Chinese ESL Students Treat AI as a Co-Author, Reshaping Academic Writing

Generative artificial intelligence has moved from novelty to daily companion for millions of students learning to write in English, and a new review argues that the relationship is far more profound than a simple matter of shortcuts and cheating. Writing in Frontiers of Digital Education, a team of researchers led by Jason Chan of MCI Management Center Innsbruck, with colleagues at the University of Hong Kong, the University of Saint Joseph in Macao, and the University of Cambridge, examines how Chinese secondary and tertiary English-as-a-second-language learners use tools such as ChatGPT, Claude, Doubao, and the Chinese automated writing evaluation platform Pigai. Their central claim is striking: students increasingly treat these systems not as mechanical aids but as coauthors in the academic writing process, and that shift demands an entirely new theoretical and pedagogical framework, which the authors call human–AI co-learning.

The context matters enormously. Chinese classrooms operate within one of the world’s most assessment-centered education systems, one that traditionally emphasizes memorization, structured learning, and mastery of prescribed forms. Decades of research have noted the tension between this transmission-oriented tradition and the dialogic, inquiry-driven learning celebrated in Western pedagogy, a tension often cited in explanations of why Chinese postgraduates sometimes struggle with critical thinking when they arrive at universities abroad. Against that backdrop, generative AI arrives as something genuinely disruptive. Because large language models respond conversationally, they invite learners to question, negotiate, and iterate rather than simply to absorb. The review argues that this dialogic quality opens new possibilities for critical thinking inside a system that has historically left little room for it.

Technically, the linchpin of this interaction is prompt engineering, and the authors are careful to frame it as both a technical skill and a pedagogical one. A well-constructed prompt specifies the rhetorical task, the audience, the genre, the desired structure, and the constraints of the output; a poorly constructed one yields generic, sometimes confidently wrong prose. Research cited in the review shows that students who experiment with chain-of-thought prompting, role assignment, and iterative refinement engage in precisely the kind of metacognitive reflection that writing instructors try to cultivate. Deciding what to ask the model, evaluating what comes back, and revising the request forces learners to articulate their own goals and standards. In other words, the act of engineering a prompt can itself become a lesson in rhetoric, genre awareness, and self-regulated learning.

The scaffolding metaphor from second-language acquisition runs throughout the analysis. Generative AI can model academic phrasing, supply grammatical corrections, generate outlines, and provide feedback at a scale no human teacher could match, effectively extending the tradition of automated written corrective feedback into a far more responsive domain. Studies of GPT-4 and of dedicated grammar-checking systems suggest that such feedback can meaningfully support language production, particularly for learners who lack access to native-speaker interlocutors or one-on-one tutoring. Just as important, the review emphasizes student agency: when learners direct the interaction, decide which suggestions to accept, and justify their choices, they retain authorship of the text while borrowing computational muscle. The tool functions as a scaffold that can be gradually dismantled as competence grows, echoing classic computer-based scaffolding research in STEM education.

Yet the authors refuse to tell a purely triumphant story, and their treatment of risk is among the most substantive parts of the paper. The first danger is epistemic dependency: a learner who accepts every AI suggestion without scrutiny outsources judgment itself, hollowing out the very cognitive processes writing is supposed to develop. The second is the erosion of critical thinking, a particularly acute concern in a context where rote learning already discourages independent argument. The third is ethical ambiguity. Models hallucinate, fabricating citations and distorted facts with fluent confidence, a phenomenon now extensively documented in the literature on AI-generated content. Students who lack critical AI literacy may not even recognize when a source is invented, and the boundary between legitimate assistance and academic misconduct, sometimes dubbed ‘AI-giarism,’ remains contested among students and institutions alike.

These risks collide with specific features of the Chinese educational environment. High-stakes examinations such as the National College Entrance Examination shape what counts as good writing, and curriculum standards set by the Ministry of Education channel classroom practice toward measurable outcomes. Automated evaluation platforms like Pigai, which has accumulated more than a decade of use in Chinese English writing instruction, fit comfortably into that assessment culture because they score quickly and consistently. But the review warns that efficiency-oriented adoption could entrench the very habits—formulaic structures, surface-level correction, answer-seeking—that generative AI could otherwise disrupt. The technology’s pedagogical value, in other words, is not intrinsic; it depends on how educators mediate its use.

