Comparing AI-assisted and teacher-led reading strategy instruction in an EFL context: a quasi-experimental study

Abstract

Grounded in a metacognitive and distributed-scaffolding framework, this quasi-experimental study examined classroom-level patterns associated with two configurations of reading strategy instruction and with business-as-usual instruction in a university EFL context. Sixty undergraduate students enrolled in an advanced reading course at a public university in Jordan participated in the study. To preserve natural classroom composition, three intact course sections with 20 students each were assigned at the class level to one of three conditions: teacher-mediated AI-assisted strategy instruction, teacher-led strategy instruction, or business-as-usual instruction without explicit strategy training. The AI-assisted section used ChatGPT as a scaffold for previewing, predicting, monitoring, questioning, inferencing, and summarizing within a technology-equipped classroom and with one short weekly AI-supported task; the teacher-led section addressed the same strategies through instructor modeling and guided practice; the comparison section followed the regular course routine. Reading comprehension was measured with an adapted 20-item, 60-point test, and metacognitive awareness was measured with a study administration version of the Metacognitive Awareness of Reading Strategies Inventory. Descriptive statistics were the primary analytic lens. Student-level ANCOVA and t-test results are reported as exploratory summaries because each condition was represented by a single intact section. The two explicit-strategy sections showed stronger reading-comprehension patterns than the business-as-usual section, and the AI-assisted section showed the highest adjusted posttest mean. For metacognitive awareness, both explicit-strategy sections improved from pretest to posttest, and the AI-assisted section showed the largest descriptive gain. The findings suggest that a teacher-managed AI-supported instructional package may extend explicit strategy instruction without displacing teacher judgment, but they should be interpreted as section-level comparative evidence rather than as isolated treatment effects. The study contributes a semester-long, classroom-level comparison in university EFL reading and clarifies how AI can be positioned as a complement to explicit strategy teaching.

1 Introduction

Second-language reading is a complex, goal-directed activity that requires readers to coordinate lexical access, syntactic parsing, inferencing, and discourse-level integration (Grabe and Yamashita, 2022; Janzen and Stoller, 1998; Koda, 2005). Readers who comprehend successfully do not rely on decoding alone. They plan before reading, monitor during reading, and evaluate the adequacy of their understanding after reading (Carrell, 1985; Carrell et al., 1989; Jiménez et al., 1995). Within this perspective, metacognitive strategy use is not an optional add-on. It is a central part of effective academic reading in EFL settings (Block, 1986; O’Malley and Chamot, 1990).

A recurring problem in the EFL reading literature is that many learners underuse higher-order strategies and depend instead on local support moves such as translation, dictionary consultation, and sentence-by-sentence decoding (Block, 1986; Carrell, 1985). More proficient readers tend to show greater flexibility in their strategic behavior, whereas less proficient readers often display weaker monitoring and less adaptive strategy selection (Green and Oxford, 1995; Jiménez et al., 1995). This concern becomes sharper in digital environments. University students report lower levels of planning and monitoring when reading online than when reading in print, which suggests that digital reading may require additional instructional scaffolding rather than less (Anggia and Habók, 2024).

Building on this premise, explicit strategy instruction has remained one of the most robust lines of intervention in reading pedagogy. Research on modeling, guided practice, gradual release, and reciprocal teaching has repeatedly shown that reading comprehension improves when strategic processes are taught directly rather than left implicit (Duffy, 1993; Gaskins et al., 1994; Palincsar and Brown, 1984; Pearson and Gallagher, 1983; Pressley and El-Dinary, 1997; Rosenshine and Meister, 1994). Recent syntheses and quasi-experimental studies continue to support this conclusion across diverse contexts (Ardasheva et al., 2017; Fathi and Afzali, 2020; Rogiers et al., 2020; Yapp et al., 2021). Strategy-oriented support has also shown positive effects beyond reading, including listening tasks that require selective attention and note-taking, which reinforces the broader pedagogical value of metacognitive instruction in EFL learning (Al-Ghazo, 2023).

At the same time, the strengths of teacher-led strategy instruction do not remove its practical constraints. Instructors must distribute attention across an entire class, manage time pressure, and balance whole-class explanation with individualized support. In such settings, not every learner receives immediate prompts at the exact point of breakdown, nor can every learner rehearse the same strategy with the same intensity across repeated tasks. These classroom realities have motivated interest in digital scaffolds that can extend, rather than replace, teacher guidance (Hinkel, 2005; Macaro, 2006).

Computer-assisted language learning has provided several foundations for this line of work. Technology-mediated environments can prompt learners to anticipate content, notice textual cues, rehearse inferencing, and reflect on comprehension while reading is still in progress (Gorsuch and Taguchi, 2010; Shih and Reynolds, 2015). Bibliometric evidence also indicates that AI-related language learning research has expanded rapidly, although designs, constructs, and instructional claims remain uneven across the literature (Mohsen et al., 2025). This expansion has been accompanied by growing evidence that AI can support a range of language-learning outcomes, including reading input, writing feedback, self-regulated learning, and metacognitive awareness (Aldamen et al., 2025; Alnemrat et al., 2025; Mehmood et al., 2025; Namaziandost, 2025).

Reading-specific studies point in a promising direction, but the evidence base is still consolidating. Digital and computer-based interventions have been associated with stronger inferential comprehension when strategy support is built into the task design (Serrano-Mendizábal et al., 2023). Chatbot-supported reading work has reported improved performance and reduced reading anxiety in foreign-language contexts (Zheng, 2024). Other recent studies suggest that AI-mediated reading environments may support self-regulated engagement and reading achievement (Allehyani et al., 2025; Shafiee Rad, 2025). Even so, many studies either examine AI as a stand-alone tool, focus on short interventions, or compare an enriched treatment with a poorly specified baseline. The question is therefore not simply whether AI is helpful. A more useful question is how AI functions when it is embedded within teacher-managed strategy instruction.

This study addresses that question by comparing two instructional configurations rather than treating AI and teachers as competing forces. The first configuration involved teacher-mediated AI-enhanced strategy instruction in which ChatGPT was used as a scaffold within a structured reading lesson. The second involved teacher-only explicit strategy instruction using the same texts, the same strategy targets, and the same instructor. A third business-as-usual condition followed the regular course routine without explicit strategy instruction. This framing is theoretically important because it positions AI as one component within a distributed scaffolding system rather than as a replacement for pedagogical expertise (Carrell and Eisterhold, 1983; Macaro, 2006; O’Malley and Chamot, 1990).

