Researchers are increasingly using ChatGPT to polish their papers, yet recent GPT releases often narrow or even retract the authors' claims. We term such unsupported modifications defensive writing and examine two possible causes: the model corrects authors' over‑claiming, or it writes for an anticipated reviewer. We asked various GPT versions and models from other developers to rewrite paragraphs from papers written before ChatGPT, or to compose text from an evidence sheet that lists a paper's method and results. Defensive writing grows with newer GPT versions. GPT‑6‑astra outright retracts the authors' claims, and even when writing from the evidence sheet it adds the most ungrounded qualifications. The findings favor the anticipated‑reviewer explanation, while correcting over‑claiming accounts for only a small portion. When models are asked merely to polish, the level of defense stays near the original; mentioning review raises it, and a single round of self‑review raises it further. Moreover, fewer than one in ten retracted claims by GPT‑6‑astra are truly overstated. AI reviewers score defensive rewrites higher, whereas human readers find them harder to read and the authors appear less certain. Combining AI writing with AI review may amplify this style.
Review