Generative AI assisted grammar instruction and Ethiopian EFL learners' self-efficacy and metalinguistic awareness

Synthesis: Wordofa and Abo (2026) test a ten-week grammar intervention in which first-year Natural Science students at Mattu University, Ethiopia, used ChatGPT inside a structured draft–prompt–reflect–revise cycle while a comparison class received conventional teacher-marked instruction. With 88 students (45 experimental, 43 comparison) and pre-test scores controlled, the ChatGPT group outperformed the comparison group on all five writing dimensions, on metalinguistic reflection, and on writing Self-Efficacy. The authors deliberately frame the tool as Scaffolding rather than as a substitute for teaching: the AI supplied form-focused Feedback and rule explanations, but students had to name the grammar pattern, explain it in their own words, and revise before the cycle repeated. The study sits in an Ethiopian Global South context largely absent from the GenAI-and-writing literature and still rarely studied, and attributes the effects to the whole instructional package rather than to ChatGPT alone, since the comparison class also got teacher feedback.

Key Findings

1. Writing performance improved on every dimension. Group membership accounted for roughly 58% of variance in combined writing outcomes: Wilks' Lambda = 0.424, F(5, 77) = 20.885, p < .001, partial η² = 0.576.
2. Grammar showed the largest single effect. Adjusted post-test means were 17.76 against 14.75, F(1, 81) = 80.83, p < .001, partial η² = 0.499.
3. Content improved although the prompt forbade content feedback. Adjusted means were 21.28 versus 18.16, F(1, 81) = 79.26, p < .001, partial η² = 0.495 — a gain the authors attribute to repeated drafting and revision.
4. Organization, vocabulary, and mechanics also moved. Adjusted means: organization 14.16 versus 12.04 (η² = 0.455), vocabulary 14.26 versus 12.08 (η² = 0.430), mechanics 3.51 versus 3.04 (η² = 0.226), all significant at Bonferroni-adjusted alpha of 0.010.
5. Metalinguistic reflection rose. Adjusted means were 3.42 against 2.98, F(1, 84) = 18.08, p < .001, partial η² = 0.177, measured on ten adapted items from the EFL Learners' Writing Metacognitive Experiences Questionnaire.
6. Writing self-efficacy rose more sharply. Adjusted means were 3.41 against 2.90, F(1, 84) = 28.54, p < .001, partial η² = 0.254, measured on twelve adapted items from the Self-Efficacy for Writing Scale.
7. Baseline equivalence and fidelity were checked. Pre-test MANOVA found no group difference (Wilks' Lambda = 0.980, F(5, 82) = 0.333, p = .892); a checklist confirmed protocol adherence in 98.5% of sessions.

A four-stage cycle built for noticing

The intervention ran once weekly for ten weeks in Communicative English Skills sessions, using ChatGPT (GPT-4o) on students' smartphones or laptops in lab sessions. Students first wrote a 150–200-word guided paragraph designed to elicit target structures — passive voice, subordination, relative clauses, modal verbs for hedging. They submitted it with an instructor-provided prompt asking for grammatical errors only, each explained with its underlying rule and a corrected sentence, leaving content and organization untouched; a second prompt generated three practice items. In the reflect stage students named the grammatical category flagged most often that week and explained the rule in their own words. The revise stage produced a final version, with both drafts retained. The same instructor taught both; the comparison class received marginal annotations and error codes instead. The design enacts Schmidt's noticing hypothesis: the AI makes form salient, and the reflection sheet forces the learner to articulate it.

Gains across all five writing dimensions

Compositions were scored on an analytic rubric weighting content at 30%, grammar at 25%, organization at 20%, vocabulary at 20%, and mechanics at 5%; two trained raters blind to group and occasion reached an intraclass correlation of 0.801. Every dimension favored the ChatGPT condition, broader than a grammar-only account predicts. Content is instructive: students asked only for grammatical feedback, so the gain cannot be traced to AI commentary on ideas. The explanation is procedural — repeated cycles gave students more passes over the whole text — and the authors read it as evidence that gains depend less on what the AI comments on than on how AI-supported revision is embedded in the writing process. Effect sizes descend from grammar through content and organization to mechanics, the smallest-weighted category.

Reflection, confidence, and the Ethiopian gap

Both instruments were shortened from validated scales and treated as single composites, and both showed significant adjusted differences. The authors read the Self-Efficacy gain through Bandura's mastery experiences: students repeatedly identified errors, weighed explanations, and revised, which may have built confidence for similar tasks. They are careful about causation: the condition bundled AI feedback, structured reflection, and revision, and the comparison class had no equivalent reflection activity, so no component can be isolated, and no mediation analysis tested the path from awareness to performance. The multilingual setting supplies the contribution: Ethiopian studies have documented grammar and tense as chronic weaknesses — a survey of 680 undergraduates at Mizan Tepi University found vocabulary and grammar most challenging — yet none examines generative AI, while African research concentrates on adoption; a survey of 1,829 Zambian students found 96.8% awareness and 74.8% perceived risk.

What this means for practice

• Instructors. Constrain the prompt. Asking for grammar errors only — with a plain-language rule and a corrected sentence — keeps the task focused and still produced gains in content and organization.
• Instructors. Require an articulation step between feedback and revision. The reflection sheet asked students to name the most frequent error category and explain the rule themselves, which separates noticing from copying a correction.
• Course designers. Schedule the cycle rather than offering the tool; ten weekly repetitions of draft–prompt–reflect–revise produced the effects, and ad hoc access is a different intervention.
• Instructors. Generate practice items from the student's own discipline, as the second prompt did, so structural practice transfers to the writing they actually have to do.

Limitations

• Assignment was by intact class, not random, because the institution required existing sections and timetables; covariate control only partly addresses possible selection effects.
• The ten-week window establishes short-term gains but says nothing about durability, and the authors call for follow-up at three or six months.
• The sample is 88 first-year Natural Science students at one Ethiopian university, which limits generalization to other disciplines, institutions, and academic years.
• Metalinguistic reflection and self-efficacy are self-reported, so the results describe perceived capability, and the design cannot separate AI feedback from reflection or revision.

Citation

Wordofa, Y. J., & Abo, A. L. (2026). Effects of generative AI assisted grammar instruction on Ethiopian EFL learners' self-efficacy, metalinguistic awareness, and writing performance. Discover Education.