arXiv · all indexed categories · August 2025–September 2026
In , of new papers show signs of AI writing, and say AI helped write them.
A small category. This category averages about {n} papers a month, so its monthly percentages move by several points from chance alone: on 80 papers, a true rate of 20% is as likely to show as 12% or 28%. The counts beside each percentage show how many papers a figure rests on.
Every preprint arXiv announced in the month, as its identifier records, and whose LaTeX source could be read, as a percentage of those papers. Across the series of the submissions had no readable source (at most in a month); they are left out of every count and percentage on this page.
Every paper gets a score from 0 to 21 based on selected markers of AI-style writing. A score of 5 or higher is classified as apparent AI writing, and 10 or higher as apparent heavy AI writing, in every area and every category. The histogram shows one month at a time, as a share of the papers scored that month. Play it or step through it; the first month stays outlined behind for comparison.
Higher score → more likely AI use. A paper counts as apparent AI writing at a score of 5 or more, and as apparent heavy AI writing at 10 or more, the last four bars.
Every month at once. Each bar is one month’s papers, divided by score: 5 or more is apparent AI writing, and 10 or more apparent heavy AI writing.
Papers showing almost no markers, scoring 0 or 1, go from % of to % of , while those at 5 or more go from % to %, and those at 10 or more from % to %. See how this is measured and what it does not mean.
| Month | Papers | Disclosed | Apparent AI writing | Apparent heavy AI writing | |||
|---|---|---|---|---|---|---|---|
| AI use | AI writing | Total | Undisclosed | Total | Undisclosed | ||
Undisclosed counts the papers with apparent AI writing, or apparent heavy AI writing, whose authors do not disclose using AI for the writing (the AI, AI* and AI+Human types). Such a paper may still disclose AI use for something else, such as code, computation or proofreading; it is the writing that goes undisclosed. Papers counts the month’s submissions with a readable LaTeX source, and every percentage is of that count; the submissions without one are left out.
Papers that disclose using AI, sorted by what their authors say the model did to the text.
The papers disclosing AI use in , by what they say about the text.
The five types partition the papers that disclose AI use, so each row adds up to the count beside it and the percentages are shares of that count, not of the month’s submissions. A handful of papers state that they used no AI at all; they are not counted as disclosing use and do not appear here. The type is assigned by fixed rules that read the authors’ text, the same rules for every category, corrected by hand where a check found them wrong, and may not be accurate.
Every month occupies two lines: the number of papers, and beneath it the mean rubric score of those same papers, shown only where there are at least two papers, and only where the other means on the page cannot be combined to give fewer papers' scores.
| Month | Papers disclosing AI use | AI writing disclosure type | ||||
|---|---|---|---|---|---|---|
| AI | AI* | AI+Human | Human | Unspecified | ||
Every new arXiv preprint in 28 mathematics categories, grouped into four areas, each counted once under its primary category.
All is every indexed category together. A category is indexed once the word usage of its area has been evaluated; each paper is counted once, under its primary category.
An area is a group of arXiv mathematics categories whose papers used words in much the same way before large language models, and which are therefore scored on the same basis. The indexed areas are:
Not in All. Papers whose primary category is outside mathematics, even when they are cross-listed to a mathematics category. Information Theory (math.IT) and Statistics Theory (math.ST), which are aliases of categories whose home is elsewhere, cs.IT and stat.TH. General Mathematics (math.GM) and History and Overview (math.HO), left out for now.
Small categories. Several categories have fewer than about 100 new papers a month. Their monthly percentages carry wide margins of error, so trends there are best read over several months, and the counts beside each percentage say how many papers a figure rests on.
Each paper’s LaTeX source is scored for markers of AI-style writing and searched for statements about AI use; the scoring was checked on papers from 2016 to 2022.
The corpus is every arXiv preprint whose primary category is an indexed mathematics category and that arXiv announced in the month, as its identifier records (2608 is August 2026). Nothing is sampled and nothing is left out. Each paper is read from the LaTeX source of its first version; later versions are not counted.
Disclosure is counted by searching that source for any mention of AI tools and classifying each statement found by the same fixed rules in every category; statements a check found the rules misreading were read by hand, and some are surely still misread. Apparent AI writing is a score built from a fixed set of markers, primarily based on writing style and vocabulary use. Style and word usage were evaluated separately for each area, against that area’s own papers from before large language models, so that one field’s ordinary vocabulary is not taken for AI writing in another; the threshold is the same everywhere. The indicators are a combination of ideas from AI detection literature, my observations, and Claude's experiments. As a control, the same scoring was run on papers of the same categories from well before large language models, 2016 to 2022: % of those papers passed the threshold ( of ). The error in the other direction is much larger: among the papers whose authors say a model generated or helped write the text, only % ( of ) reach the threshold, and the other % are missed. The apparent AI writing counts are therefore likely a significant undercount. Apparent heavy AI writing asks for a score of 10 or more, twice the threshold: % of the control papers reach it ( of ), and % of the papers whose authors say a model generated or helped write the text ( of ).
I am not publishing the precise scoring method because:
Feel free to contact me if you would like to discuss the method in more detail.
A score is not a verdict, statements about AI use are read automatically and not verified, and using AI is not misconduct.
A score is not a verdict. It counts markers that generative AI tends to produce. It is not evidence that a model wrote a paper, and it says nothing about whether the mathematics is correct, original or worth reading. A paper written entirely by a person can score high, and a paper written largely by a model can score low.
Disclosure categories are interpreted based on text. Nothing an author states has been verified, or attempted to be verified. Authors write disclosures in different ways, some more vague than others, and judgement calls were largely automated. For instance, a paper written by a person from a model’s draft and a lightly edited model draft may be described identically, and some papers whose authors state the text is theirs carry the same markers as papers that disclose drafting. That may mean the prose was written by a person working from a model’s write-up, or that there is more model writing than the statement suggests. Disclosure categories, and most disclosure statements themselves, do not distinguish these.
Using AI is not misconduct. Nothing here is intended as an accusation, and I did not personally check or read most papers with high scores. The point of the study is the shape of the trend and the gap between use and disclosure, not any individual paper. That said, I personally believe authors should try to accurately and clearly disclose their use of AI at this time.
One repository. This is mathematics on arXiv. Other fields and other repositories may look nothing like it.