The Use of Artificial Intelligence in English Language Teaching and Learning: Impact of AI Tutoring on Student Achievement

A. Definition and significance of artificial intelligence

Artificial intelligence (AI) refers to computer systems that perform tasks ordinarily associated with human intelligence: perceiving and interpreting input, reasoning toward a conclusion, learning from experience, and producing and understanding language. Its constituent fields include machine learning, natural language processing, computer vision, and robotics. In education, the term is used more narrowly and more practically to describe systems that adapt to a learner, automate assessment or feedback, generate instructional material, or act as an interlocutor in dialogue (Holmes et al., 2019).

It is useful to distinguish three paradigms of AI in education. In the first, AI directs the learner, who receives instruction sequenced by the system; in the second, AI supports the learner, who works with the system on a task; in the third, AI empowers the learner, who sets goals and directs the system as a tool (Ouyang & Jiao, 2021). English language teaching has moved quickly from the second paradigm into the third, largely because generative models allow learners to use AI to produce and refine language rather than merely to consume exercises.

The significance of AI for education lies in four properties that no previous instructional technology combined: it operates at scale, it adapts to the individual, it responds within seconds, and its marginal cost per interaction is negligible. Together these properties address the oldest structural constraint on schooling, that attention is rationed by the number of teachers (Bloom, 1984). They also introduce new risks, which are concentrated at the same points. Automated judgment can be wrong in ways that learners cannot detect; personalisation depends on data about learners that must be governed responsibly; and the same scale that makes a tool equitable in principle can widen advantage where devices, connectivity, and training are unequally distributed (Holmes et al., 2019; Selwyn, 2019).

B. Historical context within educational frameworks

AI in education is not a new arrival. Programmed instruction and computer-assisted instruction in the 1960s and 1970s encoded a sequenced curriculum and branching remediation, with the PLATO system delivering lessons to large numbers of students long before the personal computer existed. The 1980s brought intelligent tutoring systems that modelled both the domain and the learner, tracking which skills had been mastered and selecting practice accordingly, and expert systems embedded comparable reasoning in diagnostic tools (VanLehn, 2011). The 1990s and 2000s added learner modelling grounded in cognitive theory, natural language processing for automatic analysis of written responses, and adaptive hypermedia. The 2010s shifted the focus outward, toward learning analytics and massive open online courses, where the contribution of AI was to detect patterns in large behavioural datasets rather than to tutor individuals.

Each of these waves produced the same lesson. Where AI was used to automate existing practice, it changed little; where it was used to do something the classroom could not, it changed learning. Reviewers of the field have repeatedly observed that the educational questions, not the technical capabilities, determine the outcome, and that the teacher has been the least studied element of implementations that depend on teachers entirely (Zawacki-Richter et al., 2019; Selwyn, 2019). The current wave, driven by large language models, differs in kind rather than degree. Models trained on very large text corpora can generate coherent prose, sustain multi-turn dialogue, adopt roles, and critique a learner's text, all at negligible cost (Kasneci et al., 2023; Kohnke et al., 2023). What was previously a closed system with a fixed domain model is now an open interlocutor, and this is what makes the present moment both promising and in need of evidence.

C. Importance for English language acquisition

English language teaching is a distinctive case for AI, because language is simultaneously the content of instruction and its medium, and because proficiency is demonstrated in production rather than recognition. A learner can know a rule and still fail to use it under the demands of real communication. Acquisition theory therefore places a premium on two conditions: abundant opportunity to produce the target language and to notice the mismatch between intention and output (Swain, 1995), and interaction in which meaning is negotiated and repaired (Long, 1996), all of this in an affective climate where the learner is willing to risk an imperfect utterance (Krashen, 1982).

Classrooms supply these conditions unevenly. Contact hours are limited, speaking turns per learner per session are few, and the fear of public error is a documented barrier to participation (Horwitz et al., 1986). An AI tutor inverts each constraint. It offers unlimited practice at a time and pace the learner chooses, it responds immediately and without social cost, and it never tires of repetition. For learners who rarely speak in class, this may be the first environment in which they practise aloud at all.

