From Scaffolding to Substitution? Generative AI, Self-Regulation, and Learner Agency in EFL Academic Writing in Benin

Synthesis: This explanatory sequential mixed-methods study examined how Generative AI enters the academic writing process among 186 undergraduate EFL students in Benin, integrating questionnaire data with thematic frequencies from 24 English lecturers. It distinguishes frequency of use from breadth of platform use, and examines support across pre-writing, drafting, revising, and proofreading alongside self-regulated writing and risk awareness. ChatGPT was the most widely used platform (86.0%), but platform breadth was unrelated to perceived writing benefit. Frequency of use, by contrast, was strongly associated with perceived benefit, writing confidence, and self-regulated writing. Support was concentrated in revising and proofreading, and a benefit-risk coupling emerged: students reporting greater benefit also reported stronger concerns about overreliance, verification, and integrity. Lecturer themes converged, emphasizing linguistic scaffolding, revision support, overreliance, and authorship concerns. The authors frame GenAI as regulated Scaffolding whose educational value depends on learners' capacity to plan, monitor, evaluate, and retain ownership of their writing. Because the design is cross-sectional and self-report, it claims neither objective writing gains nor causal effects.

Key Findings

1. Intensity, not platform breadth, matters. Students used an average of 2.51 platforms, but platform number was unrelated to perceived benefit (rho = .017) and self-regulation (rho = -.034).
2. ChatGPT dominated but was not the whole picture. ChatGPT was reported by 160 of 186 students (86.0%), followed by Grammarly (60.8%), Gemini (32.3%), QuillBot (28.5%), Microsoft Copilot (22.0%), and Claude (21.0%).
3. Support concentrated in revision and proofreading. Revising and proofreading both scored M = 4.16, significantly above pre-writing (M = 3.79) and drafting (M = 3.71) in a Friedman test, chi-square(3) = 105.17.
4. Frequency was strongly associated with outcomes. Frequency correlated with perceived benefit (rho = .615), writing confidence (rho = .695), self-regulated writing (rho = .869), and even challenges (rho = .879).
5. Benefit and risk rose together. Students reporting greater benefit also reported stronger concerns (rho = .522), and the mean challenge score rose from 2.40 among rare users to 4.25 among frequent users.
6. Academic level did not explain the pattern. Perceived benefit (F(2, 183) = 0.08), self-regulation, confidence, and challenges did not differ significantly across L1, L2, and L3.
7. Lecturers converged on a scaffolding-substitution tension. Twenty-one of 24 lecturers (87.5%) identified linguistic scaffolding and 22 (91.7%) identified revision support, while 20 (83.3%) raised overreliance and 19 (79.2%) raised integrity concerns.

What GenAI actually supports in EFL writing

The study separates tool accumulation from engagement intensity. Students reported using several platforms, yet the number of platforms was unrelated to perceived benefit and self-regulation; frequency of use, in contrast, was strongly associated with both. Support was unevenly distributed across the writing process: grammar revision was the strongest single stage-related item (M = 4.24), followed by proofreading accuracy (M = 4.16) and vocabulary revision (M = 4.08). This suggests students experience GenAI most consistently as a revision and language-polishing resource, even though substantial use also occurs earlier. Crucially, cognitive support was also high (M = 3.98), and lecturers frequently identified idea generation and organization. Once AI participates in deciding what to say and how to structure it, questions of authorship and learner Learner Agency become pressing. The pedagogical challenge is not to classify GenAI as either a grammar tool or a ghostwriter, but to recognize that it moves along that continuum within a single writing task.

Regulation, agency, and the benefit-risk paradox

The most striking result is the positive coupling of benefit and concern. Students who felt AI supported their writing more strongly were also more likely to report concerns about overreliance, verification, Academic Integrity, and reduced critical engagement. This is not the pattern expected from a simple technology-acceptance model in which positive evaluation crowds out concern. Instead, intensive users occupy a double position: they value AI because it solves real writing problems, while also recognizing risks that emerge when reliance becomes routine. Frequency of use was very strongly associated with the self-regulation index, but the cross-sectional design cannot tell whether frequent use develops regulation, whether self-regulated students use AI more strategically, or whether both reflect another factor. The authors place regulation at the center of the scaffolding–substitution boundary: AI assistance is defensible when learners remain responsible for goals, monitoring, evaluation, and final decisions, and becomes more substitution-like when output is accepted without verification.

Assessment and pedagogy under AI mediation

The findings support a move from generic AI policy toward process-oriented AI literacy. Students need explicit practice in deciding what forms of assistance are appropriate at different stages of writing: generating alternatives in pre-writing while justifying topic and argument choices, distinguishing AI suggestions from their own claims during drafting, and comparing AI Feedback with original wording during revision. For source-based writing, the authors argue that verification should be compulsory, with students checking citations and factual claims against traceable sources, since fabricated references are a documented risk. Because a polished final text is increasingly insufficient evidence of independent competence, Assessment should capture process evidence when stakes are high—short oral defenses, annotated revision histories, process notes, AI-use statements, in-class writing components, or reflective commentaries. These measures are framed not as detection devices but as ways to preserve the educational meaning of writing and to protect critical engagement. The authors also note that under-resourced contexts can make AI support unusually attractive, and that lecturer agency is crucial when institutions lack mature guidance.

What this means for practice

• Instructors. Design stage-specific AI guidance: require justification of topic and argument choices in pre-writing, distinguish AI suggestions from students' own claims in drafting, and have students explain which revisions they accepted or rejected.
• Assessment designers. Capture process evidence—annotated revision histories, process notes, oral defenses, and AI-use statements—because finished-text quality is not a transparent measure of independent competence.
• Instructors. Make verification compulsory for source-based writing, since fabricated and inaccurate citations are a documented risk, and build AI-literacy training that addresses overreliance and authorship.
• Researchers. Move to process-tracing designs with prompts, interaction logs, draft histories, and blinded scoring, since the current data are self-reported and cannot observe substitution directly.
• Institutions. Recognize that under-resourced contexts make AI support attractive while connectivity, language background, and institutional policy may moderate the relationships observed.

Limitations

• The student data are self-reported: perceived writing benefit is not a writing score, and the associations do not demonstrate that GenAI improves writing performance.
• The design is cross-sectional and observational, with no control group, pre-test/post-test writing scores, or direct analysis of student texts.
• The qualitative dataset consisted of final theme frequencies rather than full interview transcripts, preventing discourse analysis or re-examination of deviant cases.
• The study lacks prompts, AI interaction logs, screen recordings, draft histories, and independently scored samples, so substitution itself was not directly observed.
• The strong correlations among frequency, regulation, support, and challenges may partly reflect common-method variance.
• The study is situated in one national and institutional EFL ecology, so the absence of academic-level differences should not be generalized beyond this context.

Citation

Toboula, C. M. Z., Sakpoliba, I. G., & Ezin, E. O. (2026). From Scaffolding to Substitution? Generative AI, Self-Regulation, and Learner Agency in EFL Academic Writing in Benin. Research Square preprint.