When hybrid (ChatGPT and teacher) feedback does not outperform teacher-only feedback: evidence from A2 EFL writing in Ecuador

1 Introduction

Despite the growing integration of artificial intelligence (AI) tools in language education, producing grammatically accurate sentences may still be difficult for lower-proficiency English as a foreign language (EFL) learners who are developing control of basic structures such as verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions (Little, 2020; Richards and Renandya, 2002). Indeed, writing is widely recognized as one of the most demanding skills in EFL learning because it requires learners to coordinate idea generation, organization, vocabulary selection, and grammatical control at the same time (Manchón, 2011; Richards and Renandya, 2002). For A2 students, this process may be even more challenging because their linguistic resources are still limited. As a result, grammatical errors can affect the clarity and acceptability of written productions, especially when students are asked to produce original sentences without relying on models, dictionaries, or digital tools.

Traditionally, teacher feedback has played a central role in supporting students' writing development. It provides contextualized and personalized guidance that can help learners identify errors, understand grammatical rules, and revise their written production (Ferris, 2011; Hyland and Hyland, 2006). Nevertheless, teacher-only feedback may be limited by time, workload, and consistency, especially when teachers need to provide individual feedback to several students within limited classroom time (Bitchener and Ferris, 2012; Lee, 2017). These limitations have encouraged the exploration of technological tools that could support feedback practices without replacing the pedagogical role of the teacher.

In recent years, the use of AI in EFL and English as a second language (ESL) education has increased considerably. Systematic reviews suggest that tools such as ChatGPT may support EFL writing by providing immediate feedback, improving linguistic accuracy, and assisting learners during different stages of the writing process (Amoush and Alhosban, 2025; Ibrahim and Kirkpatrick, 2024; Lo et al., 2024; Mali, 2025; Slamet and Basthomi, 2025; Teng, 2024a; Xiao et al., 2026). Similarly, recent reviews have reported that AI-supported tools can contribute to engagement, achievement, and learner satisfaction when they are integrated through appropriate pedagogical mediation (Kundu and Bej, 2025; Slamet and Basthomi, 2025). However, these studies also warn that AI-generated feedback may lead to overreliance, superficial revision, or limited critical engagement if students are not guided by teachers (Alnemrat et al., 2025; Crosthwaite and Sun, 2026; Ding et al., 2025; Khojasteh et al., 2025; Tran, 2025).

Consequently, a hybrid approach that combines AI-generated feedback with teacher feedback has emerged as a possible solution (Lo et al., 2024). This approach aims to integrate the immediacy of AI with the expertise and contextual understanding of teachers. Some authors suggest that ChatGPT-supported and hybrid feedback may enhance writing performance, grammatical accuracy, learner engagement, and revision practices when it is used under appropriate pedagogical guidance (Alnemrat et al., 2025; Alsofyani and Barzanji, 2025; Guo and Wang, 2024; Han and Li, 2024; Khojasteh et al., 2025; Tran, 2025; Zhang et al., 2025). However, it is important that those reported benefits be examined across different classroom contexts, since the effectiveness of AI-supported feedback may depend on how it is integrated into instruction and how students use it to revise their writing.

Therefore, the present study adopts a quantitative quasi-experimental approach to examine the impact of hybrid feedback, understood as ChatGPT-generated feedback combined with teacher feedback, compared with teacher-only feedback on the grammatical accuracy of A2 EFL students in a technical higher education institute in Quito, Ecuador. By focusing on specific grammatical structures such as verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions, this study seeks to provide empirical evidence on whether hybrid feedback may contribute to grammatical development in a lower-proficiency, on-site EFL context. Accordingly, the study is guided by the following research questions:

• Does the students' overall grammatical accuracy improve from pretest to posttest?
• How do grammatical-accuracy gains across specific grammatical structures differ between hybrid feedback and teacher-only feedback?
• Are there statistically significant differences in grammatical accuracy gains between students who receive hybrid feedback and those who receive teacher-only feedback?

2 Literature review

2.1 Corrective feedback and teacher feedback in EFL writing

Written corrective feedback is central to EFL writing because it may help learners notice gaps between their current production and more accurate target-language forms (Ellis, 2009; Ferris, 2011; Hyland and Hyland, 2006). Beyond error correction, it can support revision, learner awareness, and gradual linguistic control (Bitchener and Ferris, 2012; Bitchener and Storch, 2016; Nassaji and Kartchava, 2021). However, its value remains debated: Truscott (2007) questioned the long-term effectiveness of grammar correction, whereas other studies suggest that focused and meaningful feedback may improve grammatical accuracy when learners have opportunities to process it through revision (Bitchener, 2008; Bitchener et al., 2005; Shintani et al., 2014). Thus, effectiveness appears to depend on the focus, clarity, timing, and uptake of feedback rather than on correction alone.

Teacher feedback is especially relevant because teachers can explain why errors occur, prioritize structures, and adapt support to learners' proficiency and instructional content (Ferris, 2010; Hyland and Hyland, 2019; Lee, 2017). Direct feedback may benefit lower-proficiency learners with limited capacity to self-correct, whereas indirect or metalinguistic feedback may promote reflection when learners can interpret it (Ellis, 2009; Ferris and Roberts, 2001; Shintani and Ellis, 2013). Classroom studies nevertheless suggest variable effects across learners, grammatical structures, and feedback conditions (Bitchener et al., 2005; Boggs, 2019; Shintani et al., 2014; Thi and Nikolov, 2022). These findings reinforce that learners need to notice, understand, and apply corrections, which may be particularly challenging for A2 students still developing control of basic sentence-level structures.

