Using AI to Teach Writing

Writing can help students develop an argument, test it against evidence, and revise their thinking. Research on writing-to-learn and writing instruction supports carefully designed writing activities, but these school-based reviews do not establish the effects of GenAI on university writing.^{1 2} Design AI use around the thinking students need to practise, then check that they can perform it themselves.

This creates both a problem and an opportunity. Students may use AI to generate submissions without engaging with the material, but a carefully designed activity can also make evaluation and revision visible. The key is deciding which parts of writing are the learning outcome and ensuring the model does not perform those parts for the student.

This article offers practical approaches for any lecturer whose course involves writing, not just those who teach writing explicitly.

Rethinking What Writing Instruction Is For

Begin with the writing capability being assessed: constructing an argument, selecting evidence, organising ideas, or communicating in a disciplinary genre. AI can produce polished prose that is inaccurate or weakly reasoned. A polished final document alone is therefore insufficient evidence of the student’s understanding, whether or not AI was used.

This distinction changes how you design and assess writing tasks. The question to ask of every writing assignment is: what thinking does this task require, and how will I know whether the student has done it? When you can answer that question clearly, you are in a position to incorporate AI as a tool while keeping the intellectual work where it belongs — with the student.

Tasks that require students to make disciplinary judgements, apply specific course sources, defend choices, or demonstrate developing thought can provide richer evidence than a generic final essay. These features do not prove authorship, but they align assessment more closely with the intended learning.

Techniques for the Classroom

AI as First Draft

When critical evaluation is the outcome, an educator-vetted AI-generated draft can become the starting point of the assignment. Students then revise it using course knowledge, research, and disciplinary judgment. Supplying the same draft to everyone avoids requiring accounts and makes the activity easier to compare.

The submission is not the AI draft — it is the revised version, along with an annotated changelog: a document where students identify each significant change they made and explain why they made it. Did they correct a factual error? Add evidence the AI omitted? Restructure an argument that was logically weak? Replace a vague generalisation with a claim they can actually defend?

Assess the annotation for accurate diagnosis, evidence use, and justified revision, rather than the number of changes. Research on prompted learning protocols supports structured reflection in its studied setting; it does not validate an AI revision log as proof of learning or authorship.³

Be explicit about what students must demonstrate. For example: require them to verify every major claim, replace unsupported reasoning, use specified course sources, and justify three consequential revisions. Avoid percentages of “student-authored” sentences, which are difficult to interpret or verify and say little about learning.

AI as Revision Partner

Where the tool and material are approved, AI feedback can be an intermediate step. Ask students to request and review feedback against the published criteria, then submit a short record of which suggestions they accepted or rejected and why. Do not require the full conversation unless it is necessary to the outcome.

In a six-week quasi-experiment with 48 university English as a new language learners, researchers found no detected difference in writing-learning outcomes between GPT-4 feedback and human-tutor feedback. A separate study of 43 learners found mixed preferences. These findings concern that setting and feedback design; they do not establish superiority, equivalence, or independent transfer across university writing tasks.⁴

Ask students to respond to the AI’s suggestions, either by acting on them or explaining why they have not. For example: “The AI suggested I define my terms earlier, but the reader needs the case context before the definition makes sense.” The response makes the student’s judgment visible. Feedback-literacy research provides a rationale for interpreting, judging, and acting on feedback, but a recorded dialogue alone does not establish learning.⁵ Use later independent work to check whether the student can apply the principle.

Whether this saves educator time depends on the quality of the feedback and the review required. Pilot the workflow before making it part of a graded task.

Generating and Defeating Counterarguments

Argumentation is difficult to teach partly because students may not see their own arguments from the outside. AI can generate possible counterarguments for students to evaluate and answer.

Ask students to write their thesis and the main lines of their argument, then prompt an AI to argue against them as forcefully as possible. Their task is to revise their essay so that it addresses the strongest of those objections. This can be done iteratively: write, generate objections, revise, generate objections again.

The output may expose a weakness, but it can also invent an objection or misrepresent the debate. Students should compare counterarguments with course sources and decide which are legitimate before revising.

A written simulation can extend this activity: ask students to respond to several rounds of objections grounded in supplied course sources. Judge their responses and subsequent independent reasoning, rather than assuming a longer interaction means deeper engagement.

Comparative Analysis of AI and Human Writing

Give students two texts on the same topic: one AI-generated, one written by a human expert or, for more advanced classes, two AI-generated texts produced with different prompts. Ask students to compare them analytically, using the evaluative framework from your discipline.

