Argumentative writing proficiency constitutes a fundamental component of academic literacy acquisition for learners of English as a Foreign Language (Zhu, 2001), involving not only formulating texts characterized with linguistic accuracy but also constructing coherent and persuasive evidence-based arguments, integrating counterarguments, and maintaining a logical flow (Ferretti and Graham, 2019; Su et al., 2023; Su et al., 2021). Effective feedback represents a pivotal mechanism in the developmental trajectory of argumentative writing competence, with instructor-mediated evaluative commentary traditionally serving as the primary intervention modality for facilitating rhetorical enhancement and structural coherence (Banihashem et al., 2022; Zhang and Hyland, 2018). However, providing effective feedback to learners is significantly constrained by structural barriers inherent in contemporary educational settings characterized with high student-to-teacher ratios and heterogeneous proficiency distribution within the same classroom. These systemic limitations transform personalized evaluative feedback into a resource-intensive pedagogical intervention, frequently resulting in delayed feedback and diminished instructional efficacy (Guo et al., 2024; Liu et al., 2024; Ouahidi, 2021; Wisniewski et al., 2020). The advent of Generative Artificial Intelligence (GenAI) in educational contexts has introduced novel avenues for feedback delivery. AI-powered tools, such as ChatGPT, can generate immediate, detailed, and personalized feedback on student writing, potentially alleviating the workload of educators and providing timely assistance to learners (Guo et al., 2024; Lee and Moore, 2024). However, there is limited empirical research on how GenAI tools may assist in scaling feedback, especially in large English as a Foreign Language (EFL) writing classes (Li et al., 2024). For example, Wang and Dang’s (2024) systematic review pointed to a significant dearth of empirical investigations examining applications of GenAI in EFL settings. Within the limited scholarly discourse, studies have begun to explore the efficacy of AI-generated feedback in comparison to traditional teacher feedback. For instance, research indicates that GenAI feedback can be as effective as teacher feedback in improving certain aspects of writing, such as coherence and cohesion (Yoon et al., 2023). However, concerns persist regarding the depth, accuracy, and contextual appropriateness of AI-generated feedback, particularly in addressing higher-order writing skills (Chan and Hu, 2023; Yoon et al., 2023). Moreover, the integration of AI into educational feedback mechanisms raises ethical considerations, notably regarding data privacy, the potential for overreliance on technology and plagiarism (Sánchez-Vera et al., 2024). Ensuring that AI tools are used responsibly, and that student data is protected is paramount (Yan et al., 2024). In the Arab world, very few studies looked at the impact of GenAI on higher education settings (Fadlelmula and Qadhi, 2024). This study aims to fill this gap by investigating the comparative effectiveness of AI-generated feedback and teacher-generated feedback on the argumentative writing performance of EFL students in Jordanian settings. Specifically, it seeks to determine whether AI feedback can match or surpass the quality and impact of traditional teacher feedback in enhancing students’ abilities to construct and refine arguments, integrate counterarguments, and maintain logical coherence in their writing. The questions this paper aim to answer are as follows:

The findings of this study hold significant implications for EFL writing instruction. By elucidating the effectiveness of AI-generated feedback, educators can make informed decisions about integrating AI tools into their pedagogical practices. Understanding the comparative advantages and limitations of AI and teacher feedback can aid in designing hybrid feedback models that leverage the strengths of both approaches. Furthermore, addressing the ethical considerations associated with AI in education will contribute to the development of guidelines and policies that ensure responsible and effective use of technology in language learning contexts.

Argumentative writing is widely recognized as one of the most cognitively and linguistically demanding genres in EFL instruction (Ferretti and Graham, 2019). It requires learners to formulate a clear stance, develop logical reasoning, integrate counterarguments, and adhere to academic discourse conventions (Ferretti and Graham, 2019; Su et al., 2023; Su et al., 2021). These requirements make argumentative writing particularly difficult for students whose linguistic proficiency is still developing (Pelenkahu et al., 2024). In the EFL context, students often struggle not only with surface-level features such as grammar and vocabulary but also with deeper genre-related demands, such as organizing their arguments logically and integrating rebuttals effectively (Mallahi, 2024). This dual challenge underscores the need for scaffolding that supports both linguistic and rhetorical development. Several studies emphasize that EFL learners commonly exhibit a “one-sided” argument structure, with minimal integration of opposing viewpoints (cf. He and Du, 2024; Qin and Karabacak, 2010; Wagner et al., 2017). Moreover, as learners tend to rely heavily on formulaic expressions, they demonstrate limited use of critical thinking strategies during the planning and revision stages of writing (Ferretti and Graham, 2019; Kuhn, 1991). These limitations are often exacerbated in mixed-ability classrooms, where less proficient students may lack the metacognitive strategies necessary to revise content meaningfully, while more proficient learners still require structured support for advanced rhetorical moves such as counterargument and rebuttal (Hazaea, 2023).

