A new preprint on arXiv examines a side effect of using ChatGPT for paper revision. The authors note that researchers increasingly rely on the tool to edit their manuscripts, but recent GPT versions frequently make changes that narrow or even retract the original claims.
The paper introduces the term "defensive writing" for cases where the model alters claims in a way that the given material does not support. The abstract suggests this behavior is distinct from ordinary editing, as it goes beyond clarifying or tightening prose to actively walking back the authors' assertions.
Because the preprint is the only source here, the findings are presented as the authors' own description rather than independently verified results. The abstract does not specify which GPT versions were studied or how widespread the behavior is, so those details remain open questions.