What happened
The site “AI in Mathematics — interactive arXiv atlas” presents an analysis of AI-related disclosures in mathematics papers on arXiv. It was published on 25 October 2026 and describes its snapshot as current to 23 October 2026. The displayed corpus contains 169,291 first submissions across 30 math.* categories, covering 2023 through 2026. For the current year, the site says its comparable annual figures run from 1 January through 21 October; 2023–2025 are described as the full available years.
The central methodological boundary is unusually explicit. The atlas says it measures disclosures, not all AI use. Its corpus includes first versions of papers whose primary category is math.*; cross-listings are not duplicated. “Disclosed use” can include help with writing, scientific tasks, or using a generative model in experiments. But the site treats mentions of AI, research assistance, and idea assistance as distinct signals. A missing disclosure, it says, is not evidence that an author did not use AI.
The presentation is an interactive chart interface rather than a conventional static report. It offers research charts, chart navigation, a table for the current chart, and downloadable SVG and PNG versions. Users can also download current CSV data and PPTX source charts. Labels and visual styling can be edited directly in the browser; those edits are stored only in the user’s browser and are incorporated into downloaded SVG or PNG files. Filters change the view of the saved snapshot, while newly submitted papers are not automatically added.
The atlas also includes country-level material, with a significant coverage qualification. At least one affiliation country was identified for 27,208 of 50,658 papers in the 2026 portion of the corpus, or 53.7%. The country is the country of an institution, not an author’s citizenship, and international papers are counted in every identified country. The site cautions that missing country data can bias comparisons. It attributes the paper metadata and texts to arXiv v1 records collected for the 2023–2026 corpus, and says country data came from OpenAlex on 24 October 2026.
Finally, the authors caution against causal readings. Correlations between mathematical areas and author-name productivity do not establish an AI effect. A first disclosure need not be the beginning of AI use. “Names” are not equivalent to reliably established individual people. The site also says its empirical forecast range is not a formal 95% interval. These limitations make the atlas a record of visible reporting behavior in a defined paper corpus, not a census of adoption or a causal study of AI’s effect on mathematical output.
What the archive adds
This is a new evidence source for a thread opened by Terence Tao’s Math 2.0 slides, which argued that mathematics may be moving from proof scarcity toward proof abundance. Tao’s material concerned the possible shape of mathematical work; this atlas instead records how authors in a defined arXiv corpus disclose AI-related assistance. It is a new fact on a known story, not evidence that either claim causes the other.
Trend signal
The atlas modestly supports Scientific agents need artifact evidence. It does not evaluate an agent, prove scientific capability, or validate mathematical results. But it packages a claim about AI use with corpus boundaries, source descriptions, downloadable data, editable and exportable charts, and stated failure modes—the ingredients that let another reader inspect, challenge, or extend the analysis rather than accept a headline number.
Connections to other events
The earlier Math 2.0 note frames a future in which formal checking can help handle a larger supply of proofs. This atlas concerns a narrower and earlier layer: what paper authors visibly disclose about AI assistance. Together, they underline that claims about AI and mathematics need separate evidence for adoption, output, proof validity, and causal impact.
Why it matters
Teams building AI systems for research often encounter a familiar temptation: treat a visible proxy as the thing being measured. The atlas provides a practical counterexample. It calls its object “disclosures,” separates categories of assistance, fixes a corpus and date range, and names the missing data that could distort country comparisons. That is useful even outside mathematics.
If a team publishes usage research, adoption dashboards, or claims about an agent’s impact, it should make the same distinctions. A model mention is not necessarily model use; model use is not necessarily meaningful contribution; and a correlation between use and output is not a causal result. Releasing the underlying table and recording how the sample was assembled makes critique possible and prevents visual polish from substituting for evidence.
Limits and uncertainty
The supplied material does not establish the atlas’s substantive chart results, the prevalence of disclosed AI use, or whether disclosures are increasing in any particular field or country. It also does not establish that AI improved, harmed, or caused changes in mathematical research productivity or quality. The analysis is limited to first versions of arXiv papers with math.* as their primary category, and its country coverage is partial. Because authors may use AI without disclosing it, the data cannot be read as a complete adoption measure. The site itself warns that first disclosure is not necessarily first use and that observed correlations are not causal effects.
Skill to build now
For the next internal AI-impact analysis, write a one-page measurement contract before making charts. Specify the population, time window, unit of analysis, what a positive signal actually means, what it cannot mean, and which records are missing. Then publish or retain the underlying table alongside the visualization. This week, apply that contract to one existing dashboard and replace any label that overstates a proxy as direct measurement.
Reader takeaway
The important contribution is not a single adoption number. It is a reusable standard for studying AI use: define the observable signal narrowly, preserve the snapshot, expose the data, distinguish categories of evidence, and state plainly which conclusions the dataset cannot support.
Open question
Will future versions preserve a comparable disclosure taxonomy as AI assistance becomes more common and less consistently acknowledged in research papers?