What makes one more persuasive than the other? Which handles evidence better? Where does one simplify something the other engages with carefully? What is present in the human-authored text that the AI version lacks, and vice versa? Use the comparison to help students articulate and discuss disciplinary criteria for quality. Research on evaluative judgment provides a conceptual rationale for comparing examples and judging work, rather than experimental proof that this particular AI comparison improves writing.⁶

For writing teachers, this can become an exercise in voice, specificity, and structure. Avoid assuming that all model output is generic or all human writing is distinctive; ask students to identify evidence for each judgment.

Structured Brainstorming

AI can support structured brainstorming by proposing angles or questions. Students must verify any claimed debate or source and make the substantive choice themselves.

The task is then to make a deliberate choice: which of these directions do they actually find most interesting or defensible, and why? Requiring students to document their brainstorming session and reflect on the choices they made surfaces the intellectual judgement that is often invisible in a polished final submission.

This may help some students begin, but it can also anchor them to the model’s framing. Ask students to generate ideas independently first or compare the AI list with a course-grounded map of the field.

Targeted Language Feedback for Multilingual Students

For students writing in an additional language, surface-level issues can obscure their understanding. Approved AI tools may offer grammar, syntax, or clarity feedback, but quality varies by language and variety and suggestions can erase a student’s intended voice.

Make this support optional unless the institution provides an accessible, approved tool. Teach students to review every edit and preserve meaning rather than mechanically accepting suggestions.

Explain which language features the rubric assesses. If independent language production is itself the outcome, translation or rewriting may replace the capability being tested. Agree any accommodations through institutional procedures and distinguish them from optional writing assistance.

Designing Writing Assignments That Work With AI

Design the assignment around the writing capability students must demonstrate. Use assessment patterns for the target capability to choose appropriate evidence, then adapt the techniques above. A few principles:

Tie the task to course learning. An assignment that asks students to “discuss climate policy” elicits little disciplinary evidence. Asking them to evaluate a specific policy using a framework taught in the course and supplied sources makes the intended reasoning clearer. This improves alignment; it does not make the task immune to outsourcing.

Collect proportionate process evidence. Drafts, selected notes, or annotated changes can support feedback and reflection. Request only what the outcome and rubric require; full transcripts, screenshots, and writing histories can collect personal or irrelevant information and do not prove authorship.

Ask students to justify disciplinary decisions. Require a position grounded in specified course sources and an explanation of why competing interpretations were accepted or rejected. AI can also generate positions and justifications; the wording of a task does not make it resistant to outsourcing. Pair the submission with proportionate, accessible evidence of the target capability.

Use reflection for learning, not authentication. Questions such as “Which claim changed most during revision, and what evidence caused the change?” can make decisions visible. Treat the response as assessable reflection, not proof that no outside assistance was used.

Check independent transfer

After supported revision, ask students to explain a consequential change or apply the same writing principle to a fresh passage without AI, under the published assessment conditions and accommodations. Where retention matters, revisit the skill later. Learning-science research distinguishes performance during practice from enduring learning measured through retention and transfer; it predates GenAI and does not validate this particular writing activity.⁷ Use Evaluating an AI teaching workflow to distinguish improvement in the draft from improvement in the writer.

Integrating AI Writing Tools Into Your Course Policy

Publish concrete AI-use rules before students begin, so they can determine which assistance is allowed, what counts as their contribution, and where to ask questions. Explain the rules alongside examples from the writing process.

Distinguish between phases of the writing process: AI might be allowed for brainstorming but not submitted prose, or allowed for feedback with a short record of consequential suggestions. AI-free writing is also appropriate when independent writing is the outcome. Use the templates in Setting course and assignment AI rules.

Whatever your policy, the most important thing is to pair it with explicit instruction in how to evaluate what AI produces. Students who understand AI’s limitations — its tendency to hallucinate sources, flatten complexity, and produce confident-sounding errors — are better placed to use it responsibly and to recognise when its outputs are not good enough to build on.

Example Prompts

The two prompts below illustrate different roles for AI in the writing process. Use them only with material approved for the selected tool, and adapt them to the published assignment criteria.

For revision feedback

For generating counterarguments

In both cases, remind students that the AI’s feedback is a prompt for their own thinking, not a set of instructions to follow mechanically. Part of the intellectual work is deciding which of the AI’s observations are worth acting on, which are not, and why — a judgement that only the student, with their own understanding of the material, is in a position to make.

References & Footnotes