Despite its pedagogical value, teacher-only feedback may be constrained by time, workload, class size, and the need to provide consistent guidance throughout drafting, correction, revision, and follow-up (Bitchener and Ferris, 2012; Lee, 2017; Richards and Renandya, 2002). These practical constraints have encouraged interest in technology-mediated and AI-generated feedback as possible ways to increase immediacy and feedback access (Shi and Aryadoust, 2024; Teng, 2024b).

2.2 AI-generated feedback in EFL writing: benefits and limitations

AI-generated feedback may address practical constraints of teacher-only feedback in EFL writing. Earlier automated written feedback (AWF) systems were mainly associated with rule-based correction, automated scoring, and surface-level language support, whereas ChatGPT may offer flexible, dialogic, and immediate feedback (Barrot, 2023; Shi and Aryadoust, 2024; Teng, 2024b). However, Shi and Aryadoust (2024), in their systematic review of AI-based automated written feedback, reported varied findings, suggesting that effectiveness may depend on the tool, instructional design, and learner engagement.

Recent reviews suggest that ChatGPT may support EFL writing through immediate, personalized feedback on linguistic accuracy, vocabulary, organization, and revision (Amoush and Alhosban, 2025; Crosthwaite and Sun, 2026; Ibrahim and Kirkpatrick, 2024; Mali, 2026; Teng, 2024b). Likewise, empirical studies suggest potential improvements when learners use ChatGPT to revise drafts and reflect on feedback (Shi et al., 2026; Song and Song, 2023; Teng, 2024b). Mali (2026) highlights the potential value of explicit grammar corrections, which is relevant to this study's focus on grammatical accuracy.

However, ChatGPT feedback may remain superficial or be accepted without critical evaluation (Crosthwaite and Sun, 2026; Slamet and Basthomi, 2025; Teng, 2024b). Other concerns include inaccurate feedback, limited contextual sensitivity, academic integrity issues, and overreliance on AI-generated corrections (Barrot, 2023; Lo et al., 2024; Ranalli et al., 2017; Tsai et al., 2024). These limitations may restrict grammatical development if learners do not understand the corrections or apply them in later writing.

Effective use of ChatGPT feedback may also depend on prompt formulation and learners' ability to interpret, evaluate, and adapt its suggestions (Barrot, 2023; Mahapatra, 2024; Mali, 2026). Therefore, lower-proficiency learners may need support to evaluate AI-generated explanations. Kundu and Bej (2025) emphasize the importance of pedagogical mediation, teacher readiness, digital literacy, and classroom conditions in AI-supported learning. These conditions highlight the need for guidance and instructional alignment.

2.3 Hybrid feedback as a complementary feedback model

Hybrid feedback combines AI-generated feedback with teacher mediation to address the limitations of single-source feedback. Teacher feedback may provide contextualized, pedagogically sensitive, and personalized guidance, but it can be limited by time, workload, and delayed delivery. In contrast, AI-generated feedback may offer immediacy, consistency, and detailed surface-level correction, but it may lack contextual sensitivity, pedagogical judgment, and the capacity to address learners' individual needs in depth (Han and Li, 2024; Khojasteh et al., 2025; Zhang et al., 2025). In this model, AI may provide immediate correction, while teachers adapt and contextualize feedback according to learners' needs.

Empirical studies suggest potential benefits in EFL and second language (L2) writing. Han and Li (2024) found that ChatGPT-supported teacher feedback addressed local errors and global rhetorical issues in a Chinese tertiary EFL context, with teachers adapting AI output before sharing it with students. Khojasteh et al. (2025) reported greater improvements in task achievement, coherence, and grammatical accuracy for hybrid feedback than for teacher e-feedback or AI feedback among Iranian intermediate EFL learners. Zhang et al. (2025) also found greater gains with hybrid feedback than with generative AI (GenAI) feedback alone, particularly in organization and critical thinking.

However, much of the evidence concerns university or intermediate-level learners completing academic or argumentative writing tasks, which may limit its applicability to A2 sentence-level writing. For instance, Zhang et al. (2025) reported that GenAI feedback was more effective for basic writing skills, whereas hybrid feedback was more beneficial for complex academic writing skills. Alnemrat et al. (2025), comparing AI-generated with teacher-provided feedback rather than a hybrid model, found no statistically significant difference and only a small effect size. Hybrid feedback may also confuse learners when comments conflict or exceed their capacity to interpret them. Han and Li (2024) highlight teacher mediation in adapting AI output, while Khojasteh et al. (2025) report that students valued AI's quick corrections but relied on the teacher to understand them. Consequently, teacher mediation may involve selecting, simplifying, and clarifying feedback according to learners' needs.