The study contributes to the literature in four more modest and clearly bounded ways. First, it offers a semester-long comparison in a university EFL context rather than a short-term pilot. Second, it compares a teacher-mediated AI-supported instructional package with teacher-only explicit strategy instruction and with a specified business-as-usual section. Third, it examines both reading comprehension and metacognitive awareness. Fourth, it specifies the business-as-usual condition and the AI prompting routine in enough detail to support replication, while treating the findings as classroom-level comparative evidence rather than as isolated treatment effects.

2 Literature review

2.1 Strategic reading and metacognitive awareness

Strategic reading research has long treated comprehension as an active process in which readers coordinate prior knowledge, textual cues, and metacognitive regulation (Carrell and Eisterhold, 1983; Koda, 2005). Readers who are strategically aware preview the text, generate questions, monitor breakdowns, repair misunderstandings, and evaluate the adequacy of a developing interpretation (Carrell et al., 1989; Mokhtari and Reichard, 2002). These behaviors are closely aligned with broader accounts of language learning strategies and self-regulation in second-language acquisition (Hinkel, 2005; Macaro, 2006; O’Malley and Chamot, 1990).

The contrast between more and less successful readers has been especially informative. Skilled readers appear to orchestrate strategies in coordinated patterns, whereas less successful readers often rely on isolated support moves without integrating them into a larger reading plan (Block, 1986; Jiménez et al., 1995). In EFL settings, this distinction is associated with stronger comprehension and greater flexibility in adapting strategies to genre, task, and difficulty (Grabe and Yamashita, 2022; Green and Oxford, 1995). Digital reading introduces an additional layer of complexity because online environments may fragment attention and reduce deliberate planning if strategic routines are not explicitly supported (Anggia and Habók, 2024).

2.2 Explicit strategy instruction in EFL reading

Explicit strategy instruction has been shown to improve reading comprehension when it combines explanation, modeling, guided practice, and reflection. Foundational work on gradual release and reciprocal teaching demonstrated that comprehension processes can be taught in ways that make expert thinking visible to learners (Palincsar and Brown, 1984; Pearson and Gallagher, 1983). Subsequent classroom studies and reviews reinforced the instructional importance of teacher modeling, scaffolded dialogue, and repeated application across texts (Duffy, 1993; Gaskins et al., 1994; Pressley and El-Dinary, 1997; Rosenshine and Meister, 1994).

More recent evidence continues to support these findings. Fathi and Afzali (2020) reported significant gains in EFL reading comprehension after explicit instruction in skimming, scanning, inferencing, summarizing, and predicting. Rogiers et al. (2020) found that strategy-oriented instruction yielded both immediate and sustained improvements, while Yapp et al. (2021) showed in a meta-analysis that strategy interventions can generate gains beyond typical developmental progress. Ardasheva et al. (2017) likewise found that strategy instruction produces positive effects across second-language and self-regulated learning outcomes. Cross-skill evidence also suggests that strategy instruction can improve performance in adjacent domains such as listening, which strengthens the claim that metacognitive support has broad value in EFL pedagogy (Al-Ghazo, 2023).

2.3 AI-mediated scaffolding and related studies

Digital tools may strengthen strategy instruction when they are designed to prompt attention at the moment a learner needs support. Earlier CALL research suggested that technology-enhanced instruction can guide learners toward more deliberate engagement with text, especially when prompts and interaction routines are clearly tied to comprehension goals (Gorsuch and Taguchi, 2010; Shih and Reynolds, 2015). More recent AI-related studies indicate that digital scaffolds may support reading performance, self-regulation, and learner engagement, although the theoretical framing is still developing (Mohsen et al., 2025).

Several recent studies are especially relevant to the present research design. Serrano-Mendizábal et al. (2023) found that strategy-rich computer-based reading interventions were associated with stronger inferential comprehension than less strategic digital support. Zheng (2024) reported that chatbot use improved foreign-language reading performance and reduced reading anxiety among secondary students. Shafiee Rad (2025) linked AI intervention in L2 reading to engagement and self-regulated learning, while Allehyani et al. (2025) documented performance gains in AI-supported English reading. Evidence from neighboring language domains also indicates that AI support can shape learner performance when it is pedagogically structured. Aldamen et al. (2025) found that AI-mediated reading input was associated with oral proficiency growth, Alnemrat et al. (2025) compared AI and teacher feedback in EFL writing, Namaziandost (2025) examined AI-enhanced language learning and metacognitive awareness, and Mehmood et al. (2025) linked AI-assisted metacognitive strategy use to stronger self-regulated learning outcomes.

What remains less clear is how AI functions when it is not treated as an autonomous tutor, but rather as a scaffold embedded in teacher-managed instruction. Many recent studies compare a digitally enriched condition with a broad or weak baseline, which makes it difficult to determine whether the advantage comes from AI itself, from explicit strategy teaching, or from increased attention to task design. The present study addresses this gap by comparing teacher-mediated AI-enhanced strategy instruction with teacher-only explicit strategy instruction and a clearly defined business-as-usual condition.

Recent quasi-experimental studies also help delimit the contribution of the present research. Alazemi (2024) examined AI-integrated formative assessment and emphasized reading progress together with online academic enjoyment, personal best goals, and academic mindfulness. Feng and Wang (2023) compared human-AI robot interaction with paper-book reading among primary school bilingual learners, which places the population, literacy goals, and learning environment in a different instructional space. Lin et al. (2025) investigated AI-supported pre-reading scaffolding and connected it to reading comprehension, motivation, and attitude. By contrast, the present study focused on a university EFL course, used a semester-long design, compared teacher-mediated AI-enhanced strategy instruction directly with teacher-only explicit strategy instruction, and included a clearly specified business-as-usual condition. The distinctive contribution therefore lies not in the broad claim that AI can help reading, but in the narrower claim that AI may extend explicit strategy instruction when it is embedded within teacher-managed scaffolding across a full academic semester.

3 Theoretical and conceptual framework

3.1 Theoretical framework

This study is grounded in two complementary theoretical strands: a metacognitive view of second-language reading and a distributed view of instructional scaffolding. In the first strand, comprehension develops more successfully when readers make their strategic work visible to themselves. They preview, set a purpose, monitor breakdowns, recover meaning, and evaluate whether an interpretation is supported by the text. This view has deep roots in L2 reading research and strategy scholarship, where planning, monitoring, and evaluation are treated as central features of effective reading rather than optional extras (Carrell et al., 1989; Macaro, 2006; Mokhtari and Reichard, 2002; O’Malley and Chamot, 1990). Recent AI research has extended this logic by showing that AI tools can support self-regulated language learning when their guidance is structured, bounded, and linked to clear pedagogical aims (Chang and Sun, 2024).

