What Happens to Citations When AI Rewrites Academic Text? A Large-Scale Paired Audit
Abstract
Working paper and evidence data. A paired audit of 213,881 in-text citation marks across 60,786 human-AI rewrite pairs from eight model configurations in six families. 96.9 percent of citation marks survived rewriting unchanged; 2.26 percent were corrupted (dropped, altered, or without a source counterpart), and the corruption rate varied 28-fold across models on identical inputs. Files: the per-citation evidence extract (one row per citation mark with verdict and match evidence; pair identifiers are opaque and no passage text is included), extraction, alignment, validation, and analysis code, figures, and the paper PDF. Source and rewrite passage texts are not released. Paper page: https://textpulse.ai/research/ai-citation-rewrite-fidelity
Fields of Study & Subjects
Scientific Computing and Data Management
Artificial Intelligence in Healthcare and Education
Computational and Text Analysis Methods
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 20 August 2026