The same features define the risks. If a system produces fluent text on request, the productive struggle through which writing develops can be bypassed, and the learner's apparent level can rise without underlying competence (Barrot, 2023). If automated feedback on pragmatics, register, and cultural appropriateness is unreliable, learners may internalise plausible but wrong models, and they are the least equipped to detect it. And if AI support is unevenly distributed, it may widen rather than narrow the gap it promises to close. These are empirical questions about effect and about condition, and they are the questions this study addresses.

A. Theoretical frameworks supporting technological integration

1. Constructivism and collaborative learning

Constructivist accounts hold that knowledge is actively built by the learner rather than transmitted, and that this building happens in social interaction. Vygotsky (1978) formalised this with the zone of proximal development, the distance between what a learner can accomplish independently and what becomes possible with assistance. Assistance is effective when it is contingent, that is, when it is calibrated to the learner's current state and withdrawn as competence develops. The construct also implies an interlocutor more capable than the learner, so the question an AI tutor raises in theoretical terms is whether a machine can serve as that interlocutor.

In second language acquisition the same social premise appears as the interaction hypothesis (Long, 1996) and the output hypothesis (Swain, 1995). Learning is driven by attempts to communicate, by requests for clarification, by repair, and by noticing the gap between what was meant and what was said. This is significant for AI tutoring because it specifies what practice must contain to be of value. Repeated drilling of a structure without negotiation of meaning does not meet the condition. A conversational agent that asks a follow-up question, reformulates, and offers alternative phrasing does.

Two qualifications follow from the theory itself. First, AI is not a peer. Collaborative learning draws part of its power from mutuality, shared goals, and the obligation to make oneself understood to someone whose understanding matters. An AI interlocutor cannot supply relatedness in the sense used by self-determination theory, even when it supplies autonomy and competence support (Ryan & Deci, 2000). Second, scaffolding must be contingent, and contingency depends on an accurate model of the learner, which is precisely where generative systems are least transparent. Constructivism therefore predicts both where AI tutoring will work and where it will not, and both predictions are testable.

2. The TPACK framework

Mishra and Koehler (2006) argue that technology integration is not additive. Effective use requires technological pedagogical content knowledge, that is, an understanding of how a particular technology interacts with the content being taught and with the methods used to teach it. The framework explains why the same tool produces different outcomes in different rooms: a tutor used to replace a worksheet, and the same tutor used to generate the raw material for a debate in the following class, are not the same intervention.

TPACK carries direct implications for research design. If outcomes depend on the teachers' integration knowledge, then studies that vary the tool while holding pedagogy constant may understate its potential, and studies that ignore pedagogy entirely may overstate it. This study addresses the issue by holding the curriculum, the teacher, the materials, and the assessment constant across conditions, and by treating the tutor as an object of instruction in its own right, with training for the teacher and explicit guidance for learners about when to trust it.

B. Current trends in artificial intelligence as an educational tool

1. Chatbots and adaptive learning platforms

Conversational agents in language learning have evolved through three overlapping generations. The first, rule-based and pattern-matching, followed the ELIZA lineage and demonstrated that learners attribute competence to systems that merely respond plausibly, an early warning about trust (Fryer & Carpenter, 2006). The second generation was goal-oriented: task-based dialogue systems and pedagogical agents that guided a learner through a scripted exchange, with a domain model that could track progress. The third generation, made possible by large language models, holds open dialogue that is not pre-scripted, adapts register and role, and can explain its corrections (Kohnke et al., 2023).

Adaptive platforms pursue the same objective by a different route. Spaced repetition algorithms determine when an item should be reviewed, item response models estimate ability, and content is selected to keep the learner near the boundary of competence. Mobile platforms have made this accessible to very large numbers of learners and have produced measurable, if modest, gains in vocabulary, grammar, and listening (Loewen et al., 2019). The limitations of these platforms are equally consistent across studies: open production and interaction are constrained, feedback concerns form more than meaning, and sustained use depends on features that resemble a game more than a lesson.

The current research frontier combines the two. Adaptive sequencing determines what to practise; a generative model conducts the practice and gives reasons. This is the configuration examined here, and it is also the configuration for which evidence is thinnest, because the tools became widely available only recently and adoption in classrooms has outrun systematic evaluation (Crompton & Burke, 2023).