The second strand concerns how support is delivered. Explicit strategy instruction assumes that reading strategies develop more reliably when they are named, modeled, rehearsed, and reflected upon across multiple tasks (Duffy, 1993; Palincsar and Brown, 1984; Pearson and Gallagher, 1983). A distributed-scaffolding perspective adds that this support does not come from a single source. It can be shared across the instructor, the learner, the task, and a mediating tool. Under that view, an AI interface is not treated as an autonomous tutor. It is best understood as an additional channel for contingent prompting inside a teacher-managed lesson. That framing is consistent with recent work on AI-supported reading scaffolding and on the need to keep generative AI support reliable, bounded, and instructionally supervised (Lin et al., 2025; Qian et al., 2026).

3.2 Conceptual framework

The conceptual framework for the present study links instructional configuration to two outcome domains through the amount, immediacy, and boundedness of strategic support available during reading. In the teacher-led condition, support was delivered through instructor modeling, questioning, guided discussion, and written reflection. In the AI-assisted condition, the same strategy targets were retained, but supervised interaction with ChatGPT added a teacher-managed prompting channel intended to increase prompt density and opportunities for strategy rehearsal. In this study, that AI-assisted configuration also included real-time device access during guided practice and one short weekly AI-supported task, so it is best understood as a composite instructional package rather than a pure tool-only contrast. In the business-as-usual condition, students completed the course through the regular instructional routine without explicit strategy labeling or repeated metacognitive rehearsal. This framing is consistent with recent work on AI-supported self-regulated language learning, AI-supported pre-reading scaffolding, and reliable generative AI scaffolding under teacher-managed boundaries (Chang and Sun, 2024; Lin et al., 2025; Qian et al., 2026).

This framework led the study to expect stronger growth in reading comprehension and reading-strategy awareness under explicit strategy instruction than under business-as-usual instruction. It also led to a narrower expectation about AI: not that AI would replace teaching, but that a teacher-managed AI-supported package might increase prompt density, immediacy, and opportunities for individualized rehearsal within the local course design. Because the AI-assisted section also included technology access and weekly AI-supported practice, the framework supports cautious section-level comparison of instructional configurations in this semester-long university EFL setting while leaving mechanism-level explanations open to further process research.

4 Methods

4.1 Research design

The study employed a quasi-experimental pretest-posttest comparative classroom design with three conditions. Because students were enrolled in existing course sections before the study began, random assignment at the individual level was not feasible. Instead, three intact sections of the same advanced reading course were assigned at the class level to one of three instructional conditions: teacher-mediated AI-assisted strategy instruction, teacher-led strategy instruction, or business-as-usual instruction. Each section included 20 students, which yielded a total sample of 60 participants. This class-level allocation preserved the natural classroom composition while allowing parallel instruction across sections. At the same time, because only one section represented each condition, section and condition were inseparable in the design. The study therefore supports cautiously interpreted section-level comparisons rather than strong student-level causal inference about isolated treatment effects.

The design was selected for two reasons. First, it aligned with the university's registration procedures and reduced disruption to the semester schedule. Second, it allowed the same instructor to teach all three sections with the same core reading materials, thereby reducing between-teacher variability. These strengths improve procedural comparability, but they do not remove the interpretive limitation created by one section per condition. Accordingly, the study is framed as a comparative classroom study anchored in the theoretical and conceptual framework outlined in Section 3.

4.2 Participants and setting

The study used a nonprobability convenience census of the accessible population, namely all students enrolled in the three available sections of the same advanced reading course during the target semester. Because enrollment had already been completed through the normal university registration process, no individual recruitment beyond those sections occurred and no student-level random assignment was feasible. The final sample included 60 undergraduate students aged 20–22. They were distributed across three intact sections of 20 students each. Placement records, prior academic performance, and instructor evaluation indicated that the students were functioning at approximately the B2 level in reading. Supplementary Appendix F provides the formal sampling statement, the recruitment flow, and the section-level allocation details. Because each instructional condition was represented by only one section, section-specific influences such as peer composition, meeting time, and classroom ecology cannot be disentangled from condition.

The intervention was conducted over a 16-week academic semester from October 2, 2025, to January 31, 2026. Each section met for 90 min twice per week, producing 32 class meetings. The AI-assisted section met in a technology-equipped classroom with stable internet access and individual devices. The teacher-led and business-as-usual sections met in comparable classrooms. The same instructor taught all three sections and followed the same syllabus, reading load, and assessment schedule across the semester. The AI-assisted section also completed one short weekly AI-supported task outside class, so the AI condition should be interpreted as a bundled instructional configuration rather than as a pure AI-only contrast.

The study was approved by the Institutional Review Board at the University of Jordan (Approval No. 437/2025). All participants received a written information sheet and consent form before data collection. Participation was voluntary, nonparticipation carried no academic penalty, and students were informed that participation or withdrawal would not affect course grades or academic standing. All score files were anonymized before analysis. The ethics approval statement and participant consent materials are reproduced in Supplementary Appendix G.

4.3 Instruments

4.3.1 Reading comprehension test

Reading comprehension was assessed through an adapted 20-item test based on the TOEFL Junior Standard reading section. To align with the course's assessment rubric, each item was weighted to produce a 60-point total score. The test targeted literal comprehension, inferencing, vocabulary in context, and synthesis across text segments. Five subject-matter experts reviewed the adapted instrument for content coverage and appropriateness. A pilot administration with 30 comparable students produced a Cronbach's alpha of 0.78, which indicated acceptable internal consistency (Dörnyei and Taguchi, 2009). Parallel forms were used at pretest and posttest to reduce practice effects. The study-use versions of the pretest and posttest forms are reproduced in Appendices A and B.

4.3.2 Metacognitive awareness of reading strategies inventory

Metacognitive awareness was measured with a study administration version of the Metacognitive Awareness of Reading Strategies Inventory, or MARSI (Mokhtari and Reichard, 2002). The instrument contains 30 items rated on a five-point Likert scale and addresses global strategies, problem-solving strategies, and support strategies. Because the course and study tasks were conducted in English, the inventory was administered in English. Minor wording adjustments were reviewed by subject-matter experts to ensure clarity and contextual appropriateness for the local EFL setting. Pilot testing yielded a Cronbach's alpha of .83, which indicated satisfactory reliability (Dörnyei and Taguchi, 2009). The English study administration version used in this project is reproduced in Supplementary Appendix C.

