AI Writing Revision: Does the Argument Still Say What You Mean?

AI writing revision can change how an argument positions its claims and responds to other viewpoints. Zhao and Ji’s (2026) study found different patterns in students’ revisions after peer feedback and revisions generated by ChatGPT-4.

The differences concerned evaluative language and the space given to alternative perspectives. The study offers a reason to examine what a revision does to an argument, even when its wording appears clearer or more polished.

When I read an AI-revised paragraph, I want to know whether its degree of certainty still matches the writer’s intention and evidence. A qualification can make a claim more cautious. An assertion can make it harder to challenge. Both choices may be appropriate, but they need to belong to the person making the argument.

AI writing revision involves choices about stance

Argumentative writing involves taking a position while responding to possible disagreement. Writers decide which claims to present confidently, which to qualify, and how to represent competing interpretations. These decisions shape the reader’s understanding of the argument.

The paper calls this relationship to other viewpoints “dialogic positioning.” In plain terms, language can open space for different interpretations or narrow that space by rejecting alternatives and presenting a claim more firmly. The authors analyse these choices using Appraisal Theory, a framework for examining evaluation in language.

Its Engagement component examines how a text positions itself in relation to other voices. Attitude concerns expressions of feeling and evaluation, while Graduation concerns their strength or intensity. These distinctions help explain why counting evaluative expressions alone tells us little about whether a revision serves the writer’s purpose.

For educators working on AI and teaching practice, this adds a useful question to revision activities: what has changed in the relationship between the writer, the claim, and the reader?

Summary of the Study

Hengxuan Zhao and Jiaojiao Ji (2026) studied 14 undergraduate English majors at a university in China. Across four argumentative writing tasks, students produced first drafts and revised them after peer feedback.

The researchers also generated ChatGPT-4 revisions from each first draft using a standardised procedure based on the writing criteria used in instruction. This produced 168 texts across the three draft types. The researchers manually coded appraisal resources and compared their use and distribution.

AI-generated revisions contained more appraisal resources overall than students’ peer-feedback revisions, which contained more than first drafts. The increases mainly involved evaluative language and expressions that adjusted intensity. The overall frequency of Engagement resources remained comparable, but their distribution differed.

Peer-feedback revisions moved towards a more contractive stance, with greater assertion and less openness to alternatives. AI revisions maintained a relatively balanced distribution of language that opened and narrowed space for other viewpoints.

These findings describe textual patterns in a small, specific writing context. They do not establish that AI revisions were better arguments or that students learned to make these choices independently. The AI versions were generated by the researchers as comparison texts, so the study also cannot show what students would accept or reject during their own AI-supported revision.

The authors acknowledge that the model and standardised prompt may have influenced the patterns. I would therefore treat the results as material for examining rhetorical choices, with no assumption that one distribution is desirable for every argument.

Practical insights: compare the choices behind the revisions

The authors propose using peer-feedback revisions and AI versions as complementary teaching resources. Students could revise their drafts after peer feedback, compare them with generated versions, and discuss how each handles disagreement and certainty. This is a proposed application of the findings; the study did not test whether the comparison improved writing development.

A useful starting point would be one paragraph. Ask students to identify an expression that strengthens or softens a claim, then explain whether the change fits their evidence and intended meaning. They could also examine how an opposing view is introduced: is it represented fairly, acknowledged briefly, or dismissed without sufficient justification?

This gives students something specific to evaluate. “Which paragraph sounds better?” can encourage a preference for fluency. “Which version makes the claim you can defend?” directs attention to the relationship between wording and reasoning. The task connects naturally with the broader concern about AI writing tools changing voice and meaning.

Teachers can also use the comparison to discuss the assumption that stronger assertion always makes a stronger argument. The authors caution that neither expansive nor contractive positioning is inherently superior.

A well-supported claim may warrant direct language. An uncertain inference may require qualification. Students need to connect those choices to what they are arguing and what their evidence permits.

My suggested addition would be a brief explanation of one accepted and one rejected revision. Students would identify the wording, describe its effect, and justify their decision.

This would make their judgement visible without requiring them to adopt the AI version. A revision becomes educationally useful when students can explain what it changes and decide whether that change expresses the argument they intend to make.

Reference

Zhao, H., & Ji, J. (2026). Human-written and AI-generated revisions: An appraisal analysis of dialogic positioning in argumentative writing. Frontiers in Education, 11, Article 1904852. https://doi.org/10.3389/feduc.2026.1904852

The AI in Education Research Digest

Keep exploring the research with me.

Get my reviews of new studies on AI, teaching, and learning, with thoughtful takeaways for your work as an educator.

Free to join. Unsubscribe anytime. Privacy policy.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top