These complementary roles align with feedback models distinguishing task-level correction from process-oriented guidance and with sociocultural perspectives on learning through tools, interaction, and scaffolding (Hattie and Timperley, 2007; Vygotsky, 1978). The effectiveness of hybrid feedback may depend on its integration, teacher mediation, and learners' application of feedback during revision. Positive student perceptions, as reported by Khojasteh et al. (2025), should be distinguished from measurable learning gains. Ding et al. (2025) suggest that specific advantages of hybrid feedback may not translate into significantly greater overall improvement. Indeed, Lo et al. (2025) found greater improvement in revised writing scores with teacher feedback than with hybrid or AI-only feedback. Thus, the present study examines whether hybrid feedback adds measurable value over teacher-only feedback across specific grammatical structures among A2 EFL learners in technical higher education in Ecuador.

2.4 Grammatical accuracy in A2 EFL writing

Grammatical accuracy supports the clarity and comprehensibility of EFL writing. Lower-proficiency learners must develop basic sentence structures while managing ideas, vocabulary, and organization (Manchón, 2011; Richards and Renandya, 2002). The Council of Europe (2020) describes A2 learners as able to write short sequences of phrases and sentences using basic connectors, although recurrent grammatical errors remain. At this level, accuracy concerns control of foundational structures for clear communication.

Focused, understandable feedback linked to revision may support grammatical accuracy. Ferris and Hedgcock (2013) identify recurrent error patterns as useful feedback targets, while Bitchener and Ferris (2012) suggest that focused correction may be more manageable than unfocused feedback. However, Manchón (2011) cautions that excessive emphasis on accuracy may discourage linguistic experimentation, and Ruegg (2015) highlights the relevance of feedback, focus, and duration. Hence, grammatical accuracy represents one dimension of writing development.

Reviews suggest that ChatGPT feedback may support grammar and other surface-level features through immediate, actionable corrections (Crosthwaite and Sun, 2026; Shi and Aryadoust, 2024; Teng, 2024b). Such revisions do not necessarily indicate broader writing development. Crosthwaite and Sun (2026) note that lower-proficiency learners may focus on surface-level changes and require teacher mediation to engage critically with AI feedback.

Accordingly, the present study examines six grammatical categories relevant for A2 sentence-level writing: verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions. These categories represent common difficulties observable in short written responses and allow comparison of improvement patterns under teacher-only and hybrid feedback. Analytic assessment can distinguish performance across grammatical categories (Bitchener and Ferris, 2012; Ferris and Hedgcock, 2013; Manchón and Matsuda, 2016). In this study, pretest and posttest responses were scored for accurate use of each target structure in tasks aligned with learners' proficiency and instructional content.

3 Materials and methods

This study employed a quantitative quasi-experimental design to examine the effect of hybrid feedback (ChatGPT and teacher feedback) compared with teacher-only feedback on the grammatical accuracy of A2-level English as a foreign language (EFL) learners. Quantitative research allows researchers to examine differences between variables through numerical measurement and statistical analysis, while quasi-experimental designs are appropriate when an intervention is examined under conditions in which random assignment to groups is not feasible (Creswell and Creswell, 2023). In the present study, participants belonged to two intact A2-level classes at a technical higher education institute in Quito, Ecuador.

Both the control and the experimental group completed the same Grammar Accuracy Writing Test before and after a six-week instructional intervention. The feedback condition was considered the independent variable, while grammatical accuracy was the dependent variable. Grammatical accuracy was examined through total pretest and posttest scores and through performance across six target structures: verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions. The pretest-posttest structure allowed changes within each group and differences in grammatical accuracy gains between the two feedback conditions to be examined. However, because the study involved intact classes rather than random assignment, differences between the groups cannot be interpreted as definitive causal effects of the feedback condition alone. Therefore, the findings should be interpreted considering the limitations of the quasi-experimental classroom design.

3.1 Participants

The participants were 41 A2-level EFL students enrolled in two on-site English classes at a technical higher education institute in Quito, Ecuador. All participants spoke Spanish as their first language and were learning English as a foreign language. Their A2 classification was based on enrollment in A2.2, the final course in the institute's A2 sequence, completion of which is an institutional graduation requirement; no separate standardized proficiency test was administered for this study. The participants were young adults aged 18-22 and were taking English courses as part of their institutional graduation requirements. Both classes followed the same A2-level curriculum and participated in the study during their regular English lessons. Because the study was conducted with two existing intact classes, participants were not randomly assigned to the feedback conditions. One class consisted of 18 students and was assigned as the control group, receiving teacher-only feedback. This group attended one-hour classes from Monday to Thursday, from 7:00 to 8:00 a.m. The second class consisted of 23 students and was assigned as the experimental group, receiving hybrid feedback combining ChatGPT-generated feedback and teacher feedback. This group attended two-hour classes on Mondays and Wednesdays, from 12:15 to 2:15 p.m. All 41 participants completed both the pretest and posttest and were included in the analyses. No participants were excluded, and there were no missing outcome data; therefore, no imputation or other missing-data procedure was required.

3.2 Instrument and scoring criteria

To measure students' grammatical accuracy, the same Grammar Accuracy Writing Test was used as the pretest and posttest, a procedure commonly employed in intervention research to measure learning gains (Mackey and Gass, 2011). The instrument consisted of 18 open-ended items designed to elicit grammatical structures appropriate for A2-level learners and aligned with the content covered during the instructional period (Supplementary Appendix 1). The test included six grammatical categories: verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions, with three items assigned to each category, as shown in Table 1.

Table 1. Structure of the grammar accuracy writing test.