4.4 Instructional procedures

All three sections followed the same advanced reading syllabus and used the same core texts. The instructional contrast lay in the form of scaffolding rather than in the reading content itself. Across the semester, instruction addressed previewing, predicting, questioning, inferencing, monitoring comprehension, summarizing, and evaluating authorial claims. In every section, the teacher introduced the week's target text and learning objective. After this common entry point, the sections diverged in the way strategic support was delivered.

4.4.1 AI-assisted strategy instruction

The AI-assisted section used the ChatGPT web interface available during the semester in a technology-equipped classroom with stable internet access and individual devices. The instructor prepared a prompt bank before the intervention and maintained the same prompt architecture throughout the term. Each AI-guided interaction began with an instructional frame directing the tool to function as a reading-strategy coach rather than an answer generator. The opening prompt instructed the model to guide students by asking questions, prompting evidence-based reasoning, and encouraging strategy selection while avoiding direct answers to worksheet or test items. Students interacted with the AI during guided practice segments of the lesson, usually for 15–20 min per session, and completed one shorter AI-supported task outside class each week. Accordingly, the AI-assisted condition combined teacher-mediated AI prompting, immediate device access during guided practice, and additional weekly AI-supported practice.

The prompt bank was organized around six recurring moves: previewing, which asked students to identify title cues, headings, visuals, and likely themes before reading; predicting, which prompted them to generate tentative expectations and state the textual basis for them; monitoring, which focused on identifying points of comprehension breakdown and proposing repair moves; questioning, which required students to produce literal and inferential questions tied to textual evidence; inferencing and vocabulary, which emphasized inferring implied meaning and interpreting unfamiliar words from context; and summarizing and evaluating, which asked students to condense the text and judge the strength of claims or evidence.

Teacher mediation remained central in this condition. The instructor modeled one example at the start of each strategy cycle, monitored student interactions with the AI, redirected answer-seeking behavior, and concluded the lesson with a whole-class debriefing focused on why certain prompts supported comprehension more effectively than others. Students were required to paste only short text excerpts or paraphrased task segments into the AI interface. They were not permitted to upload full worksheets or ask the tool to complete comprehension answers. When such requests occurred, the instructor redirected students to the strategy prompt bank and asked them to restate the question in metacognitive terms, such as “What clues should I examine here?” or “Which strategy fits this difficulty?”

4.4.2 Teacher-led strategy instruction

The teacher-led section addressed the same strategies, texts, and learning objectives without AI support. The instructor explicitly named each strategy, demonstrated its use through think-aloud modeling, and guided students through pair and whole-class discussion. Students completed paper-based or teacher-led guided-practice tasks that paralleled the AI condition in content and difficulty. In this section, scaffolding was provided through teacher questioning, peer discussion, short written reflections, and follow-up feedback during and after reading.

4.4.3 Business-as-usual instruction

The business-as-usual section followed the regular course routine without explicit strategy training and without AI support. Students completed vocabulary preparation, silent reading, teacher-led explanation of difficult sentences or ideas, comprehension questions, and brief follow-up writing or discussion tasks. The teacher answered questions about content and language, but did not teach or rehearse a formal set of reading strategies, use a prompt bank, or require students to articulate strategy choice during the lesson. This condition therefore represented the standard advanced reading class in the department rather than a minimal-treatment control.

To strengthen procedural transparency, the semester unfolded in the following sequence. Table 1 summarizes the semester phases and the condition-specific realization of each phase. Weeks 1 and 2 focused on orientation, previewing, goal setting, and baseline reflection on reading difficulties. Weeks 3 and 4 focused on prediction and question generation before and during reading. Weeks 5 and 6 targeted comprehension monitoring and repair strategies when understanding broke down. Weeks 7 and 8 emphasized inferencing, contextual vocabulary interpretation, and tracking textual clues. Weeks 9 and 10 focused on summarizing paragraph-level and text-level meaning. Weeks 11 and 12 addressed integrating ideas across longer texts and distinguishing major from minor information. Weeks 13 and 14 focused on evaluating claims, author stance, and evidence. Week 15 provided cumulative review and independent practice. Week 16 was reserved for posttesting and final reflection.

4.5 Implementation fidelity and replicability

The intervention documentation was designed to make the treatment visible enough for replication. The AI-assisted section followed a stable prompt architecture across the semester rather than a changing or improvised use of the tool. The prompt bank included recurring frames for previewing, predicting, monitoring, questioning, inferencing, vocabulary inference, summarizing, evaluating, and metacognitive reflection. Each guided interaction was bounded by three standing rules: the tool functioned as a strategy coach rather than an answer generator; students could paste only short excerpts or paraphrased task segments; and direct answer requests were redirected into metacognitive prompts. Supplementary Appendix D reproduces the prompt bank and the teacher mediation protocol.

A session-level fidelity log was also maintained across the 32 meetings in each condition. The log recorded whether the target text was introduced, the planned focus was completed, teacher modeling occurred, guided practice was completed, the whole-class debrief occurred, and answer-seeking was redirected when observed. Implementation was stable across the semester. In the AI-assisted section, the only documented deviations were brief connectivity delays during guided practice in a small number of sessions. This additional documentation is included so that the editor and readers can see not only what the treatment was supposed to look like, but how it was enacted during the semester. These fidelity records document implementation stability, but they do not remove the interpretive limitation that each condition was represented by one section and that the AI-assisted section bundled AI prompting with technology access and weekly out-of-class practice.

4.6 Data analysis

Descriptive statistics were computed for all pretest and posttest measures. Reliability for the reading comprehension test and MARSI was recalculated at both testing points. All analyses were conducted using IBM SPSS Statistics 28 and were cross-checked against the analysis-ready dataset and reproducibility files. Because each instructional condition was represented by a single intact section, descriptive section-level patterns were treated as the primary evidentiary basis of the study. Student-level inferential analyses were retained as exploratory summaries of the observed patterns, not as decisive tests of isolated treatment effects.

As exploratory summaries, posttest reading comprehension was examined through analysis of covariance, with instructional condition as the independent variable and the pretest reading score as the covariate. Baseline student-level comparability was inspected with one-way analyses of variance on the pretest measures. For the exploratory reading ANCOVA, homogeneity of regression slopes was tested through the pretest-by-condition interaction, variance homogeneity was inspected with Levene's tests, and residual distribution was examined through skewness, kurtosis, Shapiro–Wilk tests, and casewise residual inspection. Table 2 summarizes the visible diagnostics, and Supplementary Appendix E reproduces the supporting tables. These checks speak to the technical fit of the exploratory student-level models, but they do not remove the design limitation that section and condition were confounded.