Feedback has been identified as one of the most pivotal pedagogical tools in improving argumentative writing skills. According to Hattie and Timperley’s (2007, p. 86) model, feedback clarifies goals (feed up), points out progress (feed back), and closes gaps between current achievement and learning outcomes (feed forward). Numerous studies confirm that high-quality, formative feedback significantly enhances learners’ writing performance, especially when it is timely, specific, and focused on meaning-level aspects (Bitchener and Ferris, 2012; Peltzer et al., 2024; Zhang and Hyland, 2018). In argumentative writing, feedback is particularly crucial for helping learners recognize logical gaps, improve coherence, and refine argumentative strategies (Guo et al., 2024). However, providing individualized feedback in large and mixed-ability classrooms remains a persistent challenge, especially in resource-constrained settings (Liu et al., 2024; Ouahidi, 2021; Wisniewski et al., 2020).

Recent advances in large language models (LLMs) such as ChatGPT have opened new possibilities for AI-supported writing instruction. Unlike traditional automated writing evaluation (AWE) tools, which primarily target grammar and syntax, LLMs can provide nuanced, contextualized, and semi-structured feedback on content-level aspects of writing (Yoon et al., 2023). ChatGPT, in particular, has shown promise in supporting learners with argument structure, organization, and even idea generation, especially when guided by carefully designed prompts (Guo et al., 2024). Studies by Guo et al. (2022), Mahapatra (2024) show that AI-supported scaffolding can improve feedback quality, revision depth, and learner motivation. These tools are especially effective when students are trained to use structured prompts that focus AI output on genre-specific goals, such as argument strength and counterargument inclusion (Mollick and Mollick, 2023). In a study on chatbot-assisted argumentative writing, Guo et al. (2022) found that students who received AI support produced stronger arguments and engaged in deeper revision than those relying solely on peer feedback. These findings are echoed in research by Mahapatra (2024), who found GenAI feedback to have a significant positive impact on student learning coupled with positive student perceptions (p. 13).

Feedback effectiveness in writing instruction is not solely determined by its quality or timing but also by how learners engage with, interpret, and apply the feedback they receive; a process often referred to as feedback uptake (Carless and Boud, 2018). The Feedback Engagement Model proposed by Winstone et al. (2017) emphasizes that feedback is inherently dialogic, requiring active learner agency for it to be pedagogically impactful. This model outlines key stages in feedback engagement, including noticing, sense-making, and implementation each of which can vary significantly depending on whether the feedback originates from a human or an AI source.

When interacting with teacher-generated feedback, learners often benefit from interpersonal trust, contextual knowledge, and nuanced scaffolding that is aligned with classroom dynamics. However, such feedback may be delayed due to time constraints and can sometimes be inconsistent in tone or focus (Zhang and Hyland, 2018). By contrast, AI-generated feedback (e.g., from ChatGPT) offers immediacy and consistency, and can be tailored through prompt engineering to target specific rhetorical or genre-related issues (Guo et al., 2024). Yet, studies show that learners may interact with AI feedback more passively, often accepting suggestions without critical evaluation or reflection (Yoon et al., 2023). This passive uptake raises concerns about surface-level revision, overreliance, and reduced metacognitive engagement. Effective uptake of AI feedback, therefore, depends not only on the technical quality of the output but also on how learners are trained to critically engage with it. Furthermore, scaffolded reflection such as asking students to justify how they revised based on AI feedback can mitigate overreliance and encourage critical thinking (Woo et al., 2024). Research suggests that embedding reflective prompts (e.g., “Which of these AI suggestions will you use, and why?”) and requiring justification for revisions can enhance cognitive engagement and support deeper learning (Mollick and Mollick, 2023; Winstone et al., 2017). Ultimately, while both AI and teacher feedback can be effective, their pedagogical value is mediated by how learners perceive their credibility and interact with the feedback in the revision process.

However, the limitations of AI feedback are also widely acknowledged. Yoon et al. (2023) caution that while ChatGPT can generate plausible feedback on coherence and logic, it may also “hallucinate” critiques, offer generic advice, or miss contextually nuanced issues. There is also the risk that students may accept AI suggestions uncritically, leading to overreliance and potential plagiarism (cf. Alshurafat et al., 2024; Esmaeil et al., 2023; Hostetter et al., 2024; Sánchez-Vera et al., 2024). These limitations underscore the importance of prompt design, feedback framing, and teacher mediation in AI-supported writing instruction.

This study explored the comparative effectiveness of AI-generated and teacher-generated feedback on EFL students’ argumentative writing performance across two ACTFL proficiency levels. Findings revealed that both feedback types led to statistically significant gains in writing scores, with no significant difference between the AI and teacher groups. This result, coupled with a small effect size for the between-group comparison, suggests that well-structured AI feedback (delivered through prompt engineering) can serve as a scalable and pedagogically meaningful alternative to traditional teacher feedback. Notably, proficiency level emerged as a significant predictor of post-revision performance, with Advanced-Low learners outperforming Intermediate-Low learners. However, the largest within-group effect sizes were observed among Intermediate-Low students, indicating their high responsiveness to structured revision support, regardless of the feedback source. The results underscore the growing potential of large language models (LLMs) in second language writing instruction, especially in contexts where teacher feedback is limited by class size or time constraints. At the same time, the findings reinforce the importance of embedding AI use within guided, ethical, and pedagogically sound frameworks. This study is not without limitations. Essays were scored by a single rater, and only immediate revision gains were assessed. Future studies should include multiple raters, examine long-term effects of AI-assisted revision, and incorporate learner reflections to better understand feedback uptake. As educational technologies evolve, it is imperative to continue evaluating how AI can complement rather than replace the human element in language teaching and learning. Despite the single-rater design, the inclusion of inter-rater reliability metrics for a representative subsample provides reasonable confidence in scoring validity and transparency.