Each item required students to produce one complete sentence rather than select an answer from predetermined options. This format was selected to obtain direct evidence of students' ability to produce the target grammatical structures independently (Ferris and Hedgcock, 2013; Hyland, 2003). The instrument was reviewed and approved by the academic coordination of the institution before its administration. Both groups completed the test individually in class using paper-based responses. Students were not allowed to use dictionaries, textbooks, notes, or digital tools, and 25 min were provided to complete the assessment.

A binary scoring system was used to evaluate grammatical accuracy. Each response received 1 point when the target grammatical structure was used correctly and 0 points when the target structure was incorrect, missing, or unclear. Therefore, the maximum possible score was 18 points. Separate subscores ranging from 0 to 3 were also calculated for each of the six grammatical categories. The first author, who was also the classroom teacher, scored all pretest and posttest responses using the predefined binary criteria. Because the first author taught both classes, scoring was not blinded to feedback condition. Potentially ambiguous responses were evaluated against the same predefined target-structure criteria; when the correctness of the target structure remained unclear, the response was assigned 0 points. No independent adjudicator was used. The same scoring criteria were applied at both measurement occasions. Spelling, capitalization, and punctuation errors were not included in the grammatical accuracy score unless they interfered with interpretation of the target grammatical structure. Because a single non-blinded rater scored the responses, inter-rater reliability could not be estimated.

3.3 Procedure and feedback conditions

The study was conducted over a six-week instructional period, from February 23 to April 2, 2026, during regular English classes. At the beginning of the study, both groups completed the Grammar Accuracy Writing Test as a pretest. At the end of the intervention, the same instrument was administered as the posttest under the same testing conditions. During the six weeks, both groups followed the same instructional content based on the Top Notch 2 textbook (Saslow and Ascher, 2015), covering one unit per week. To avoid excessive writing workload, students completed one handwritten writing task each week, resulting in six tasks during the intervention (Supplementary Appendices 2-7). Each task was related to the grammatical content studied in class and followed three main stages: drafting, feedback, and revision. The main difference between the groups was the feedback condition. Figure 1 summarizes the implementation procedure applied in both groups. Further details of the intervention stages are presented in Supplementary Appendix 8.

3.4 Control group: teacher-only feedback condition

Students in the control group received teacher-only feedback. On Mondays, they wrote sentences related to the content studied in class and handed in their handwritten work to the teacher. The teacher reviewed the sentences and provided written corrections focused on the grammatical structures addressed in the study. On Tuesdays, the teacher returned the worksheets with handwritten corrections, all of which were provided in English. Students reviewed and analyzed the written feedback using their textbook and class notes, while the teacher provided individual oral clarification in English and Spanish when required, depending on the complexity of the grammatical point or error. This individual clarification was also available as needed during the subsequent rewriting stage. On Wednesdays, students rewrote their original sentences by hand, taking into account the English-language written corrections and the teacher's oral clarification in English and Spanish when required, and handed in the revised worksheets to the teacher. The teacher reviewed the tasks again to identify persistent grammatical errors. On Thursdays, these persistent errors were addressed through whole-class feedback delivered in both Spanish and English, depending on the complexity of the grammatical point. This procedure was repeated in the last 10 min of every class throughout the six-week intervention.

3.5 Experimental group: hybrid feedback condition

Students in the experimental group received hybrid feedback combining ChatGPT-generated feedback and teacher-provided feedback. Students used the default model available through ChatGPT Free. The model was not fixed by the researchers, and model identifiers were not recorded for each interaction; therefore, the intervention may have involved more than one model version. Because the availability of the ChatGPT model may vary over time, the study refers to the tool as the “free version of ChatGPT” rather than attributing the intervention to a single, fixed model. On Mondays, students completed the weekly handwritten writing task during the final 25 min of class. During the first 8 min, they wrote their sentences independently without using textbooks, notes, or AI tools. After completing the initial draft, students photographed their handwritten sentences and uploaded them to ChatGPT together with a structured prompt designed for A2 learners (Supplementary Appendix 9). The prompt instructed ChatGPT to provide grammar feedback only, using simple and clear English and focusing on verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions. Accordingly, the AI-generated feedback that students received and read was in English. The prompt also instructed the tool not to change students' ideas, add new vocabulary, or rewrite the text for them.

Students then had approximately 10 min to read the English-language AI feedback, compare it with their original sentences, and identify suggested corrections related to the six target grammatical categories. During this review, students also consulted their textbook and class notes, while the teacher provided individual oral clarification in English and Spanish when required, depending on the complexity of the grammatical point or error. Because the prompt prohibited complete rewriting, students were expected to apply relevant corrections themselves rather than copy an AI-written text. Students used the final 7 min to rewrite their sentences by selecting and applying relevant corrections from the AI feedback together with any teacher clarification they had received. Students then handed their revised worksheets to the teacher, who reviewed them to identify persistent grammatical errors for the Wednesday whole-class feedback. Thus, the experimental condition represented a teacher-mediated hybrid process rather than AI-only feedback. On Wednesdays, these recurrent errors were addressed through whole-class teacher feedback in both Spanish and English, depending on the complexity of the grammatical point, during the final 15 min of class. This procedure was repeated across the six weeks of the intervention.

