AI joins the rewrite: systematic review maps reformulation feedback in L2 writing

Reformulation, a feedback technique in which a learner’s draft is rewritten while its meaning is preserved, has quietly shaped second language writing instruction for more than four decades. A new systematic review now maps the full landscape of this practice, including its most recent transformation: the arrival of generative artificial intelligence as a source of rewritten texts. Published in the Asian-Pacific Journal of Second and Foreign Language Education, the review synthesizes 26 empirical studies published up to January 13, 2026, and finds a field in transition, dominated for years by human reformulators but now beginning to absorb tools such as ChatGPT.

The review, conducted by Jie Yin and Rozita Che Omar, followed the PRISMA framework for systematic reviews. The authors searched four major databases, Web of Science, Scopus, ERIC, and ProQuest, and retrieved 670 records. After duplicates were removed, 258 records underwent title and abstract screening against eligibility criteria: the study had to involve second language writing, investigate reformulation as a feedback practice, and report original empirical data. Twenty-six studies survived the screening, and none were excluded at the full-text stage. The authors note that their work differs from earlier reviews of positive evidence feedback or model texts by offering a synthesis focused specifically on reformulation, covering both human-generated and emerging AI-generated practices.

Reformulation itself is defined in the review as a draft-based, meaning-preserving rewriting of a learner’s original text into a more target-like version, whether produced by a human or an AI system. The authors distinguish it from direct written corrective feedback, which explicitly corrects discrete linguistic errors, and from model texts, which present one or two good writing samples to all learners. Reformulation, by contrast, is individualized: it preserves each learner’s intended meaning while presenting it in more accurate grammatical, lexical, cohesive, and stylistic form. Because reformulated texts are built directly on learners’ own writing, the review explains, they are assumed to promote deeper cognitive engagement as learners compare versions and search for better linguistic alternatives.

The theoretical backbone of the field, as the review describes it, is a three-stage process: composing, in which learners produce an original draft; comparing, in which they examine a reformulated version to notice differences; and rewriting, in which they revise based on what they noticed. This sequence aligns with the writing-to-learn perspective in second language acquisition, drawing on the Noticing Hypothesis and the Output Hypothesis. The review also frames reformulation through a sociocultural lens, in which the reformulated text functions as a mediating artifact and collaborative dialogue, teacher guidance, or peer discussion provides scaffolding within learners’ zone of proximal development.

Who has been studied, and where? The review found that research has concentrated heavily on university-level learners, who constituted 73 percent of participants across the corpus, with most at intermediate proficiency or higher. Geographically, 45 percent of studies were conducted in Asia and 27 percent in Europe, with additional contributions from North America and Australia. Most studies took place in English as a Foreign Language settings, 64 percent, followed by English as a Second Language settings at 32 percent. The authors contrast this demographic profile with research on model texts, which has focused more on young learners, and they caution that the relative neglect of younger or lower-proficiency populations raises questions about how generalizable current findings are to broader pedagogical settings.

Methodologically, the picture is strikingly uniform. Quasi-experimental designs dominated the learner-focused studies, accounting for 83 percent, and 83 percent of studies were situated in classroom-based environments. Narrative writing tasks were the disproportionate focus, appearing in 67 percent of studies, a preference the authors attribute to the salient lexical concreteness and temporal cohesion of narrative tasks. Quantitative designs appeared in 54 percent of studies and mixed-methods approaches in 33 percent, with the remainder being case studies. Individual writing was the most frequent mode of production, though collaborative writing received substantial attention at 38 percent. Sample sizes varied enormously, from 2 participants in an intensive case study to 162 in a comparative study, with a mean of 39.29.

The review also flags persistent methodological gaps. Only six learner-focused studies reported effect-size information, which the authors note limits interpretation because statistical significance alone does not indicate the magnitude or practical importance of an intervention effect. Only six studies employed delayed posttests, meaning most evidence concerns short-term outcomes and offers little insight into whether learners retain reformulated features over time or transfer them to new writing tasks. A quality appraisal informed by the Mixed Methods Appraisal Tool rated 19 studies as high quality and 7 as moderate, with none rated low; the authors emphasize that these ratings were used to qualify interpretation rather than to exclude studies, and that the synthesis should be read as a descriptive account rather than conclusive evidence of reformulation’s effectiveness.

The clearest sign of change lies in who or what produces the reformulation. All studies published before 2024 relied on human reformulators, including native speakers, teachers, and researchers, who remain the primary providers at 92 percent of the corpus. AI-generated reformulation first appeared in 2024 with one study and expanded to three studies in 2025. Among these, one learner-focused study reported generally positive learner perceptions of AI-generated reformulation, while two feedback-focused studies evaluated the linguistic quality of AI output itself. One of these compared 140 reformulations of narrative texts, 35 human-generated and 105 AI-generated, focusing on text length, lexical coverage, lexical diversity, and vocabulary incorporation. The other used a much larger corpus of 1,200 argumentative texts to examine syntactic complexity, lexical sophistication, lexical diversity, and cohesion. Together they suggest that AI-generated reformulations can be linguistically advanced and lexically diverse, though quality may vary in cohesion, prompt specificity, and similarity to learners’ original drafts.

On the central question of whether reformulation works, the review reports a divided field. Reformulation-only studies, 11 in total, were largely process-oriented, with learners frequently noticing lexical and grammatical features during composing and comparing stages; one study reported noticing rates of 86 to 92 percent for grammatical and lexical features. Reformulation-integrated studies, 15 in total, compared reformulation with error correction, model texts, other corrective feedback types, editing feedback, and self-correction, and produced mixed findings. Error correction was more frequently reported to lead to more accurate revisions than reformulation, and one study found that only error correction improved writing on new tasks, while another found reformulation improved complexity and delayed fluency gains. Comparisons with model texts suggested complementary strengths: model texts promoted noticing of covert or lexical features, whereas reformulation directed attention toward grammatical, spelling, or overt features. One study found that combining reformulation with a model text and pair discussion may enhance uptake.

The review’s authors conclude that reformulation research stands at an inflection point. Generative AI tools may offer immediate, scalable, individualized draft-based feedback, potentially broadening access in resource-constrained contexts, but the evidence base remains preliminary and limited with respect to output quality, prompt design, teacher mediation, and ethical considerations, including data privacy when learners’ drafts are submitted to AI systems and the risk of learners overrelying on AI feedback at the expense of independent noticing and revision. The authors position AI-generated reformulation as an emerging extension of the practice rather than a replacement for human reformulators, and they call for future studies that systematically compare human-led, AI-led, and AI-assisted reformulation across diverse genres, learner groups, and instructional contexts, with delayed posttests to capture whether the cognitive engagement reformulation is meant to spark translates into durable learning.

Subject of Research: Systematic review of human- and AI-generated reformulation as feedback in second language writing

Article Title: A systematic review of human- and emerging AI-generated reformulation in L2 writing

Article References: Yin, J., & Omar, R. C. (2026). A systematic review of human- and emerging AI-generated reformulation in L2 writing. Asian-Pacific Journal of Second and Foreign Language Education, 11(1), Article 91. https://doi.org/10.1186/s40862-026-00465-x

Image Credits: AI Generated

DOI: 10.1186/s40862-026-00465-x

Keywords: reformulation, L2 writing, generative AI, ChatGPT, written corrective feedback, systematic review, PRISMA, noticing, uptake, EFL, quasi-experimental design, model texts