2. Findings from previous empirical studies

The most substantial evidence base concerns intelligent tutoring systems. Kulik and Fletcher (2016), reviewing fifty controlled evaluations, reported a median effect size of 0.66 in favour of ITS over conventional teaching, and Ma et al. (2014) reached similar conclusions in a meta-analysis of cognitive outcomes, with the largest effects where systems modelled the learner and provided step-level feedback. VanLehn (2011) compared human tutoring, ITS, and other tutoring systems and concluded that ITS approaches the effectiveness of human tutoring, with human tutors retaining only a modest advantage. These findings matter for the present study because they establish that machine tutoring can work, and because their limitation is equally clear: the evidence is drawn overwhelmingly from mathematics, science, and procedural domains, where the correct answer is well defined and can be modelled formally.

Evidence in language learning is more mixed. App-based studies show statistically significant gains from additional mobile practice, but with small effects, substantial attrition, and heavy reliance on self-selected samples (Loewen et al., 2019). Studies of automated writing evaluation report improvements in accuracy and revision behaviour, alongside concerns that learners over-trust the automated judgment and revise toward superficial changes (Godwin-Jones, 2022). Research on chatbots in language learning before generative AI found positive effects on willingness to communicate and on fluency practice, with acknowledged limits in error correction and discourse coherence (Fryer & Carpenter, 2006).

Since 2022, the literature on generative AI in language education has expanded quickly, but mostly in the form of conceptual analysis, classroom reports, and small exploratory studies (Kohnke et al., 2023; Barrot, 2023). Reviews of the wider field converge on a methodological diagnosis: samples are small, durations are short, comparison groups are frequently absent, engagement is measured informally or inferred from usage, and learners' own accounts are reported only anecdotally (Zawacki-Richter et al., 2019; Crompton & Burke, 2023; Zhang & Zou, 2022). A systematic review of technology use in second and foreign language learning similarly notes that reported effectiveness is often judged by perceived usefulness rather than by measured proficiency.

Three consequences follow, and they structure this study. First, the increment attributable to AI tutoring over ordinary instruction in English remains unquantified with a comparison group. Second, engagement is claimed more often than it is measured, and almost never across its behavioural, emotional, and cognitive dimensions at once. Third, the conditions under which the tutor is integrated, including what the teacher does with the practice it generates, are rarely described in enough detail for replication. The present study addresses each in turn.

A. Participant selection criteria

Participants were 120 learners enrolled in a two-semester English for Academic Purposes course at a public university. Intact classes were used, and the two classes were allocated to conditions by random draw: 60 learners formed the AI-tutoring condition and 60 formed the comparison condition. Inclusion criteria were enrolment in the course, a placement result at the intermediate level corresponding to CEFR B1, age of 18 years or above, and written informed consent. Exclusion criteria were attendance below 80 per cent of sessions, incomplete pre-test or post-test data, and previous enrolment in a course that used the same AI tutor, on the grounds that familiarity would confound the measurement of first exposure. Applying these criteria left 57 learners in the AI condition and 56 in the comparison condition for the analysis of proficiency, with the full samples retained for the descriptive reporting of survey data.

For the qualitative strand, fifteen learners were selected purposively from the AI condition using a maximum-variation strategy on measured gain, so that the interview sample included high, typical, and low gainers in approximately equal numbers, together with variation in gender and first language. This decision follows the recommendation of Dörnyei (2007) that qualitative samples be selected for information richness rather than representativeness, and it guards against the common bias of interviewing only the learners who liked the tool.

B. Research design rationale

1. Comparison groups

The study used a mixed-methods design with a quasi-experimental core: a non-equivalent comparison group, pre-test and post-test structure, with an embedded qualitative strand that explains the quantitative results rather than standing apart from them (Creswell & Plano Clark, 2018). A quasi-experimental rather than a randomised design was chosen because randomising individual learners would require dividing an existing class, changing the social conditions of instruction, and creating unequal access to a tool within a single cohort, which is both an ethical and a validity problem in an educational setting (Cohen et al., 2018).

Both conditions received the same syllabus, materials, teacher, task sequence, and assessment. The single difference was the presence of the AI tutor. In the AI condition, learners worked with a conversational tutor for 20 to 25 minutes in each session and were assigned between-session practice with the same tutor; the tutor conducted role-plays and discussion, gave corrective feedback on written and spoken production, explained errors on request, and reviewed previously learned items at spaced intervals. In the comparison condition, the equivalent time was used for paired and small-group practice with the same task objectives, the teacher's feedback, and independent study with the course materials. Fidelity was checked through a structured observation protocol completed after each session in both conditions.

