assessment is established. In addition to current validity, the validity of a test can be verified from other perspectives, such as predictive validity, which concerns whether scores of a test can predict an outcome, or divergent validity, which refers to whether a test is uncorrelated with a test that is hypothesized to be unrelated to the test in point.

Among the four studies discussed below (Table 3), all examined concurrent validity by comparing GenAI ratings with human ratings or an automated writing evaluation system such as Grammarly; one examined predictive validity by probing whether GenAI ratings for linguistic accuracy can predict students’ overall writing quality. The studies to be reviewed fall into two categories investigating GenAI’s validity in assessing overall writing quality (Bucol and Sangkawong 2024; Shin and Lee 2024) and specific aspects of writing such as accuracy, complexity, and fluency (Mizumoto et al. 2024; Yamashita 2024). Bucol and Sangkawong (2024) sought to validate GenAI’s ratings for a descriptive writing task completed by EFL learners at a university in Thailand. 10 teachers were asked to rate 10 students’ essays; five teachers used ChatGPT and five used their own judgments based on the same criteria, which were related to task achievement, grammar, lexical selection, logic, and mechanics. The correlation between the two types of ratings was r = 0.65, p = 0.04. Despite the authors’ conclusion that ChatGPT demonstrated significant benefits such as efficiency, consistency, etc., a correlation of 0.65 shows a certain degree of inconsistency between human and ChatGPT ratings and is not satisfactory for assessment purposes. The authors presented snapshots of human-ChatGPT exchanges exemplifying ChatGPT’s failure to consult certain criteria and its omission of errors. The article did not report any prompt engineering where ChatGPT and human ratings would have been calibrated to be aligned with each other. It is also unclear if criteria that are not straightforward such as “task achievement” and “logic” were defined and exemplified. Perhaps prompt engineering and training would have led to greater consistency. These speculations were confirmed by another study, conducted by Shin and Lee (2024).

In Shin and Lee’s study, 50 persuasive essays (80–120 words) written by Year 2 students from several Korean high schools were rated by two experienced high school teachers who were certified writing examiners. The same essays were also rated by a scoring system built using ChatGPT. Both human and GenAI (ChatGPT) raters were provided with detailed guidance on what to do in the rating process, a detailed scoring rubric, sample essays, and sample scores. The authors reported piloting the scoring tool to ensure it could score essays according to instructions. Although details about the piloting process were not provided, clearly some training may have occurred to improve the tool’s scoring performance. The results showed high consistency between the two raters and ChatGPT, with correlations ranging between 0.78 and 0.87 and only one correlation falling below 0.80. The higher human-GenAI consistency seems to be primarily due to the rigorous training GenAI received that included not only the scoring rubric but also sample scored essays; the accurate ratings provided by the two human raters; the clearly defined rating criteria; and the larger score range (1–5) than Bucol and Sangawong (2024) (0.5–2).

While the two above studies examined GenAI’s ability to evaluate overall writing quality, Mizumoto et al. (2024) and Yamashita explored GenAI’s effectiveness in rating specific aspects of L2 writing. Mizumoto et al.’s study focused on accuracy and found an overall correlation of 0.79 between GenAI and human ratings of ESL writing accuracy; the correlation was stronger than the correlation between the ratings of GenAI and Grammarly, r = 0.69. GenAI’s correlation with overall writing was also stronger than Grammarly’s correlation with overall writing: 0.63 versus 0.55. Although a correlation of 0.79 between GenAI and human raters is acceptable, it is below 0.80, which the authors stated may have been due to the disparity in the interpretation of what constitutes an error. For example, while human raters ignored punctuation, GenAI consistently counted punctuation-related errors. The researchers also acknowledged that the inconsistency was perhaps partly because they used a zero-shot prompt that lacked elaboration and exemplification of the rating criteria. Yamashita (2024) compared GenAI and human evaluations of complexity, accuracy and fluency, and found that the GenAI-human rating correlation was 0.67 for complexity, 0.82 for accuracy, and 0.75 for fluency. Overall, these correlations are relatively weak, albeit acceptable. There are several caveats about Yamshita’s prompt. First, explanations on the three rated dimensions are not straightforward. For example, complexity is defined as the extent to which “you think the speaker/write uses morphologically and/or complex words, phrases, expressions, constructions, and grammar” (p. 6). But it is unclear what specifically complex words and grammar mean. In the literature (e.g., Li and Fu 2018), syntactic complexity has been operationalized in various ways, such as subordination, average sentence length, etc., and so has lexical complexity, which can be indexed by lexical variety (using different words), lexical sophistication (using less frequent words), etc. Also, in this study, GenAI was informed that the essays were written by nonnative speakers of English and should be graded accordingly. However, GenAI was not provided with any further information about the norms for nonnative speakers. Fluency has been operationalized as the number of words a text contains inn writing research. But in Yamashita’s study, fluency was primarily defined as dysfluencies, such as pauses, hesitations, false starts, etc., information that is not relevant to the rated texts.

