AI analysis on AI impact on hep-th publications.
By Philip Chang
Jerry Ling shared me a link to a post on Not Even Wrong: Looks Like it is Happening…. Woit noted that arXiv hep-th submissions seem to have roughly doubled around December 2025. December 2024 had 780 submissions, December 2025 had 1192. He was speculating about AI-generated papers and I wanted to check.
I’ve been using Claude Code a lot these days. And he said that more analysis could be done with AI, so I did. I asked it to verify Woit’s numbers and in just a few minutes it confirmed them. (One month was off by 2, 686 vs 684, probably a typo.)
But I noticed there were a different search option. So I asked it to re-run the same searches but with “Submission date (original)” instead of “Submission date (most recent)”. Woit used the default, “most recent”, which counts a paper based on when its latest version was uploaded. But this means if someone revises a 2020 paper in December 2025, it shows up as a December 2025 submission.
With this change, the “doubling” largely goes away!
I then asked it for essentially two more prompts; do the study as far as back as 2018, and do it for hep-ex, hep-ph, and cs.AI. It did all of those and made a nice summary of the entire analysis in just about 10 minutes.
It’s scary what these AI can do these days.
Data, scripts, and plots: sgnoohc/arxiv-submission-analysis.