Three features of this design address known threats. Pre-test equivalence was verified before analysis and controlled statistically with analysis of covariance, reducing the risk that pre-existing class differences account for the results. The intervention ran for twelve weeks, long enough for initial novelty to recede and for its effect to become visible in the usage data. The qualitative strand was drawn from learners at all levels of gain, so that the explanation of the results is not supplied solely by the most successful users.

2. Available resources and study phases

The study required the following resources: two intact classes taught by the same teacher; licences for the AI tutor for the 60 learners in the experimental condition; a computer laboratory with reliable connectivity for the in-session tutor work, with mobile access for between-session practice; two research assistants, one for test scoring and one for interview transcription and first-cycle coding; and a one-and-a-half-day training workshop for the teacher covering task design with the tutor, interpretation of its feedback, and the boundary between practice and assessment. Contingency arrangements were made for connectivity failure, in which sessions were completed with the printed equivalent of the same tasks.

C. Instruments

• Proficiency test. A parallel-form institutional test covering grammar and vocabulary, reading, and listening, plus a timed written task (250 words, argumentative prompt) and a recorded speaking task (three-minute monologue and a two-minute follow-up exchange). Writing and speaking were double-rated by two trained raters using a five-criterion analytic rubric (task response, coherence, lexical range, grammatical range and accuracy, pronunciation and fluency), with disagreement resolved by discussion. Inter-rater agreement was .84 by Cohen's kappa for writing and .81 for speaking.
• Engagement questionnaire. A 24-item instrument adapted from established measures of course engagement and work engagement, capturing behavioural engagement (effort, attendance, time on task), emotional engagement (interest, enjoyment, anxiety reverse-scored), and cognitive engagement (strategy use, depth of processing) on a five-point Likert scale. Internal consistency in this sample was acceptable for all three subscales.
• Perception survey. Twenty-five items covering perceived usefulness, ease of use, feedback quality, confidence, and preference relative to traditional instruction, with three open-ended items.
• Semi-structured interviews. Conducted in the learners' preferred language, 30 to 40 minutes, audio-recorded and transcribed verbatim, then translated for reporting with back-translation checks on all quoted extracts.
• Interaction logs. Session timestamps, duration, number of learner turns, use of translation and explanation requests, and task types completed.

A. Descriptive statistics for survey responses

Survey and test data were entered into a statistical package and screened for out-of-range values, duplicates, and missing responses before analysis. Descriptive statistics were computed for every item and for each composite: means and standard deviations for the test scores and for the three engagement subscales, and frequencies with percentages for the individual Likert items. Distributions are reported rather than means alone, because a mean of 4.0 on a five-point item conceals the difference between consensus and a polarised sample, and that difference matters when the item concerns trust in automated feedback.

Internal consistency was assessed for each subscale before composites were formed, and all reported values met conventional thresholds. Missing data were limited in the endline surveys and were handled by available-case analysis, with the number of valid responses reported for each statistic; learners missing either the pre-test or the post-test were excluded from the proficiency analysis, as specified in the selection criteria, and their exclusion is reported in the participant flow.

Inferential analysis proceeded in three steps. Within-group change was tested with paired-samples t-tests. Between-group differences in gain were tested with an independent-samples t-test on gain scores, and then with analysis of covariance using the pre-test as covariate, which is the more conservative test of the treatment effect because it adjusts for initial differences and reduces the influence of regression to the mean. Assumptions were checked before each test: normality of the distributions, homogeneity of variance, and homogeneity of regression slopes for the covariance analysis. Effect sizes are reported throughout, as Cohen's d for mean comparisons and partial eta squared for the covariance analysis, on the principle that statistical significance without magnitude is not interpretable (Cohen et al., 2018). The significance level was set at .05, with a Holm correction applied across the engagement subscale comparisons to control the family-wise error rate. Relationships between tutor usage and gain were examined with Pearson correlations, and are reported descriptively, with the explicit caveat that dosage is not randomly assigned and the association is therefore not causal.