Initial written corrective feedback was provided entirely in English in both groups: handwritten teacher corrections in the teacher-only condition and ChatGPT-generated feedback in the hybrid condition. While reviewing this feedback, students in both groups consulted their textbook and class notes and received individual teacher clarification in English and Spanish, their first language, as needed according to the complexity of the grammar or error. Subsequent whole-class teacher feedback on persistent errors was also provided in English and Spanish in both groups. Although the intervention stages were distributed differently across the week, both conditions involved a total of 40 min per week. Thus, the study compared a teacher-only feedback cycle with a teacher-mediated hybrid feedback cycle rather than teacher feedback with AI feedback in isolation.

3.6 Data analysis

The data were analyzed using descriptive and inferential statistics to examine changes in students' grammatical accuracy within and between the two feedback conditions. Descriptive statistics, including the mean, median, and standard deviation, were calculated for pretest, posttest, and gain scores. Gain scores were calculated as posttest minus pretest for each participant. All 41 cases had complete pretest and posttest data. The distribution of total gain scores showed no significant departure from normality for either the teacher-only group (Shapiro–Wilk W = 0.91, p = 0.077) or the hybrid-feedback group (W = 0.93, p = 0.119). A median-centered Levene test also supported homogeneity of gain-score variances (p = 0.307).

To address the first research question, paired-samples t-tests were conducted separately for the teacher-only and hybrid-feedback groups. Within-group standardized mean changes were identified as Cohen's d_z and were calculated as the mean paired difference divided by the standard deviation of the paired differences. To address the second research question, gain scores were calculated separately for the six grammatical categories. Because each category consisted of only three items and produced discrete scores with limited variability, two-sided Mann–Whitney U tests were used to compare category-level gains. Holm's sequential procedure was applied across the six category tests to control the familywise error rate. To address the third research question, an independent-samples t-test compared total gain scores, and pooled-standard-deviation Cohen's d quantified the between-group difference. The contrast was consistently defined as hybrid feedback minus teacher-only feedback.

As a sensitivity analysis for baseline imbalance, posttest scores were also analyzed with an ANCOVA that included feedback group and pretest score as predictors. The homogeneity-of-regression-slopes assumption was evaluated with a group×pretest interaction and was supported, F(1, 37) = 0.22, p = 0.641. ANCOVA residuals showed no significant departure from normality (Shapiro–Wilk p = 0.835) or homogeneity of variance (median-centered Levene p = 0.168). Two-sided tests used α = 0.05. Ninety-five percent confidence intervals were reported for mean changes, between-group differences, and standardized effects. Analyses were conducted in Python 3.12.13 using NumPy 2.3.5 and SciPy 1.17.0.

3.7 Ethical considerations

This study followed established ethical principles for educational research involving human participants, in accordance with the guidelines of the American Educational Research Association (2011). Written institutional authorization was obtained before the study was conducted to ensure that the research procedures were consistent with institutional policies (Supplementary Appendix 10). Moreover, the instrument was reviewed and approved by the academic coordination of the institution (Supplementary Appendix 11). Furthermore, all participants received information about the purpose of the study, the procedures involved, and the voluntary nature of their participation. Written informed consent was obtained from all 41 participants before data collection (Supplementary Appendix 12). Since all participants were between 18 and 22 years old, parental or guardian consent was not required.

The intervention was conducted during regular English classes and involved activities consistent with the students' usual instructional context, including writing, feedback, and revision. These procedures were intended to minimize potential risk or disruption to participants while maintaining the normal classroom environment. At the end of the study, participants were provided with a summary of the findings in an accessible format. These procedures were followed to promote transparency, voluntary participation, and responsible educational research throughout the study.

4 Results

Before the main comparisons were conducted, the distribution of the overall gain scores was examined. The Shapiro–Wilk test showed no significant departure from normality in either the control group (W = 0.91, p = 0.077) or the experimental group (W = 0.93, p = 0.119). Levene's test also indicated that the homogeneity of variance assumption was met (p = 0.307). These results supported the use of parametric tests for the overall score comparisons.

4.1 Changes in overall grammatical accuracy

To address the first research question, pretest and posttest scores were compared within each group. Table 2 presents the descriptive statistics for overall grammatical accuracy. Both groups increased their mean scores from pretest to posttest. The control group increased from M = 7.17 (SD = 4.15) to M = 9.61 (SD = 3.84), with a mean gain of 2.44 points. The experimental group increased from M = 9.17 (SD = 3.52) to M = 10.78 (SD = 4.24), with a mean gain of 1.61 points. Paired-samples t-tests were conducted to determine whether the changes within each group were statistically significant.

Table 2. Descriptive statistics for overall grammatical accuracy.

As shown in Table 3, the teacher-only group showed a statistically significant increase from pretest to posttest, with a mean gain of 2.44 points, 95% CI [1.14, 3.75], t(17) = 3.96, p = 0.001, Cohen's d_z = 0.93, 95% CI [0.37, 1.48]. The hybrid-feedback group also showed a statistically significant increase, with a mean gain of 1.61 points, 95% CI [0.27, 2.94], t(22) = 2.50, p = 0.020, Cohen's d_z = 0.52, 95% CI [0.08, 0.95]. Thus, improvement was observed in both conditions, with a larger descriptive mean gain and standardized within-group change in the teacher-only group.