Changes in metacognitive awareness were examined in two steps. First, paired-samples t-tests were used within each section to summarize change from pretest to posttest at the student level. Second, between-section gain patterns were inspected through gain-score contrasts. Because Levene's test indicated unequal variances for MARSI gain scores, Welch's t-tests were used for the pairwise gain comparisons reported in the text and Supplementary Appendix E. Statistical significance was set at p < .05. Partial eta squared and Cohen's d were reported as effect-size indicators following common interpretive guidance in behavioral research (Cohen, 1988). All inferential results are interpreted cautiously and subordinated to the design-based limitation noted above.

5 Results

5.1 Preliminary analyses and descriptive statistics

The reliability analyses indicated acceptable internal consistency for both instruments. For the reading comprehension test, Cronbach's alpha was .78 at pretest and .81 at posttest. For MARSI, Cronbach's alpha was .83 at pretest and .86 at posttest. Table 3 presents the descriptive statistics for reading comprehension and metacognitive awareness across the three sections. Student-level one-way analyses of variance at pretest did not detect large baseline differences on either measure, but with one intact section per condition these checks should not be interpreted as demonstrating true condition equivalence.

Assumption checks for the exploratory student-level analyses were satisfactory for the reading ANCOVA. No baseline difference was detected at the student level for pretest reading comprehension, F(2, 57) = 0.08, p = .921, or pretest MARSI, F(2, 57) = 0.57, p = .569. The homogeneity-of-regression-slopes test was not significant, F(2, 54) = 0.68, p = .510. Residuals were approximately normal by Shapiro–Wilk, W = .978, p = .345, and no standardized residual exceeded |2.06|. Levene's test did not indicate heterogeneity of variance for reading posttest scores, W = 0.17, p = .847. For MARSI gain scores, however, Levene's test was significant, W = 14.06, p < .001, which is why the gain-score contrasts were examined with Welch's t-tests. Table 2 summarizes these diagnostics, and Supplementary Appendix E provides the supporting tables. These diagnostics support the technical fit of the exploratory models, but they do not resolve the design limitation created by one section per condition.

5.2 Effects on reading comprehension

As an exploratory student-level summary, an ANCOVA was conducted with posttest reading comprehension as the dependent variable, instructional condition as the independent variable, and pretest reading comprehension as the covariate. After baseline adjustment, the section/condition term was statistically significant, F(2, 56) = 20.99, p < .001, partial η² = .43. Within the constraints of the design, this result indicates a substantial difference in adjusted posttest reading performance across the three course sections.

The adjusted posttest means were ordered AI-assisted > teacher-led > business-as-usual. Exploratory Bonferroni-adjusted pairwise comparisons followed the same pattern: the AI-assisted section exceeded the teacher-led section by 4.54 points and the business-as-usual section by 8.59 points, while the teacher-led section exceeded the business-as-usual section by 4.05 points. All three pairwise differences were statistically significant in the student-level model. Because each condition was represented by a single intact section and the AI-assisted section bundled AI prompting with technology access and weekly AI-supported practice, these results are interpreted as section-level comparative patterns rather than as isolated treatment effects.

5.3 Effects on metacognitive strategy awareness

Within-section paired-samples t tests indicated meaningful differences in metacognitive awareness over time. The AI-assisted section showed a significant increase in MARSI scores from pretest to posttest, t(19) = 6.66, p < .001, d = 1.49. The teacher-led section also improved significantly, t(19) = 5.13, p < .001, d = 1.15. The business-as-usual section did not show a statistically significant change, t(19) = 1.15, p = .264.

Exploratory between-section gain comparisons pointed in the same general direction, but with a more cautious pattern than the reading results. Welch's tests showed that the AI-assisted section gained more than the business-as-usual section, t(22.84) = 5.99, p < .001, and that the teacher-led section also gained more than the business-as-usual section, t(24.76) = 4.35, p < .001. The AI-assisted section showed the largest descriptive gain, but the difference between the AI-assisted and teacher-led sections did not reach conventional significance, t(36.45) = 1.94, p = .060. As with the reading results, these contrasts are treated as exploratory section-level patterns rather than as clean estimates of isolated condition effects.

5.4 Summary of findings

The results show a consistent section-level pattern across the semester. The two explicit-strategy sections showed stronger reading-comprehension patterns than the business-as-usual section, and the AI-assisted section showed the highest adjusted reading mean. For metacognitive awareness, both explicit-strategy sections improved from pretest to posttest, while the business-as-usual section changed little. Because the design used one intact section per condition and the AI-assisted section represented a bundled instructional package, the findings are best read as comparative classroom evidence rather than as direct proof of isolated treatment effects or underlying mechanisms.

6 Discussion

6.1 Reading comprehension as a function of instructional configuration

The reading-comprehension results are consistent with earlier work suggesting that explicit strategy instruction can support comprehension more effectively than leaving strategic activity implicit within routine reading lessons (Duffy, 1993; Fathi and Afzali, 2020; Palincsar and Brown, 1984; Rogiers et al., 2020; Rosenshine and Meister, 1994). In the present study, the two explicit-strategy sections showed stronger reading patterns than the business-as-usual section, which is compatible with the broader view that reading develops through an interaction among the reader, the text, and the support available during task execution (Carrell and Eisterhold, 1983; Grabe and Yamashita, 2022). Because the design included one section per condition, however, this pattern should be read as classroom-level comparative evidence rather than as a definitive condition effect.

The stronger adjusted pattern in the AI-assisted section is consistent with the possibility that a teacher-managed AI-supported instructional package increased opportunities for timely prompting and individualized rehearsal. Students in that section received teacher modeling, repeated guided interactions during moments of uncertainty, immediate device access during guided practice, and one short weekly AI-supported task. This pattern resonates with recent work on digital and AI-supported reading, including studies of computer-based strategy instruction and chatbot-supported comprehension (Allehyani et al., 2025; Serrano-Mendizábal et al., 2023; Shafiee Rad, 2025; Zheng, 2024). At the same time, the present data do not permit a definitive claim about why the AI-assisted section showed the strongest pattern, because the study did not collect process data and did not isolate AI from the other bundled features of the section.