B. Content analysis approach for qualitative data

Interview transcripts and the open-ended survey responses were analysed using reflexive thematic analysis following the six phases described by Braun and Clarke (2006). The first phase involved repeated reading of the transcripts and the recording of initial impressions. The second generated initial codes across the full dataset, both semantic codes close to what participants said and latent codes concerned with what those statements implied about trust, identity, and the learner's relationship to the tool. The third phase grouped codes into candidate themes and reviewed them against the coded extracts. The fourth refined the themes, testing each for internal coherence and for distinctness from the others, and discarding a fifth candidate theme on self-regulation that did not hold across the dataset. The fifth phase defined and named the final themes. The sixth produced the report, with themes illustrated by verbatim extracts.

Trustworthiness was addressed in four ways. Two coders independently coded 20 per cent of the transcripts, reaching agreement of .79 by Cohen's kappa on a shared codebook, after which the remaining transcripts were coded by one coder with the second auditing the coded output. An audit trail recorded all analytic decisions, including changes to the codebook. Emerging themes were member-checked with five interviewees, who confirmed three themes, refined one, and prompted a change of emphasis in a fourth toward the practical limits of the tutor rather than its acceptability in principle. Negative cases were sought deliberately: extracts from learners who stopped using the tutor are reported alongside the majority view.

The two strands were then integrated. A joint display was constructed in which each quantitative finding was set beside the qualitative evidence that explained, qualified, or contradicted it, and the discussion is organised around that display rather than reporting the strands separately.

A. Proficiency outcomes

Both conditions improved from pre-test to post-test, and the improvement was statistically significant in each. The AI-tutoring condition gained more than twice as much on the composite measure as the comparison condition: a mean gain of 14.4 points against 6.3 points.

The difference in gain between conditions was significant before adjustment, t(111) = 6.02, p < .001, d = 1.10, and remained significant after controlling for pre-test score, F(1, 110) = 24.6, p < .001, partial eta squared = .17. The assumptions of the covariance analysis were satisfied. An effect of this magnitude corresponds to roughly seventeen per cent of the variance in post-test performance attributable to condition after initial differences were accounted for, which places the intervention in the range reported for intelligent tutoring systems in other domains (VanLehn, 2011; Kulik & Fletcher, 2016).

Disaggregating the composite by skill shows where the difference was concentrated. Gains in reading and listening were comparable across conditions and modest in both. Gains in grammar and vocabulary favoured the AI condition moderately. The largest differences appeared in the productive skills, where the AI condition gained more than twice as much as the comparison condition in writing and nearly three times as much in speaking.

B. Engagement

Engagement was higher in the AI-tutoring condition on all three dimensions, and the difference was significant for behavioural and emotional engagement after correction for multiple comparisons. The cognitive engagement difference was smaller and did not reach significance.

The behavioural difference is corroborated by the usage data. Learners in the AI condition spent more time on out-of-class English practice than the comparison group in every week of the intervention, but the difference was not stable. Practice time rose through the first seven weeks, peaked at 92 minutes per week, and then declined steadily to 83 minutes by the final week, while the comparison group's practice time remained close to 48 minutes throughout. The trajectory matters more than the mean: a rising curve that turns flat or downward is the pattern normally associated with novelty, and it argues against reading the engagement advantage as durable.

Practice time in the AI condition correlated moderately with composite gain, r(55) = .48, p < .001, indicating that learners who used the tutor more gained more, but the relationship was far from determinate: several learners with high usage made typical gains, and one of the largest gains in the condition came from a learner in the middle of the usage distribution. Because dosage was not randomly assigned, this association cannot be read as a causal dose-response relation; it may equally reflect that more motivated learners both used the tool more and studied more in other ways.

C. Perceptions of AI tutoring

Responses to the perception survey were favourable on the properties that concern practice and feedback, and markedly less favourable on the properties that concern judgment and nuance. Items were collapsed to agreement percentages for reporting.

D. Qualitative findings

Four themes were developed from the interviews and the open-ended survey responses. They are presented with the frequency with which each appeared across the fifteen interviewees, and with the negative cases noted where they qualify the theme.

