Over 30% of new ArXiv submissions now read as AI-written
13 hours ago
- A study measured AI-written content in arXiv papers using a calibrated detector with a 0.4% false-positive rate on pre-ChatGPT human text.
- About a third (32%) of recent arXiv papers from 2023-2026 read as machine-written, with peaks near 39% in early 2026, showing a significant rise post-ChatGPT.
- Field analysis reveals wide variation: Computer science leads at 65% flagged, while mathematics is lowest at 0.7%, partly due to detector limitations with notation-heavy prose.
- Limitations include small control sample sizes per field, potential detector blind spots (especially in mathematics), and inability to distinguish between heavy AI assistance and full generation.
- The detector provides a lower bound on AI adoption; reported prevalence is conservative due to incomplete coverage of diverse models and prompts used by authors.