That caution matters because the AI-assisted condition was not teacher-free and was not a pure tool-only contrast. The teacher selected the texts, introduced the strategy goals, modeled appropriate use of the tool, monitored student interactions, redirected answer-seeking behavior, and led post-task reflection. In addition, the AI-assisted section differed in technology access and out-of-class practice. The more precise contrast is therefore between teacher-only explicit instruction and a teacher-managed AI-supported instructional package. This framing speaks directly to the concern that AI and teachers should not be treated as mutually exclusive instructional agents. Within the present design, AI functioned as one scaffold nested inside human pedagogy rather than as an independent instructor.

6.2 Metacognitive awareness and guided reflection

The MARSI results point in the same general direction as the reading results, but less sharply. Both explicit-strategy sections improved significantly from pretest to posttest, whereas the business-as-usual section did not. The AI-assisted section showed the largest descriptive increase, which is compatible with the idea that frequent prompting can make strategy use more visible to learners during reading. This pattern remains compatible with the broader literature linking metacognitive awareness to successful reading performance (Carrell et al., 1989; Mokhtari and Reichard, 2002; O’Malley and Chamot, 1990).

At the same time, the difference between the two explicit-strategy sections did not reach conventional significance in the exploratory gain comparison. That detail matters. It suggests that the study provides clearer classroom-level evidence for the value of explicit strategy instruction than for a decisive AI advantage on self-reported metacognitive awareness. A careful interpretation is therefore that the AI-assisted section may have afforded more visible metacognitive rehearsal, but the present dataset does not justify a stronger causal claim about direct metacognitive superiority.

The teacher-led section also improved substantially, which remains pedagogically important. Guided explanation, think-aloud modeling, and reflective discussion have long been recognized as effective means of supporting comprehension strategy development (Gaskins et al., 1994; Pressley and El-Dinary, 1997). The contrast between the two explicit-strategy sections should therefore not be interpreted as evidence that teacher-led instruction was inadequate. A more defensible conclusion is that teacher-led instruction remained effective, while the AI-assisted section showed the strongest descriptive pattern within this particular bundled classroom configuration.

6.3 Theoretical contribution

The main theoretical contribution of the study lies in how the comparison is framed. Much of the emerging AI literature risks presenting a simplified opposition between human instruction and technological support. The framework proposed in Section 3 suggests a different interpretation. In this study, the strongest section-level pattern was observed not in AI in isolation, but in a teacher-managed classroom configuration in which AI was used to multiply prompts for planning, monitoring, and reflection during reading. This perspective fits a distributed view of scaffolding in which pedagogical work is shared across the instructor, the learner, the task, and the mediating tool (Carrell and Eisterhold, 1983; Macaro, 2006). It also aligns with recent work that situates AI within self-regulated language learning and bounded pedagogical scaffolding rather than autonomous tutoring (Chang and Sun, 2024; Lin et al., 2025; Qian et al., 2026).

This contribution remains modest and context-bound. The study does not settle larger theoretical questions about the long-term internalization of reading strategies or the boundary between productive guidance and overreliance on automated support. It does, however, provide a semester-long comparison showing that AI may be studied more productively as part of an instructional configuration than as a stand-alone treatment label. In the present manuscript, that claim is intentionally limited to classroom-level comparative evidence from three intact sections.

This framing also clarifies how the study differs from several recent quasi-experimental investigations of AI-supported reading. Alazemi (2024) centered AI-integrated formative assessment and a broader set of affective and self-regulatory outcomes. Feng and Wang (2023) examined human-AI robot interaction with primary school bilingual learners rather than university EFL readers. Lin et al. (2025) concentrated on AI-supported pre-reading scaffolding and learner motivation. The present study instead held the instructor, texts, semester length, and strategy targets constant while varying classroom configurations across AI-assisted, teacher-led, and business-as-usual sections. The contribution therefore lies in the bounded pedagogical question of how AI may function inside explicit strategy instruction, not in a broad claim that AI itself independently caused the observed differences.

6.4 Pedagogical implications

Several pedagogical implications emerge from the findings, but they should be read conditionally. First, the observed section-level patterns reinforce the instructional value of making reading strategies a visible and sustained component of university EFL teaching rather than assuming that students will acquire them incidentally through repeated exposure to texts. Because both explicit-strategy sections showed stronger patterns than the business-as-usual section, the study suggests that comprehension growth is more likely when learners are systematically taught how to preview, predict, monitor, question, infer, summarize, and evaluate texts across an extended period of instruction.

Second, the findings indicate that EFL reading courses may benefit from moving beyond routines centered mainly on vocabulary preparation, silent reading, and comprehension checking. A more robust instructional model would incorporate explicit modeling, guided rehearsal, reflective discussion, and repeated opportunities for students to justify their strategic choices in relation to specific reading difficulties. Within this perspective, strategy instruction is not an optional supplement to reading instruction, but a central pedagogical means of supporting deeper comprehension and stronger metacognitive awareness.

Third, the findings suggest that AI may be most educationally useful when it functions within a teacher-managed instructional design rather than as an autonomous tutor. In the present study, the strongest pattern was observed in a section that combined teacher-designed prompts, clear task boundaries, classroom monitoring, immediate device access during guided practice, and one short weekly AI-supported task. Accordingly, the pedagogical takeaway is not that AI should replace teachers, but that carefully bounded AI use may extend opportunities for strategic prompting and guided reflection when teacher judgment remains at the center of instruction.

6.5 Limitations and directions for future research

The findings should be interpreted in light of several limitations. The most important is design-based: the study used one intact section per condition rather than multiple sections per condition or random assignment. As a result, section and condition were confounded, and unmeasured differences in peer composition, meeting time, classroom ecology, or instructor expectations may have contributed to the observed patterns. The sample was also relatively small and drawn from a single institutional context, which limits generalizability across proficiency levels, age groups, and educational settings.

A second limitation concerns treatment bundling and the absence of process-oriented data. The AI-assisted section differed not only in its use of ChatGPT, but also in immediate device access during guided practice and one short weekly AI-supported task outside class. The manuscript also did not analyze interaction logs, screen captures, think-aloud protocols, or classroom observation codes. As a result, the study cannot determine which specific features of the AI-supported package mattered most, how often students relied on different prompts, or whether some learners benefited more than others because of digital literacy, prior AI familiarity, or motivation. These questions remain especially important given the rapid growth of AI-assisted language learning research and the continued need for sharper theoretical models in the field (Mohsen et al., 2025).