• Theme 1: Calibrated pace and endless repetition. Thirteen interviewees described the tutor's central value as the ability to repeat, slow down, or advance without embarrassment or waiting for others. Learners who described themselves as slow in class were the most emphatic. The theme is consistent with the behavioural engagement scores, and the negative cases were informative: two learners reported that the tutor's willingness to repeat indefinitely made them less attentive on the first attempt, because they knew another was available.
• Theme 2: Low-stakes practice and reduced anxiety. Eleven learners linked the tutor to a reduction in the fear of error, several describing it as the first context in which they had spoken English aloud without preparation. This theme maps onto the emotional engagement difference and onto the widely reported pattern that willingness to communicate rises when the interlocutor does not judge. It also qualifies it: the reduction was described as confined to the tutor, and only four learners reported carrying the confidence gained into class discussion.
• Theme 3: Limits of automated judgement. Twelve learners described instances in which they distrusted or discovered an error in the tutor's feedback. The reported failures cluster in pragmatics, collocation, and register, exactly the areas where judgment depends on context rather than rule, and learners reported responding with a verification strategy: asking the teacher before using a form in assessed work. Notably, learners who had been given explicit instruction about the tutor's limits during the training week described this strategy as routine, while learners who had not reported it as a discovery.
• Theme 4: The teacher as the final authority. Ten learners articulated a division of labour in which the tutor is responsible for practice and the teacher for explanation and judgment. This theme accounts for the otherwise puzzling survey result that 91 per cent found the tutor's feedback timely while only 21 per cent would trade class time for more of it. Learners were not rejecting the tutor; they were assigning it a subordinate role, and no interviewee proposed that it replace instruction.

A. Interpretation against previous literature

The proficiency result is consistent with, and extends, the evidence from intelligent tutoring systems. Studies of ITS in mathematics and science report moderate to large effects, a median of 0.66 across fifty controlled evaluations (Kulik & Fletcher, 2016), and conclude that machine tutoring approaches human tutoring in effectiveness (VanLehn, 2011; Ma et al., 2014). This study finds a comparable magnitude in a domain those reviews largely excluded, and with a technology that depends on no formal domain model at all. That the effect survived adjustment for pre-test performance suggests it is not an artefact of the initial equivalence of the groups, though the quasi-experimental design cannot exclude unmeasured differences between classes.

The distribution of gains across skills is the more informative finding, and it aligns closely with acquisition theory. Writing and speaking gains were two to three times larger in the AI condition, while reading and listening gains were similar across conditions. This is what the interaction and output hypotheses predict: the advantage of a conversational tutor lies in the volume of production it permits and in the immediate, interpretable feedback that follows production (Swain, 1995; Long, 1996). Reading and listening, by contrast, are receptive skills that a conventional curriculum already supplies in quantity, so the incremental value of the tutor is smaller. The result argues against describing AI tutoring as a general-purpose improvement to language instruction, and for describing it as an intervention with a specific mechanism, concentrated where practice is scarce.

The engagement findings are consistent with the motivational accounts of the literature but complicate them. Higher behavioural and emotional engagement, with little change in cognitive engagement, suggests that the tutor increased effort and reduced fear more than it deepened processing. This distinction is important because the literature often treats engagement as a single quantity, and the practical implications differ: effort without depth may produce fluency gains while leaving strategic learning untouched (Kohnke et al., 2023). The decline in practice time after week seven, together with the plateau and fall visible in the usage data, supports the long-standing caution that short-term engagement gains may be novelty effects, and it is the reason the effect on proficiency must be interpreted with the trajectory in view rather than with the peak (Zawacki-Richter et al., 2019). That the gains nonetheless persisted to the endline test suggests either that the habit was established before the decline, or that the concentration of practice in the first half of the semester was sufficient for the observed improvement, or that gains were carried by factors other than volume of use. The design cannot distinguish these, and a delayed post-test would be needed to test retention.

The perception findings replicate the concerns of the conceptual literature with unusual clarity. Learners valued immediacy and low-stakes practice, and reported distrust precisely where the literature predicts it: pragmatics, collocation, and register (Barrot, 2023; Godwin-Jones, 2022). The two results together, high satisfaction with feedback timing and low willingness to trade class time for more tutor use, are not contradictory. They describe an acceptance of the tool as a practice partner combined with a refusal to accept it as an authority, and they match the argument that the teacher's role in AI-supported instruction is not diminished but redefined (Selwyn, 2019). The finding that learners who were explicitly taught about the tutor's limitations developed verification strategies as a matter of routine, while others discovered them by accident, is a practical result with an immediate implication for implementation.