7 Discoursal Comparison Between GenAI and Human Writing

One promising area of research is the comparison between GenAI- and human-generated texts in terms of linguistic, rhetorical, and semantic features. In the literature, there are essentially two types of methods to examine this topic: direct and indirect. In direct methods, human writers and GenAI are asked to write essays based on the same prompt, and a corpus analysis is conducted to identify the differences and commonalities between the two sets of essays (Jiang and Hyland 2024). One principle for this method is to use the exact same writing prompt for both human writers and GenAI in order to ensure the two types of written output are comparable. It is also necessary to specify the identity GenAI assumes (e.g., “You are a second-year student at the Department of English of X University”). In indirect methods, differences between GenAI and human writing are inferred based on GenAI’s feedback on human writing (Wang and Wang 2025) or its adaptations to human-generated texts. In the case of adaptations, two approaches have emerged: one is to ask only GenAI to adapt (Huang and Deng 2025), and the other is to ask both GenAI and human editors to adapt the same ESL essays and compare their adaptations (Tu 2025). A comparison of human and GenAI writing is beneficial in at least two ways. The results may (1) facilitate GenAI plagiarism detection by helping the tool recognize features of GenAI writing, and (2) raise writers’ awareness of the features of GenAI writing and help them avoid those features to distinguish their own writing from GenAI writing. Avoiding features of GenAI writing is particularly important for L2 writers because L2 writing is more easily misclassified as GenAI-generated compared with texts authored by native speakers, as previously discussed. The following sections provide a snapshot of the findings of the research comparing GenAI and human writing. Although not all studies involve L2 writers, the findings have implications for L2 writing.

8 Implications, Future Directions, and Guiding Principles

In the final section of the article, I discuss the implications of the research, tentative conclusions that can be drawn, future directions for major strands of research, and underlying principles for GenAI use based on the findings of L2 writing research. Starting with GenAI literacy, it is crucial to understand GenAI’s mechanisms, affordances, and limitations. GenAI literacy is an important research topic that needs to be conceptualized, instruments must be developed to measure it, and research must be conducted to examine the effects of interventions on its improvement. Due to the importance of GenAI literacy and users’ lack of knowledge about its affordances and limitations, GenAI literacy should be included as part of GenAI interventions L2 writing research (Koltovskaia et al. 2024). It should also be a pedagogical requirement that must be addressed in institutional policies, course syllabi, and program curricula where independent courses can be offered on GenAI literacy. One aspect of GenAI literacy is ethics related to GenAI use, and two core components of the GenAI ethics are bias and ethical use. Biases might be political, ideological, cultural, and social, and one bias related to L2 writing is the bias against ESL writers – they are often misclassified as GenAI writers. Ethical use is a central topic of GenAI use because one of the greatest challenges posed by GenAI is preventing plagiarism and fraud. Both bias and ethical use require empirical investigation that can be conducted from various perspectives.

Moving on to the process of using GenAI, one crucial initial step is prompt engineering because the quality of GenAI relies on the prompt. A good prompt must be clear, specific, and contextualized. It is important to specify all the parameters of the expected output, define and exemplify concepts, provide enough “fodder” – training materials – to GenAI so it can recognize patterns and learn what it is expected to do. When eliciting feedback on writing, the user needs to give specific instructions on what types of feedback are expected, provide criteria for the expected feedback, and vet the quality of the generated feedback. The research on corrective feedback suggests that teachers and GenAI may work together and draw on their strengths in feedback-providing, for example, GenAI may focus on organization and teachers on content. Peer feedback may be drawn into the picture and may focus on content, together with teacher feedback, given the effects of content-related peer feedback on L2 writing (Vuogan and Li 2023). It must be pointed out that most research on GenAI feedback is observational, examining learners’ attitudes, use, and engagement with GenAI feedback, and there is insufficient experimental research on GenAI’s effects on the development of L2 writing ability.

