July 20, 2026
Publish or Per-ish
Over 30% of new ArXiv submissions now read as AI-written
Scientists are side-eyeing papers as readers split between panic, shrugs, and memes
TLDR: A new study says roughly a third of recent arXiv papers now look AI-assisted, with computer science far ahead of other fields. Commenters instantly split into two loud camps: people panicking that science is being machine-styled into blandness, and people shrugging that good research still matters more than who wrote the sentences.
A fresh arXiv analysis just dropped a stat guaranteed to start fights in every academic group chat: more than 30% of new papers now read as if artificial intelligence helped write them, with computer science soaring to about 65%. The researcher behind the study tried to head off the usual detector drama by setting the system so it almost never wrongly flags older, pre-ChatGPT papers. In plain English: the author is saying this spike is too big to dismiss as a glitch.
But the real fireworks were in the comments. One camp basically yelled, “So what?” Their argument: if the paper is useful, clear, and adds something valuable, why should anyone care whether a chatbot polished the sentences? That cooler-headed take immediately collided with the doom crowd, who warned that even if authors aren’t fully outsourcing the work, AI style is starting to infect academic writing itself. In other words, the fear isn’t just fake papers — it’s a whole generation of researchers learning to sound like the machine.
And then came the drama queen entry: one commenter called it a “Weimar moment of science,” which is the kind of over-the-top comparison that practically begs to be screenshotted. Others were oddly impressed, praising the detector for catching even hybrid human-AI writing. So the mood was a glorious mess: alarm, resignation, jokes, and a giant philosophical brawl over whether robotic writing is ruining science or just making it more efficient.
Key Points
- •The study calibrated its academic-writing detector to a 0.4% false-positive rate using 2021-2022 pre-ChatGPT papers as the control baseline.
- •Researchers sampled 12,750 arXiv papers across ten field groups from January 2023 to July 2026, plus control months in 2021 and 2022.
- •The share of papers flagged as AI-written remained at 0.4% in 2021-2022, then rose after ChatGPT’s release to about 32% in the latest complete quarter and peaked near 39% in early 2026.
- •Field-level rates varied widely, with computer science highest at 65.0% over the 12 months to July 2026 and mathematics lowest at 0.7%.
- •The article states that low scores in mathematics may reflect detector blind spots rather than low AI use because math prose differs from the scientific English the detector was trained on.