The findings have significant implications for EFL writing instruction across diverse educational contexts. Large language models can serve as effective feedback partners, delivering genre-specific guidance that supports student revision while freeing instructors for higher-order pedagogical activities. Standardized prompt design focusing on genre conventions can mitigate AI inaccuracies and misdirection, while the scalable nature of AI feedback enables differentiated instruction in large, mixed-proficiency classrooms. However, AI should complement rather than replace teacher feedback, functioning as a preliminary revision tool that prepares students for subsequent instructor guidance. This tiered approach preserves essential human pedagogical expertise while leveraging technological capabilities to enhance instructional effectiveness and accessibility across varied resource contexts.

While this study offers valuable insights into the comparative effects of AI and teacher feedback, several limitations must be acknowledged. First, although most essays were scored by a single trained rater, inter-rater reliability was established through a dual-rating process on a stratified subsample of 36 essays. The inclusion of a second rater, use of a validated analytic rubric, and high transparency in scoring procedures help mitigate concerns of bias and subjectivity. Pre-test scores showed moderate agreement (r = 0.653), while post-test scores revealed more variability (r = 0.316), a common pattern in subjective assessment of revised writing. Several factors may have contributed to this issue. First, the nature of the post-test responses, produced after exposure to individualized feedback and revision, may have led to more diverse writing structures and strategies, increasing subjectivity in rating. Second, although raters used a shared rubric, differences in interpretation may have emerged when evaluating revisions. Discrepancies were resolved through discussion to ensure shared understanding of rubric dimensions. Nonetheless, future studies should consider full-scale dual scoring or use intra-class correlation (ICC) to assess agreement more robustly. To address this in future research, we recommend the implementation of full dual scoring for all writing samples rather than a subset, alongside rigorous rater calibration sessions prior to and during scoring. These practices can help improve consistency and minimize subjectivity in evaluating student writing, especially in post-intervention contexts where performance tends to be more heterogeneous.

Second, the study examined short-term effects of feedback on a single revision cycle. Longitudinal data would be needed to evaluate the durability of learning gains and the impact of sustained feedback engagement over time. Third, another limitation of this study is the absence of direct measurement of student feedback uptake. While our approach was guided by theoretical models of feedback engagement (e.g., Winstone et al., 2017), we did not collect empirical data on how students interpreted or applied the formative feedback provided. As a result, the study cannot account for individual differences in feedback engagement or clarify the specific ways in which feedback influenced revisions. Future research should consider incorporating qualitative and process-oriented methods, such as think-aloud protocols, student interviews, or digital revision tracking, to capture how learners interact with feedback and make use of it during revision. Lastly, a further limitation of this study lies in its use of a single argumentative writing prompt as the basis for data collection. While argumentative writing is an important academic genre, relying on a single task restricts the extent to which findings can be generalized to other types of writing, such as narrative, expository, or reflective genres. Moreover, genre-specific features may influence how students interpret and apply feedback, meaning that the observed effects of the intervention may not transfer uniformly across contexts. Future research should replicate this study using a broader range of writing genres and disciplinary tasks to assess whether the effectiveness of rubric-aligned, AI-generated feedback varies by genre or subject area.

Although post-test scores showed lower inter-rater correlation, the inclusion of a second rater on a stratified subsample mitigates the risk of single-rater bias and supports the reliability of scoring. While ICC estimates indicated moderate agreement (ICC = 0.61), full-scale dual scoring across all essays would further enhance the reliability and generalizability of the findings, particularly in the context of post-revision assessment where scoring tends to be more variable.

Future research endeavors should investigate several critical dimensions to advance understanding of AI-assisted writing instruction. Longitudinal studies examining the cumulative effects of iterative AI-supported writing cycles would provide valuable insights into sustained performance trajectories and skill development patterns over extended periods. Additionally, comparative analyses of differentiated AI prompt strategies, including structure-oriented, error-correction focused, and tone-modification approaches, would elucidate the relative efficacy of targeted feedback modalities in facilitating specific aspects of revision behavior. Investigations into student attitudes and confidence levels regarding AI-generated versus instructor-provided feedback represent another essential research direction, particularly given the implications for pedagogical acceptance and implementation. Furthermore, the incorporation of qualitative methodologies would substantially enhance the depth of understanding regarding student engagement with AI feedback systems. Specifically, stimulated recall interviews and reflective journaling protocols could illuminate the cognitive processes underlying students’ interpretation, evaluation, and application of AI-generated suggestions, a domain that remains significantly underexplored in current literature. Such mixed-methods approaches would provide crucial insights into the mechanisms through which AI feedback influences writing development and inform evidence-based best practices for educational implementation.