~32% of Recent arXiv Papers Show Signs of AI-Generated Writing, New Study Finds

What happened

A new quantitative study analyzed arXiv preprints and found that ~32% of recent academic papers show characteristics of machine-generated text, with sharp differences between disciplines.

Context and impact

The study uses a calibrated detector with a 0.4% false-positive rate on pre-ChatGPT papers as a control baseline. The flagged share grew in two distinct waves — after ChatGPT (Q4 2022) and after GPT-4 (Q1 2023). This has serious implications for scientific integrity, peer review, and the reliability of academic research output.

Details

• ~32% of total arXiv preprints show signs of AI-generated text (most recent complete quarter)
• Computer Science: 65% flagged articles
• Mathematics: only 0.7% flagged — nearly unaffected
• Methodology: calibrated detector with 0.4% false-positive rate on pre-GPT baseline
• Growth in two distinct waves: first after ChatGPT, second after GPT-4
• Raises serious questions about peer review processes and academic integrity