Over 30% of new ArXiv submissions now read as AI-written
unslop.run
July 20, 2026
5 min read
🔥🔥🔥🔥🔥
60/100
Summary
A study analyzed 12,750 arXiv papers and found that approximately one-third of new submissions appeared to be machine-written. The measurement method revealed limitations, including a tendency for detection tools to falsely identify some genuine human writing as AI-generated.
Key Takeaways
- Approximately one-third of new arXiv papers are identified as machine-written based on a study of 12,750 papers from 2023 to 2026.
- The flagged share of machine-written papers increased from 0.4% in 2021 and 2022 to about 32% in early 2026, with peaks reaching nearly 39%.
- The field of computer science has the highest flagged share at approximately 65%, while mathematics has the lowest at around 0.7%.
- The study's detector has a false-positive rate of 0.4%, effectively distinguishing between genuine human writing and AI-generated text.
Coverage & discussions
Hacker News · unslop.run
Community discussions
Hacker News discussion ↗
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- The spike in AI-detected papers is a wake-up call for researchers — it shows how deeply AI has influenced academic writing styles.
- The detailed analysis of arXiv submissions adds a fascinating layer to our understanding of AI's impact on scientific communication.
- Commenters highlight the potential for LLMs to enhance clarity and precision in scientific writing, which could democratize access to effective communication.
Concerns
- Critics are worried that reliance on LLMs for writing could dilute the originality and authenticity of research papers, raising ethical concerns.
- The fact that 40% of new papers are flagged as AI-written suggests a troubling trend where AI styles dominate human expression in academia.
- There are doubts about the reliability of AI detectors, with some suggesting they might misclassify high-quality human writing as machine-generated.