Experimental studies need to systematically manipulate the type and amount of feedback provided in each treatment condition so that any differences between treatment groups can be attributed solely to the manipulated (independent) variable instead of extraneous, confounding variables. A finding that GenAI feedback is more effective than teacher feedback is difficult to interpret if GenAI and teachers provided different types and amounts of feedback. GenAI can provide different types of feedback, which previous research has found to have differential effects on learning gains (Kang and Han 2015; Li and Vuono 2019). Experimental research is needed to examine the extent to which the results can be replicated in GenAI contexts, and to investigate new variables or opportunities created by GenAI such as integrated versus detached feedback. Theoretically, integrated feedback, which is embedded in the text, is contextualized and provides an immediate comparison between errors and correct forms. Thus, it is more effective than detached feedback, which takes the form of a list of corrections at the end of the text. Another example topic with a significant potential in GenAI contexts is the timing of corrective feedback (Li 2020; Li et al. 2025). Feedback timing can be operationalized in various ways, such as focusing on different aspects in different drafts during the cycle of a writing assignment or correcting errors while learners are engaged in the process of composing versus after the writing task is completed (Cheng and Zhang 2024; Shintani and Aubrey 2016).

One area that appears related to but actually separate from feedback is writing assessment; the former focuses on helping students learn and the latter on evaluating writing quality or performance. One recurring theme of the research is that GenAI ratings are more consistent with human ratings when assessment criteria are clearly defined and sufficiently illustrated and when enough training is provided to GenAI. Research is needed on ways to increase the validity of GenAI assessments, and principles of effective assessments should be applied in the examination of the construct validity of GenAI assessments, such as reliability, concurrent validity, divergent validity, convergent validity, and predictive validity. The findings of this stream of research may provide valuable implications for large-scale standardized assessments as well as curriculum-based or classroom assessments.

Finally, research on the comparison of the discourse and linguistic features of GenAI and human writing shows that GenAI writing is more formal, rigid, impersonal, academic, linguistically complex, and formulaic, whereas human writing is more personal, creative, engaging, and less predictable. These findings help us distinguish human and GenAI writing, contributing to more accurate detection of GenAI use in plagiarism inspection and enabling us to make informed decisions in our attempts to maintain our writing identity and our identity as a unique human being.

To conclude this article, I propose three principles that may guide the use, research, and theory construction of GenAI. The first and overarching principle is individualization. Individualization may be understood as the personalization of writing style including content, language, organization, and other aspects of writing. Personalization contrasts with commonality, patterning, and typicality – defining characteristics of GenAI output; therefore, personalization sets human writing apart from GenAI writing. As reported in previous sections, GenAI writing has been found to lack personal (authorial) voice, creativity, and logical reasoning – characteristics of human writing. Another way to apply the concept of individualization is to tailor GenAI output to fit user characteristics such as L2 writers’ proficiency level, age, personal traits and dispositions (anxiety, motivation, working memory, etc.), and so on. Individualized assistance is particularly important for GenAI use because GenAI does not provide tailored assistance and it provides the same output to all users without considering their individual differences.

The second principle is domain-specificity. Here the scope of domain varies across contexts and may refer to an area of learning such as L2 learning, a subarea of learning such as L2 writing, and a particular task within the subarea. Some examples may help illustrate the importance of domain-specificity. For example, using GenAI to assess TOEFL writing involves training GenAI using the TOEFL rating rubric and scored TOEFL essays. Asking GenAI to compose a text necessitates the provision of task-specific information regarding the goal, genre, audience, assumed writer’s identity, etc. in addition to the exact writing prompt.

The third principle is agency, which means that users should take a critical stance vetting the quality of GenAI and exercising autonomy instead of overreliance on GenAI. Users should take control over the whole process by actively improving GenAI literacy, learning or using strategies for effective prompt engineering, analyzing GenAI output to determine its acceptability instead of passively accepting it, and reflecting on the process for future improvement. Agency also means maintaining one’s identity when taking advantage of GenAI output, such as determining whether a change suggested by GenAI fits one’s own style or makes one modify their writing in a dramatic way to the extent that their identity is lost.