Table 3. Within-group pretest-posttest comparisons.

d_z is Cohen's standardized mean change for paired observations.

4.2 Comparison across grammatical categories

The second research question examined whether gains differed between the two feedback conditions across the following six target grammatical structures: verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions. Table 4 presents the descriptive gain scores for each grammatical category. As shown in Figure 2, the experimental group showed slightly higher mean gains in tense use (M = 0.48 vs. 0.28) and subject-verb agreement (M = 0.09 vs. 0.06). In contrast, the control group showed higher mean gains in word order (M = 0.33 vs. 0.30), articles (M = 0.28 vs. 0.00), prepositions (M = 0.72 vs. 0.26), and conjunctions (M = 0.78 vs. 0.48).

Table 4. Descriptive gain scores by grammatical category.

Two-sided Mann–Whitney U tests were conducted to compare category-level gain scores between the groups. As shown in Table 5, none of the raw comparisons reached statistical significance: tense use, U = 192.50, p = 0.701; word order, U = 222.50, p = 0.660; subject-verb agreement, U = 196.00, p = 0.751; articles, U = 244.50, p = 0.302; prepositions, U = 256.00, p = 0.183; and conjunctions, U = 242.00, p = 0.338. After applying Holm's correction across the six tests, all adjusted p-values were 1.000. In summary, although the groups showed different descriptive gain patterns across categories, the analyses detected no statistically significant advantage for either feedback condition in any of the six grammatical structures. These descriptive differences should therefore be interpreted cautiously and not as evidence that one feedback condition was more effective for a particular grammatical category.

Table 5. Between-group comparisons by grammatical category.

Holm-adjusted p-values control the familywise error rate across the six category comparisons.

4.3 Between-group differences in overall gain scores

The third research question examined whether overall improvement in grammatical accuracy differed between students who received teacher-only feedback and those who received hybrid feedback. As shown in Table 6, the teacher-only group obtained a mean gain of 2.44 points (SD = 2.62), while the hybrid-feedback group obtained a mean gain of 1.61 points (SD = 3.09). With the contrast defined as hybrid minus teacher-only, the unadjusted mean difference was −0.84 points, 95% CI [−2.68, 1.00], t(39) = −0.92, p = 0.364, Cohen's d = −0.29, 95% CI [−0.91, 0.33]. An ANCOVA sensitivity analysis of posttest scores controlling for pretest scores yielded adjusted posttest means of 10.49 [95% CI (9.13, 11.86)] for the teacher-only group and 10.09 [95% CI (8.89, 11.30)] for the hybrid-feedback group. The adjusted group difference was −0.40 points, 95% CI [−2.25, 1.45], F(1, 38) = 0.19, p = 0.664, partial η² = 0.005. The pretest covariate was associated with posttest score, B = 0.78, 95% CI [0.54, 1.02], F(1, 38) = 44.05, p < 0.001, partial η² = 0.537. In summary, the teacher-only group showed a descriptively larger unadjusted mean gain, but the estimated between-group effect was small and not statistically significant. The ANCOVA likewise detected no statistically significant difference after controlling for pretest scores. Because the confidence intervals for both estimates included zero and effects favoring either condition, these results indicate uncertainty about the comparative effect of the two conditions rather than evidence that they were equivalent.

Table 6. Between-group comparison of overall gain scores.

Gain scores were calculated as posttest minus pretest. The between-group contrast and Cohen's d were calculated as hybrid feedback minus teacher-only feedback; therefore, negative values indicate a difference in favor of the teacher-only condition. The mean difference was calculated using unrounded group means.

5 Discussion

5.1 Improvement in overall grammatical accuracy

The pretest-posttest improvement observed in both groups is consistent with previous research indicating that written corrective feedback may contribute to grammatical development when learners have opportunities to notice, process, and apply corrections during revision (Bitchener et al., 2005; Bitchener and Ferris, 2012; Shintani et al., 2014). In the present study, both groups followed repeated cycles of drafting, feedback, and revision during the six-week intervention. Their gains may therefore reflect these opportunities to review grammatical choices, together with regular classroom instruction. However, familiarity with the repeated test may have contributed to gains in both groups. Scoring by one non-blinded rater also leaves open the possibility of scoring bias despite the predefined binary criteria. These limitations restrict attribution of the observed gains to the feedback routines alone.

Teacher support may have contributed to improvement in both groups. Previous research has shown that teachers can provide contextualized explanations, identify recurrent errors, and adapt corrective information according to learners' linguistic needs (Ferris, 2010; Lee, 2017). Similarly, Thi and Nikolov (2022) found that automated and teacher feedback can complement one another, but teacher feedback remains important for explaining and contextualizing linguistic problems that automated systems may identify only at the surface level.

Within the hybrid routine, ChatGPT's immediate response may have helped students notice errors before receiving teacher clarification. This possibility is consistent with research describing AI-generated feedback as particularly responsive to sentence-level and linguistic features (Crosthwaite and Sun, 2026; Mali, 2026; Shi and Aryadoust, 2024).

5.2 Feedback across specific target grammatical categories

The category-level gains showed different descriptive patterns across the two groups, but none of the six comparisons reached statistical significance after Holm correction. These patterns should not be interpreted as evidence that either feedback condition was more effective for a particular grammatical structure. Median gain scores were zero for most categories, particularly in the hybrid group. However, each category contained only three items, which may have limited sensitivity to small changes in performance. Variation in responses to feedback has been reported in previous research, where effects of feedback may differ according to the grammatical feature, the type of correction, and learners' ability to understand and apply the feedback they receive (Bitchener et al., 2005; Boggs, 2019; Shintani and Ellis, 2013).