Finally, the study offers indirect support for the TPACK position. The intervention varied the tool while holding curriculum, teacher, materials, and assessment constant, and the tutor's tasks were designed so that their output became material for the following lesson rather than standing alone. In the language of the framework, the treatment was not the technology but the technology-content-method configuration (Mishra & Koehler, 2006). Its effect size should therefore be read as the effect of that configuration, and it does not license the inference that access to the same tutor, without the design work around it, would produce comparable gains.

B. Limitations

• Design. Allocation was by intact class rather than by individual random assignment, so unmeasured differences between the classes cannot be entirely excluded, despite verified pre-test equivalence and the use of analysis of covariance.
• Scale and context. The study was conducted at one institution, with learners at a single proficiency level in one academic context, over a single semester. Findings transfer to similar settings rather than generalising across levels, languages, and educational systems.
• Duration and retention. Twelve weeks is long enough for novelty to begin to recede but not long enough to establish durable change, and the absence of a delayed post-test means nothing can be said about retention or about whether writing gains survive the withdrawal of the tool.
• Self-report. The engagement questionnaire and the perception survey are self-reported, and learners in the experimental condition knew they were receiving something additional, which may inflate both engagement and satisfaction independently of the tool's effect on learning.
• Reproducibility. The tutor is a generative system whose behaviour depends on the model version in use, on prompt configuration, and on the safety and grounding settings applied. The study describes its configuration, but exact replication with a later model version cannot be assumed.
• Rater and teacher effects. Raters were trained and agreement was verified, but they were not blind to the general purposes of the study. The teacher taught both conditions, which controls for teacher quality but does not eliminate the possibility that expectancy influenced the experimental condition.
• Usage data. Practice-time data combine logs and self-report, and the association between dosage and gain is observational; it is reported descriptively and should not be interpreted causally.
• Ethical limits of scope. The study did not examine whether long-term reliance on a generative tutor reduces learners' independent writing competence, an outcome that requires a longitudinal design and that the literature has not yet settled (Barrot, 2023).

This study asked whether AI tutoring improves English language achievement, how it affects engagement, and how learners judge it against traditional instruction. The three initial objectives were addressed directly. In answer to the second objective, learners in the AI-tutoring condition made significantly greater gains in proficiency than the comparison group, with the advantage concentrated in writing and speaking. In answer to the first, their engagement was higher in its behavioural and emotional dimensions but not in its cognitive dimension, and the practice-time trajectory suggests that the behavioural advantage is not automatically sustained. In answer to the third, learners described the tutor as a valuable practice partner and declined to treat it as an authority, locating its limits in the places where automated judgment is weakest.

The core conclusion is that AI tutoring functions as a supplement rather than a substitute, and that its contribution is specific rather than general. It supplies the practice that individual attention has always been able to supply and group instruction has never been able to, and it does so most effectively for the productive skills. Where practice is already abundant, or where judgment rather than repetition is required, its incremental value is small. The conditions surrounding the tool, including training for the teacher, guidance for learners about its limits, and tasks that connect tutor practice to subsequent instruction, are not incidental to the effect; on the evidence of the qualitative strand they are part of it.

For educators and institutions, the implication is to integrate rather than to adopt. A tutor introduced without changes to task design, without instruction in how to test its feedback, and without a defined place in the lesson is unlikely to reproduce the results reported here, and the resources it consumes are better spent elsewhere. A tutor integrated as a source of unlimited low-stakes production, with the teacher retained as the final authority and the resulting language used in class, addresses a constraint that no other available measure addresses at comparable cost.

Several directions for further work follow from the limitations. A delayed post-test would establish whether the writing and speaking gains persist once the tool is withdrawn, which is the question that bears most directly on whether AI practice builds competence or scaffolds performance. Multi-site randomised designs with larger samples, and comparisons across proficiency levels, first-language backgrounds, and subject contexts, would test the boundaries of the effect reported here. Comparisons between feedback conditions, for example corrective feedback alone against corrective feedback with explicit explanation, would identify the mechanism rather than the technology. Research on writing development without the tool, and on the design of assessment that remains valid when generative systems are available, is needed to guard against the deskilling that learners in this study already anticipated. Finally, professional development deserves study in its own right, because the evidence here and throughout the literature is that the teacher determines whether a tutor improves learning or merely accompanies it.