A third limitation is that the study focused on outcomes measured at the end of one semester. It did not examine whether the observed patterns persisted after instruction ended or transferred to new reading genres and tasks. Future work should use designs with multiple sections per condition, collect process data, and compare different prompt architectures or levels of teacher mediation while holding access and practice time constant. Such work could build on the present findings while also drawing on recent reading-specific AI studies to identify when AI support is most productive and for whom (Serrano-Mendizábal et al., 2023; Shafiee Rad, 2025; Zheng, 2024).

7 Conclusion

This quasi-experimental comparative classroom study examined whether teacher-mediated AI-enhanced strategy instruction and teacher-only explicit strategy instruction were associated with different patterns of reading development across a 16-week semester in a university EFL course. The observed section-level pattern showed stronger reading-comprehension outcomes in the two explicit-strategy sections than in the business-as-usual section, with the AI-assisted section showing the highest adjusted reading mean. For metacognitive awareness, both explicit-strategy sections improved from pretest to posttest, and the AI-assisted section showed the largest descriptive gain.

These results suggest that AI can contribute productively to university EFL reading instruction when it is used as a constrained scaffold within a teacher-managed pedagogical design. In the present study, the potential value of AI did not lie in replacing instruction, but in supporting additional opportunities for prompting, guided practice, and reflective engagement with text as part of a broader instructional package.

These conclusions should still be interpreted cautiously. Because each condition was represented by a single intact section and the AI-assisted section also included technology access and weekly AI-supported practice, the study does not show that AI independently teaches reading better than teachers do. Rather, the findings indicate that a teacher-managed AI-supported classroom configuration may function as a useful extension of explicit strategy instruction in an advanced university EFL course. In that sense, the study supports a complementary view of AI in pedagogy, one in which AI appears most valuable when it expands opportunities for guided practice and reflection while leaving instructional judgment in the hands of the teacher.

References

• 1

  Al-GhazoA. (2023). The impact of note-taking strategy on EFL learners’ listening comprehension. Theory Pract. Lang. Stud.13 (5), 1146–1153. 10.17507/tpls.1305.06

  • CrossRef
  • Google Scholar

• 2

  AlazemiA. F. T. (2024). Formative assessment in artificial integrated instruction: delving into the effects on Reading comprehension progress, online academic enjoyment, personal best goals, and academic mindfulness. Lang. Test. Asia14 (1), 44. 10.1186/s40468-024-00319-8

  • CrossRef
  • Google Scholar

• 3

  AldamenH.AlmashourM.Al-DeaibesM.AlSharefeenR. (2025). Testing Krashen’s input hypothesis with AI: a mixed-methods study on reading input and oral proficiency in EFL. Front. Educ.10, 1614680. 10.3389/feduc.2025.1614680

  • CrossRef
  • Google Scholar

• 4

  AllehyaniB.AlmashyA.JamshedM.BanuS. (2025). Measuring the impact of meta-AI on English reading comprehension score enhancement: a study within social media application. Theory Pract. Lang. Stud.15 (2), 263–272. 10.17507/tpls.1502.31

  • CrossRef
  • Google Scholar

• 5

  AlnemratA.AldamenH.AlmashourM.Al-DeaibesM.AlSharefeenR. (2025). AI vs. teacher feedback on EFL argumentative writing: a quantitative study. Front. Educ.10, 1614673. 10.3389/feduc.2025.1614673

  • CrossRef
  • Google Scholar

• 6

  AnggiaH.HabókA. (2024). University students’ metacognitive awareness of reading strategies (MARS) in online Reading and MARS’ role in their English reading comprehension. PLoS One19 (11), e0313254. 10.1371/journal.pone.0313254

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  ArdashevaY.WangZ.AdesopeO. O.ValentineJ. C. (2017). Exploring effectiveness and moderators of language learning strategy instruction on second language and self-regulated learning outcomes. Rev. Educ. Res.87 (3), 544–582. 10.3102/0034654316689135

  • CrossRef
  • Google Scholar

• 8

  BlockE. (1986). The comprehension strategies of second language readers. TESOL Q.20 (3), 463–494. 10.2307/3586295

  • CrossRef
  • Google Scholar

• 9

  CarrellP. L. (1985). Facilitating ESL reading by teaching text structure. TESOL Q.19 (4), 727–752. 10.2307/3586673

  • CrossRef
  • Google Scholar

• 10

  CarrellP. L.EisterholdJ. C. (1983). Schema theory and ESL reading pedagogy. TESOL Q.17 (4), 553–573. 10.2307/3586613

  • CrossRef
  • Google Scholar

• 11

  CarrellP. L.PharisB. G.LibertoJ. C. (1989). Metacognitive strategy training for ESL reading. TESOL Q.23 (4), 647–678. 10.2307/3587536

  • CrossRef
  • Google Scholar

• 12

  ChangW.-L.SunJ. C.-Y. (2024). Evaluating AI’s impact on self-regulated language learning: a systematic review. System126, 103484. 10.1016/j.system.2024.103484

  • CrossRef
  • Google Scholar

• 13

  CohenJ. (1988). Statistical Power Analysis for the Behavioral Sciences, 2nd Edn. New York: Routledge. 10.4324/9780203771587

  • CrossRef
  • Google Scholar

• 14

  DörnyeiZ.TaguchiT. (2009). Questionnaires in Second Language Research: Construction, Administration, and Processing, 2nd ed.New York: Routledge. 10.4324/9780203864739

  • CrossRef
  • Google Scholar

• 15

  DuffyG. G. (1993). Rethinking strategy instruction: four teachers’ development and their low achievers’ understandings. Elem. Sch. J.93 (3), 231–247. 10.1086/461724

  • CrossRef
  • Google Scholar

• 16

  FathiJ.AfzaliM. (2020). The effect of second language Reading strategy instruction on young Iranian EFL learners’ reading comprehension. Int. J. Instr.13 (1), 475–488. 10.29333/iji.2020.13131a

  • CrossRef
  • Google Scholar

• 17

  FengY.WangX. (2023). A comparative study on the development of Chinese and English abilities of Chinese primary school students through two bilingual Reading modes: human-AI robot interaction and paper books. Front. Psychol.14, 1200675. 10.3389/fpsyg.2023.1200675

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  GaskinsI. W.GuthrieJ. T.SatlowE.OstertagJ.SixL.ByrneJ.et al (1994). Integrating instruction of science, Reading, and writing: goals, teacher development, and assessment. J. Res. Sci. Teach.31 (9), 1039–1056. 10.1002/tea.3660310914

  • CrossRef
  • Google Scholar

• 19

  GorsuchG.TaguchiE. (2010). Developing reading fluency and comprehension using repeated Reading: evidence from longitudinal student reports. Lang. Teach. Res.14 (1), 27–59. 10.1177/1362168809346494

  • CrossRef
  • Google Scholar

• 20

  GrabeW.YamashitaJ. (2022). Reading in a Second Language: Moving from Theory to Practice, 2nd Edn. Cambridge: Cambridge University Press.