For A2 learners, immediate correction may help make an error visible, but noticing an error does not necessarily mean that learners have fully understood the grammatical rule or can transfer the correction to later tasks. This distinction is consistent with recent research on AI-generated feedback, which suggests that automated feedback can support surface-level revision while still requiring learner engagement and teacher mediation for deeper understanding (Crosthwaite and Sun, 2026; Mali, 2026; Teng, 2024b).

5.3 Hybrid feedback compared with teacher-only feedback

The central finding is that the estimated between-group difference in overall grammatical accuracy gains was small and not statistically significant. This result is particularly relevant because the comparison was not between AI feedback and no feedback, but between hybrid feedback and an already structured teacher-feedback condition. Both conditions included initial written corrective feedback in English. Students reviewed the feedback using their textbook and class notes and received individual teacher clarification in English and Spanish as needed. Both groups also received bilingual whole-class feedback on persistent errors. The initial written feedback differed in source and timing: handwritten teacher corrections in the following class for the teacher-only group and immediate ChatGPT-generated feedback in the hybrid group. Accordingly, the comparison concerns two teacher-mediated feedback cycles.

Recent research presents mixed evidence regarding whether combining AI and teacher feedback produces an additional advantage. Khojasteh et al. (2025) reported stronger writing performance among learners receiving hybrid feedback than among those receiving teacher e-feedback or AI feedback separately. Similarly, Zhang et al. (2025) found advantages for hybrid feedback, particularly in more complex aspects of L2 writing. These findings suggest that combining AI-generated feedback with teacher guidance may be beneficial in some instructional contexts.

However, other studies provide a more cautious picture. Ding et al. (2025) found that combined ChatGPT and teacher feedback could produce advantages in specific aspects of writing, including grammar, but did not demonstrate significantly greater overall writing improvement. Lo, Chan, and Wong (2025) reported that teacher feedback produced greater improvement than hybrid and AI-only feedback. Alnemrat et al. (2025) found no statistically significant difference between AI-generated and teacher-generated feedback. Taken together with the present findings, this evidence suggests that the effectiveness of hybrid feedback may depend on the learners, task, proficiency level, duration of intervention, and the way AI is integrated into the feedback process.

In the hybrid condition, teacher clarification in English and Spanish may have helped students understand the AI-generated feedback and apply it during revision, thereby contributing to the observed gains. However, individual teacher clarification and whole-class feedback were provided in both groups. The hybrid group's improvement cannot be attributed to ChatGPT alone because teacher support and opportunities for review and revision were also part of the intervention. For A2 learners, additional feedback may not automatically produce greater learning if students still require teacher support to understand and transfer corrections. This interpretation is consistent with Han and Li (2024), who emphasize that the value of ChatGPT-supported teacher feedback depends on teacher mediation rather than on the AI tool functioning independently.

The hybrid-feedback group began the study with a higher pretest mean than the teacher-only group (9.17 vs. 7.17). The ANCOVA sensitivity analysis controlling for pretest score did not detect a statistically significant adjusted group difference. Because participants were not randomly assigned, and only one intact class represented each condition, adjustment for pretest scores cannot separate the potential influences of class membership, class schedules, and feedback timing from the feedback condition. The small sample limits precision, and the confidence intervals include effects favoring either condition; therefore, statistical nonsignificance does not establish equivalence. Accordingly, the larger unadjusted gain in the teacher-only group should not be interpreted as definitive evidence that teacher feedback is universally more effective than hybrid feedback.

Overall, the findings support a complementary rather than replacement-oriented view of AI-generated feedback. ChatGPT may provide immediate corrective support and additional opportunities for students to review grammatical errors, but teacher mediation appears to remain central to the feedback process. The results therefore do not suggest that hybrid feedback lacks pedagogical value; rather, they indicate that its added value should not be assumed simply because an AI component is included. For lower-proficiency learners, the quality, clarity, and instructional integration of feedback may be more important than the number of feedback sources involved.

6 Limitations and future research

Although both groups received a total of 40 min of intervention-related activities per week, their class schedules and the distribution of these activities differed. The teacher-only condition was implemented during the final 10 min of class from Monday through Thursday, whereas the hybrid condition was implemented during 25 min on Monday and 15 min on Wednesday. These differences in timing and spacing may have influenced how students reviewed and applied corrections. Feedback source and immediacy also varied together: teacher corrections were returned in the following class, whereas ChatGPT feedback was received during the same session. Consequently, possible scheduling effects cannot be separated completely from the feedback conditions. The six-week intervention also assessed only short-term changes; maintenance and transfer to later writing were not established. Future research should include delayed assessments and new writing tasks.

Some items of the Grammar Accuracy Writing Test appeared to create confusion between related linguistic concepts, such as adverbs of frequency and time expressions. Although students could produce grammatically acceptable sentences, their responses did not always contain the specific grammatical structure intended by the item. Therefore, future versions of the instrument could use more specific questions to isolate individual grammatical structures while maintaining the open-ended format and the A2-level language demands of the test (Supplementary Appendix 13).

