{"id":23317,"date":"2026-09-17T14:42:42","date_gmt":"2026-09-17T12:42:42","guid":{"rendered":"https:\/\/djalil.chafai.net\/blog\/?p=23317"},"modified":"2026-09-26T23:18:22","modified_gmt":"2026-09-26T21:18:22","slug":"ai-effect-on-arxiv-math-pr","status":"publish","type":"post","link":"https:\/\/djalil.chafai.net\/blog\/2026\/09\/17\/ai-effect-on-arxiv-math-pr\/","title":{"rendered":"AI effect on arXiv math.PR ?"},"content":{"rendered":"

\"\"<\/a><\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
GPT-4<\/strong><\/td>\nMarch 2023<\/td>\n<\/tr>\n
GPT-4o<\/strong><\/td>\nMay 2024<\/td>\n<\/tr>\n
o1 (full version) first reasoning model<\/strong><\/td>\nDecember 2024<\/strong><\/td>\n<\/tr>\n
GPT-4.5<\/strong><\/td>\nFebruary 2025<\/td>\n<\/tr>\n
GPT-5<\/strong><\/td>\nAugust 2025<\/td>\n<\/tr>\n
GPT-5.1<\/strong><\/td>\nNovember 2025<\/td>\n<\/tr>\n
GPT-5.2<\/strong><\/td>\nDecember 2025<\/td>\n<\/tr>\n
GPT-5.3-Codex<\/strong><\/td>\nFebruary 2026<\/td>\n<\/tr>\n
GPT-5.4<\/strong><\/td>\nMarch 2026<\/td>\n<\/tr>\n
Claude Sonnet 4.6<\/strong><\/td>\nFebruary 2026<\/td>\n<\/tr>\n
GPT-5.5<\/strong><\/td>\nApril 2026<\/td>\n<\/tr>\n
Claude Fable 5<\/strong><\/td>\nJune 2026<\/td>\n<\/tr>\n
GPT-5.6<\/strong><\/td>\nJuly 2026<\/td>\n<\/tr>\n
Claude Opus 5<\/strong><\/td>\nJuly 2026<\/td>\n<\/tr>\n
GPT-6 Astra<\/strong><\/td>\nSeptember 2026<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\"Time<\/a>
Time series analysis (with AI)<\/figcaption><\/figure>\n

\nThis post is about a natural question: we all feel a rise in the number of submissions on arXiv. But what is the quantitative reality, and what are the plausible conclusions? To explore this question, I have prompted ChatGPT to produce a gentle Python script to extract the monthly submissions to arXiv.math.PR, over the last ten years, using the arXiv API. This is more subtle than one can imagine. I have then prompted ChatGPT to produce the graphics and a time series analysis, keeping in mind a possible Covid effect and AI effect. What do we see? A global increase, with a long pause plausibly due to Covid, and a recent (2026) rise plausibly due to AI.\n<\/p>\n

\nWe bring together a descriptive analysis of monthly counts of articles classified under probability on arXiv, exploratory detection of trend changes, sensitivity checks, an examination of additional breaks over 2023\u20132026, and a discussion of possible explanations, including the Covid-19 pandemic and the rise of artificial intelligence. The analysis covers 120 consecutive months and 41 013 articles, including both primary and secondary classifications; the complete monthly counts are reproduced in Appendix A. The main model, a continuous piecewise-linear trend with monthly seasonal effects, selects breakpoints in October 2020, March 2023 and February 2026. These dates are not invariant across models: the most stable signal is a sharp acceleration in late 2025\u2013early 2026, preceded by a slowdown and then a recovery whose dates are harder to pinpoint. Within the 2023\u20132026 window, an additional inflection in mid-2023 remains possible, whereas the segmentations of 2024\u20132025 are less stable. The aggregate series cannot attribute these developments to a particular cause or establish the persistence of the latest regime.\n<\/p><\/blockquote>\n

Further reading.<\/strong><\/p>\n