The public release of ChatGPT in November 2022 put a capable text generator in the hands of anyone with an internet connection. For most school and university subjects this was a change in the tools available. For the teaching of writing it reached the object of instruction itself, because the essay, the report and the literature review are both what students are taught to produce and the means by which their learning is judged. Early responses in the scholarly press were position papers that listed opportunities and threats (Kasneci et al., 2023) or concentrated on cheating (Cotton et al., 2024). A rapid review of the first months of publication dealt mostly with the tool's performance on tasks in different subjects and with its potential uses and risks (Lo, 2023).

Studies conducted with students and teachers have since begun to appear. They are small, they are concentrated in higher education and in second language writing, and they vary a good deal in rigour. They are nonetheless a different kind of source from commentary, and teachers who must decide what to permit, what to teach and what to assess are entitled to know what they show. This article reviews that body of work. Its title promises what classroom studies show so far, and the qualification is meant seriously: the tools change faster than studies can be published, and several conclusions below may not hold for later systems.

The review is narrative. It covers peer-reviewed studies and widely cited reports, published since the arrival of general-purpose chat models, that examine generative artificial intelligence in the teaching, learning or assessment of academic writing. Studies of the software alone are included only where they bear directly on a classroom decision, as with scoring and detection. Earlier research on automated writing evaluation and on the psychology of composition is drawn on for context. No systematic search was carried out and no counts of sources are offered; the selection reflects the studies that have shaped discussion in the field, and other reviewers might reasonably weigh them differently.

Any judgement about a writing tool rests on a view of what learning to write involves. The cognitive process model of Flower and Hayes (1981) described composing as the recursive interplay of planning, translating ideas into text and reviewing, all under the writer's own monitoring. Bereiter and Scardamalia (1987) distinguished novices, who tell what they know in the order it occurs to them, from experts, who transform their knowledge in the effort to meet a rhetorical goal. On this account the difficulty of writing is where the learning happens. The meta-analysis of instruction for adolescents by Graham and Perin (2007) found the largest gains from teaching strategies for planning, revising and summarising, which are precisely the operations a language model can now perform on a student's behalf.

Two concepts frame the risk and the promise. Cognitive offloading, the use of external aids to reduce mental demand, is efficient for the task in hand and can leave the underlying capacity undeveloped (Risko and Gilbert, 2016). Scaffolding, a term from the sociocultural tradition in educational psychology, is help that lets a learner do with assistance what will later be done alone. The same prompt to a chatbot can be either, and the difference lies in what the student does next. Much of the classroom evidence can be read as an attempt to find out which of the two is occurring. Figure 1 sets out the uses that studies have examined across the stages of composing.

The studies are grouped below by the part of the writing process they address. Table 1 gives an overview of what each group reports and how far the evidence extends.

Four weaknesses recur. The first is that performance with the tool is taken as evidence of learning. A better essay written with assistance shows that the assistance works; whether the student writes better unaided is a separate question that needs a later, unassisted task to answer. The second is duration. Interventions typically last a few weeks, which is short against the years over which writing develops. The third is reliance on self-report from volunteers in courses where the tool was permitted, who are unlikely to represent students in general. The fourth is the narrowness of the settings: university students of English dominate, and school pupils, writers in other languages and writers in the disciplines are thinly represented.

A further difficulty is specific to this technology. A study names a model version that may be withdrawn before the article is printed, so exact replication is rarely possible. And the models themselves carry known problems that classroom studies seldom measure: they generate plausible text without reference to truth and reproduce patterns in their training data (Bender et al., 2021), which in academic writing appears as fabricated citations and confident error.

The most useful studies would compare forms of use, since comparing use with non-use is rapidly becoming artificial. Whether students who ask a model to critique their own draft develop differently from students who ask it to produce a draft is a question with direct consequences for teaching, and it can be examined with delayed, unassisted writing as the outcome. Studies should record the interaction itself, the prompts and the uses made of the responses, because the label "used ChatGPT" covers behaviours as different as copying and arguing. Research is also needed on assessment designs that remain informative when a tool is available: staged submissions, oral defence of written work and writing done under observation are widely recommended and little evaluated.

Teachers are part of the evidence as well. How instructors read machine feedback, when they trust it and how it changes the time they give to each student are questions that only a few studies have raised, and the answers will decide whether feedback tools ease teaching loads or add a further layer of checking.

Classroom research on generative artificial intelligence in writing instruction is at an early stage, and its firmest findings are modest ones. Machine feedback on drafts is good enough to be useful where human feedback is scarce and is not yet its equal. Students welcome the tools and are uneasy about them. Detection software cannot bear the weight institutions have placed on it and treats writers of English as an additional language unfairly. On the central matter, whether students who learn to write with these tools become better or worse writers, the studies so far do not give an answer. Teachers should plan on that footing, and researchers should treat the unassisted writing of students, measured after the tool is taken away, as the outcome that matters most.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.