ChatGPT and student engagement in L2 writing: a systematic review towards a smart learning environment framework

Synthesis: Arthi and Gandhimathi (2026) synthesize 13 empirical studies of ChatGPT in university second-language L2 writing and conclude that its effect on Student Engagement is conditional: the tool can sustain planning, revision, and confidence, and it can produce over-reliance, Cognitive Offloading, and frustration with impersonal Feedback. Three dimensions organize the evidence — behavioral participation, cognitive evaluation of suggestions, and affective response — and each shows dual effects. Six studies (46.15%) reported a predominantly positive influence, while seven (53.85%) reported both benefits and drawbacks. From these patterns the authors propose the ChatGPT-Enabled Smart Learning Environment (CESLE) framework, which pairs a three-dimensional engagement model with disengagement indicators and adds three normative constructs under the label Ethical Human–AI Synergy: algorithmic stewardship, ethical collaborative intelligence, and an AI + teacher arrangement. Across the corpus, structured tasks, guided feedback, and opportunities to evaluate AI suggestions were more often linked to positive outcomes, while dependency, incorrect output, and emotional disconnection marked the dual-effect cases.

Key Findings

1. The tool both supports and impedes engagement on every dimension. The same studies report motivation, reflection, and revision alongside over-reliance, inaccurate output, and emotional distance; the balance depends on learner characteristics, task design, and teacher guidance.
2. The split is close to even. Six studies (46.15%) were classified as predominantly positive and seven (53.85%) as dual-effect — an unweighted count the authors decline to treat as a pooled effect estimate.
3. Behavioral engagement shows in planning and revision, and frays in uptake. Learners planned, drafted, monitored, and revised in response to feedback, but also accepted suggestions uncritically or skimmed inaccurate feedback.
4. Cognitive engagement turns on evaluation. Recursive thinking, critical evaluation, and self-regulation appeared in some studies; overload, reduced effort, difficulty spotting wrong feedback, and heavier demands on lower-proficiency learners in others.
5. Affect is mixed and partly relational. Gains include Motivation, confidence, satisfaction, and reduced writing anxiety; costs include frustration at inaccurate feedback, diminished motivation under monotonous feedback, and missing emotional support.
6. Students preferred AI for negative feedback. One study found learners favored critical feedback from the tool over the same feedback from a person, calling the environment less threatening — a "fun adventure."
7. Methods are spread thin. Five studies were mixed-methods (38%), four qualitative (31%), and four quantitative (31%); two comparisons found no significant cognitive-engagement difference between feedback conditions, one found greater improvement over time.

Behavioral and cognitive engagement

Behavioral findings cluster around process visibility. Learners who used ChatGPT while writing planned, drafted, revised, monitored their text, and responded to generated Feedback, and several studies reported higher behavioral engagement after ChatGPT-supported instruction. The disengagement side is concrete: over-reliance, uncritical acceptance, skimming. Screen-recorded revisions counted as observed evidence; learners' accounts counted as perceptions. Cognitive findings are where the review is most cautious. Positive cases show recursive thinking, strategic decision-making, critical evaluation, planning, organization, and reflection. Adverse cases show overload, reduced mental effort, difficulty judging inaccurate output, and heavier linguistic and technological demands on lower-proficiency learners. The authors' sharpest move appears in their discussion: reduced load is not increased engagement, because when the tool performs reasoning or evaluation on the learner's behalf, the saved effort may be offloading rather than Scaffolding.

Affective engagement and the low-threat environment

Affect runs both ways, and much of it is relational. Gains include motivation, confidence, appreciation, satisfaction, and reduced writing anxiety, with learners describing ChatGPT as a tutor, a "writing buddy" that carries them past writer's block, and a partner in building meaning. Costs include frustration and confusion at inaccurate feedback, overload and mental stress, diminished motivation under monotonous feedback, and the missing emotional nuance of human response, alongside educators' worries about plagiarism and Academic Integrity. The comparative case is pointed: students preferred negative feedback from the AI to the same feedback from a person, reporting affirmative emotions and direct involvement in a less threatening space. The review reads this as a low-threat feedback environment rather than proof that AI feedback is better, since the same studies document apathy in students who wanted the personalization a teacher or peer supplies. What separates reflective use from passive reliance, the authors argue, is Feedback Literacy.

The CESLE framework

The constructive output is a conceptual framework, not a tested model. CESLE rests on the tripartite engagement account and expands it toward a smart learning environment where ChatGPT acts as a conversational writing partner, and toward paired engagement and disengagement indicators. Behavioral indicators include accepting, questioning, or rejecting suggestions, annotating, seeking validation, and selectively integrating feedback against earlier drafts. Cognitive indicators include self-talk as metacognitive monitoring, micro-level analysis, and lateral verification. Affective indicators include enjoyment, appreciation, and reduced anxiety against frustration and uncertainty. On this empirical layer the authors place three normative proposals: algorithmic stewardship, meaning critical oversight of AI output; ethical collaborative intelligence, meaning coordination of learner, AI, peer, and teacher contributions while preserving human judgment and authorship; and an AI + teacher arrangement where AI supplies immediate, scalable support while the teacher supplies contextual judgment, emotional support, and ethical guidance. The arrows, they stress, are conceptual relationships, not tested causal pathways.

What this means for practice

• Instructors. Build tasks around evaluation rather than production: require students to accept, question, or reject each suggestion and justify the choice against their own draft, since uncritical acceptance is the most documented behavioral risk.
• Instructors. Do not read a lighter workload as deeper learning. When the tool generates ideas or judges text for the student, redirect the released attention to planning, revision, and reflection.
• Program designers. Keep a human feedback channel alongside the AI one; students valued the low-threat space but also reported emotional disconnection, the gap the AI + teacher arrangement addresses.

Limitations

• The base is 13 studies: 471 records from Scopus and Web of Science yielded 128 abstracts, 27 full texts to assess, 10 inclusions, and three snowball additions. Only English-language peer-reviewed articles from January 2023 onward were eligible, so dissertations, conference papers, and non-English work are absent.
• Those study-level percentages are unweighted counts of study-level classifications, not pooled effects, implying nothing about equivalence in design, sample size, or evidential strength.
• Several included studies rely on self-reported engagement, and positive views may reflect the technology's newness rather than sustained involvement; appraisal also found unclear allocation concealment, infeasible blinding, potential confounding, and one unreported ethical approval.

Citation

Arthi, M. P., & Gandhimathi, S. N. S. (2026). ChatGPT and student engagement in L2 writing: a systematic review towards a ChatGPT-enabled smart learning environment framework. Frontiers in Education.