  • Google Scholar

• 21

  GreenJ. M.OxfordR. (1995). A closer look at learning strategies, L2 proficiency, and gender. TESOL Q.29 (2), 261–297. 10.2307/3587625

  • CrossRef
  • Google Scholar

• 22

  HinkelE. (2005). Handbook of Research in Second Language Teaching and Learning, 1st Edn. Mahwah: Routledge. 10.4324/9781410612700

  • CrossRef
  • Google Scholar

• 23

  JanzenJ.StollerF. L. (1998). Integrating strategic reading into L2 instruction. Read. For. Lang.12 (1), 251–269. 10.64152/10125/66962

  • CrossRef
  • Google Scholar

• 24

  JiménezR. T.GarcíaG. E.PearsonP. D. (1995). Three children, two languages, and strategic reading: case studies in bilingual/monolingual reading. Am. Educ. Res. J.32 (1), 67–97.

  • Google Scholar

• 25

  KodaK. (2005). Insights into Second Language Reading: A Cross-Linguistic Approach. Cambridge: Cambridge University Press. 10.1017/CBO9781139524841

  • CrossRef
  • Google Scholar

• 26

  LinC.-C.LinT.-H.TangC.-K. (2025). Enhancing English reading comprehension, learning motivation and attitude through AI-supported pre-reading scaffolding. J. Comput. Assist. Learn.41 (6), e70150. 10.1111/jcal.70150

  • CrossRef
  • Google Scholar

• 27

  MacaroE. (2006). Strategies for language learning and for language use: revising the theoretical framework. Mod. Lang. J.90 (3), 320–337. 10.1111/j.1540-4781.2006.00425.x

  • CrossRef
  • Google Scholar

• 28

  MehmoodW.GondalS.FaizM. S.KhurshidA. (2025). AI-assisted metacognitive strategies for improving self-regulated learning among high school students. Crit. Rev. Soc. Sci. Stud.3 (2), 2333–2349. 10.59075/vk7et188

  • CrossRef
  • Google Scholar

• 29

  MohsenM. A.AlthebiS.QadhiS. (2025). Mapping the evolution of computer-assisted language learning research: a 44-year bibliometric overview. Eur. J. Educ.60 (1), e70051. 10.1111/ejed.70051

  • CrossRef
  • Google Scholar

• 30

  MokhtariK.ReichardC. A. (2002). Assessing students’ metacognitive awareness of reading strategies. J. Educ. Psychol.94 (2), 249–259. 10.1037/0022-0663.94.2.249

  • CrossRef
  • Google Scholar

• 31

  NamaziandostE. (2025). Integrating flipped learning in AI-enhanced language learning: mapping the effects on metacognitive awareness, writing development, and foreign language learning boredom. Comput. Educ. Artif. Intell.9, 100446. 10.1016/j.caeai.2025.100446

  • CrossRef
  • Google Scholar

• 32

  O’MalleyJ. M.ChamotA. U. (1990). Learning Strategies in Second Language Acquisition. Cambridge University Press. 10.1017/CBO9781139524490

  • CrossRef
  • Google Scholar

• 33

  PalincsarA. S.BrownA. L. (1984). Reciprocal teaching of comprehension-fostering and comprehension-monitoring activities. Cogn. Instr.1 (2), 117–175. 10.1207/s1532690xci0102_1

  • CrossRef
  • Google Scholar

• 34

  PearsonP. D.GallagherM. C. (1983). The instruction of Reading comprehension. Contemp. Educ. Psychol.8 (3), 317–344. 10.1016/0361-476X(83)90019-X

  • CrossRef
  • Google Scholar

• 35

  PressleyM.El-DinaryP. B. (1997). What we know about translating comprehension-strategies instruction research into practice. J. Learn. Disabil.30 (5), 486–488.

  • Pubmed Abstract
  • Google Scholar

• 36

  QianK.LiuS.LiT.RakovićM.LiX.GuanR.et al (2026). Towards reliable generative AI-driven scaffolding: reducing hallucinations and enhancing quality in self-regulated learning support. Comput. Educ.240, 105448. 10.1016/j.compedu.2025.105448

  • CrossRef
  • Google Scholar

• 37

  RogiersA.MerchieE.Van KeerH. (2020). Learner profile stability and change over time: the impact of the explicit strategy instruction program “learning light”. J. Educ. Res.113 (1), 26–45. 10.1080/00220671.2019.1711005

  • CrossRef
  • Google Scholar

• 38

  RosenshineB.MeisterC. (1994). Reciprocal teaching: a review of the research. Rev. Educ. Res.64 (4), 479–530. 10.3102/00346543064004479

  • CrossRef
  • Google Scholar

• 39

  Serrano-MendizábalM.VillalónR.MeleroÁIzquierdo-MagaldiB. (2023). Effects of two computer-based interventions on reading comprehension: does strategy instruction matter?Comput. Educ.196, 104727. 10.1016/j.compedu.2023.104727

  • CrossRef
  • Google Scholar

• 40

  Shafiee RadH. (2025). Reinforcing L2 reading comprehension through artificial intelligence intervention: refining engagement to foster self-regulated learning. Smart Learn. Environ.12 (1), 23. 10.1186/s40561-025-00377-2

  • CrossRef
  • Google Scholar

• 41

  ShihY.-C.ReynoldsB. L. (2015). Teaching adolescents EFL by integrating think-pair-share and reading strategy instruction: a quasi-experimental study. RELC J.46 (3), 221–235. 10.1177/0033688215589886

  • CrossRef
  • Google Scholar

• 42

  YappD. J.de GraaffR.van den BerghH. (2021). Improving second language reading comprehension through reading strategies: a meta-analysis of L2 reading strategy interventions. J. Sec. Lang. Stud.4 (1), 154–192. 10.1075/jsls.19013.yap

  • CrossRef
  • Google Scholar

• 43

  ZhengS. (2024). The effects of chatbot use on foreign language reading anxiety and reading performance among Chinese secondary school students. Comput. Educ. Artif. Intell.7, 100271. 10.1016/j.caeai.2024.100271

  • CrossRef
  • Google Scholar