Each category consisted of three items and produced subscores ranging from 0 to 3. Therefore, category-level results should be interpreted cautiously, since small changes in grammatical performance may not always be captured by a three-item subscore. Future studies could increase the number of items within each category to improve sensitivity to small changes in grammatical performance.

The study involved a small sample of 41 students drawn from two intact classes, and participants were not randomly assigned to the feedback conditions. The groups also differed at baseline, with the hybrid-feedback group obtaining a higher pretest mean than the teacher-only group (9.17 vs. 7.17). Although the ANCOVA sensitivity analysis controlled statistically for pretest scores, this adjustment cannot eliminate unmeasured confounding associated with the intact-class design. The wide confidence intervals around the between-group estimates indicate limited precision. Therefore, the nonsignificant findings should not be interpreted as evidence of equivalence, and generalization beyond similar on-site A2 EFL contexts should be approached cautiously. Future studies should include larger samples across multiple classes per condition and, where feasible, randomly assign classes to feedback conditions.

The study also used the same Grammar Accuracy Writing Test as the pretest and posttest. Familiarity with the items may have contributed to pretest-posttest gains in both groups, limiting the extent to which improvement can be attributed solely to the feedback interventions. Future studies could use parallel pretest and posttest forms that elicit the same grammatical structures through different prompts. In addition, the study focused specifically on grammatical accuracy. Therefore, the findings should not be interpreted as evidence of overall writing development. Future research could examine whether the effects of hybrid and teacher-only feedback differ when both sentence-level accuracy and higher-order writing features are considered.

A further limitation concerns the combined nature of the hybrid-feedback condition. Both groups received individual teacher clarification and whole-class feedback in English and Spanish. In the hybrid condition, this support accompanied the initial written feedback generated by ChatGPT. This teacher mediation may have contributed to how students understood and applied the AI-generated corrections; therefore, the observed outcomes cannot be attributed to ChatGPT alone. The findings should therefore be interpreted as a comparison of complete teacher-mediated feedback routines. Future studies could standardize and systematically record teacher clarification or use additional comparison conditions to examine the separate and combined contributions of AI-generated and teacher-provided feedback.

In addition, all pretest and posttest responses were scored by the first author, who was also the classroom teacher and therefore knew each participant's feedback condition. Although predefined binary scoring criteria were used, reliance on a single non-blinded rater may have introduced scoring bias, and inter-rater reliability could not be estimated. The absence of systematic feedback-use data also limits explanations of individual improvement and how students accepted, rejected, or applied feedback. Future research could use blinded independent scoring, report inter-rater agreement, and combine quantitative performance measures with learner interviews, questionnaires, revision analysis, or feedback-use records. Further studies could also investigate different sequences of AI and teacher feedback and the extent to which learners' proficiency level influences their ability to use AI-generated feedback.

7 Conclusion

This study compared hybrid feedback (ChatGPT-generated and teacher-provided feedback) with teacher-only feedback in relation to the grammatical accuracy of A2 EFL learners in an on-site technical higher education context in Quito, Ecuador. Students' overall grammatical accuracy improved from pretest to posttest in both groups. The teacher-only group showed a descriptively larger mean gain and a larger within-group effect size, while the hybrid-feedback group also demonstrated statistically significant improvement, although with a more moderate effect size. These findings indicate that grammatical accuracy improved under both feedback conditions during the six-week intervention.

When the two feedback conditions were compared across the six grammatical structures (verb tense usage, word order, subject-verb agreement, articles, prepositions, and conjunctions), different descriptive patterns were observed. The hybrid-feedback group showed slightly higher gains in tense use and subject-verb agreement, whereas the teacher-only group showed higher gains in word order, articles, prepositions, and conjunctions. However, none of these differences reached statistical significance after Holm's correction. Therefore, the findings do not indicate a consistent advantage for either feedback condition at the grammatical-category level.

Similarly, neither the unadjusted gain-score comparison nor the ANCOVA sensitivity analysis controlling for pretest score found a statistically significant difference between feedback conditions. The unadjusted between-group effect was small (d = −0.29), and the adjusted mean difference was −0.40 points. Nevertheless, the confidence intervals were sufficiently wide to include effects favoring either condition; therefore, the findings do not establish equivalence.

Overall, the study did not detect greater short-term gains when ChatGPT was used as the source of the initial written feedback instead of handwritten teacher corrections. Because both conditions included individual and whole-class teacher mediation in English and Spanish, the comparison concerned two teacher-supported feedback cycles and did not isolate the independent effect of ChatGPT. ChatGPT-generated feedback may offer immediate written support within teacher-mediated routines, but its added contribution to learning remains uncertain. These conclusions should be interpreted cautiously given the small sample, intact-class design, baseline differences, differences in class schedules and feedback timing, repeated use of the same assessment, and reliance on a single non-blinded rater.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feduc.2026.1978336/full#supplementary-material

Anonymized participant-level scores and statistical calculations for overall grammatical accuracy, category-level gains, and the ANCOVA sensitivity analysis. Supplementary material at the source

Footnotes

Edited by: Nazime Tuncay, North Cyprus, Cyprus

Reviewed by: Efstathios Stefos, National University of Education, Ecuador

Hồng Nguyễn, Can Tho University, Vietnam