That is why the authors advance human–AI co-learning as a theoretical framework rather than a slogan. In their account, the learner and the model form a reciprocal system: the student supplies intent, disciplinary context, and critical evaluation, while the machine supplies fluency, alternatives, and tireless feedback. Each interaction potentially improves the human partner’s understanding, even though the model itself does not learn from the exchange in any pedagogical sense. Co-learning, properly understood, means the human side of the loop must be taught—explicitly, deliberately, and critically. Prompt engineering becomes part of digital literacy; rhetorical awareness becomes the standard against which AI output is judged; and metacognition becomes the habit of asking not just ‘is this correct?’ but ‘why is this appropriate, and did I decide it?’

The practical recommendations follow directly. The review calls for critical AI literacy to be built into curricula, so that students learn the capabilities and failure modes of large language models alongside their own writing development. It calls for educator mediation, noting recent intervention studies showing that language teachers can develop professional generative AI competence when given structured training, and that teacher guidance dramatically changes how productively students use these tools. And it calls for culturally responsive pedagogy—an approach that reconciles traditional Chinese learning practices, with their genuine strengths in disciplined practice and textual mastery, with reflective engagement in digital environments, rather than treating memorization and critical inquiry as enemies. Freirean dialogue, the authors suggest, need not require abandoning the intellectual virtues of the Chinese classroom.

The paper’s closing reframing may prove its most quotable idea: generative AI should be understood not as a shortcut but as a scaffold. A shortcut moves a student from question to answer while bypassing learning; a scaffold supports the climb while the learner’s own muscles do the work. By redesigning writing instruction around prompting, evaluation, and revision, the authors propose a pedagogical model in which learners reclaim authorship and engage more deeply in academic inquiry than either pure human drafting or pure machine generation would allow. For the hundreds of millions of ESL learners worldwide now staring at a chat window instead of a blank page, the stakes of getting that distinction right could hardly be higher.

Beneath the framework lies a body of empirical work the review draws on selectively. Eye-tracking studies of second-language learners engaging with automated feedback reveal that attention is uneven: students often skim surface corrections while overlooking deeper rhetorical comments, a finding that cautions against assuming AI feedback is absorbed simply because it is delivered. Similarly, evaluations of GPT-4 as a source of written corrective feedback show genuine grammatical competence but inconsistent judgment about discourse-level issues, reinforcing the review’s insistence that a human evaluator remain in the loop.

The theoretical lineage is also broader than the co-learning label suggests. The authors situate their argument within classic scaffolding theory, which holds that support should be contingent—offered when needed and withdrawn as competence develops—and within genre-based approaches to academic writing that treat text forms as socially situated choices rather than templates. Their invocation of Bloom’s taxonomy underscores a worry echoed across the AI-in-education literature: if learners delegate the lower-order tasks of drafting and polishing, the higher-order work of analysis and evaluation must be deliberately designed into instruction, or it will simply not occur.

Practical constraints receive attention too. Surveys of privacy and security in generative AI document real risks when students paste unpublished manuscripts into commercial systems, and the review’s emphasis on institutional mediation implicitly includes data governance. At the same time, teacher-facing studies suggest optimism is warranted: structured professional development can move language instructors from anxiety about misconduct toward confident integration, and students themselves report more productive use when teachers model critical prompting rather than banning the tools outright.

What emerges is a picture of a field still defining its terms. Whether human–AI co-learning becomes a durable theoretical contribution or a transitional description will depend on longitudinal evidence the review, by its own admission as a conceptual synthesis, cannot yet supply.

Subject of Research: Human–AI co-learning in generative AI–assisted academic writing among Chinese ESL learners

Article Title: Human–AI Co-Learning in Academic Writing Among Chinese ESL Learners

Article References: Chan, J., Wong, J., Ieong, J., & Pang, H. (2026). Human–AI Co-Learning in Academic Writing Among Chinese ESL Learners. Frontiers of Digital Education, 3(2), Article 15. https://doi.org/10.1007/s44366-026-0089-8

Image Credits: AI Generated

DOI: 10.1007/s44366-026-0089-8

Keywords: generative AI, prompt engineering, human–AI co-learning, Chinese ESL learners, academic writing pedagogy, AI literacy, ChatGPT, scaffolding, critical thinking, digital literacy, student agency, automated writing evaluation