Reproduce “Defensive Writing in GPT Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junchi Liao, Xiao Huang, Yucheng Liu, Jiawen Deng, Fuji Ren

The paper examines AI revisions of scientific claims and defines defensive writing as qualifications unsupported by supplied material. Across 77 paragraphs from 20 papers, GPT-6-astra added 1.30 ungrounded qualifications per 100 words in result analyses, versus 0.04 in originals. Checks found many retractions corrective, but most sampled additions redundant; AI reviewers rewarded them.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.