AI Math Slop - Math Papers on arXiv Jump 75% in a Year

From August 2025 to August 2026, the number of math papers posted on arXiv rose 75%, while the share of sampled abstracts classified as Likely AI saw a 68% increase over the same time period.

Mathematicians are concerned about this rise and what it will mean for the field. 

This study's findings arrive during ongoing disputes about the role AI should play within Mathematics. 

The Mathathon dispute

Mathathon is a planned AI-assisted Mathematics research competition at Caltech with OpenAI and Anthropic initially as sponsors.

The open letter

Current and former Caltech mathematicians called for the event to be suspended, warning of a flood of poorly checked results:

Read the open letter.

The backlash

Business Insider reported that OpenAI had withdrawn its sponsorship. 

Dan Roberts announced the decision on X:

The organizers’ response

The organizers acknowledged shortcomings in their initial plan and added an arXiv publication requirement. 

Read the organizers’ response.

The Navier–Stokes dispute

In another ongoing dispute OpenAI’s claimed Navier–Stokes solution prompted questions about credit and access to unpublished work. Mathematician Tristan Buckmaster raised concerns; OpenAI responded and public reaction has been strong.

Public reaction from Mehdi (@BetterCallMedhi). Opening excerpt; profanity blurred.

The Mathematics Paper Boom, Month by Month

Originality.ai counted new arXiv math papers by month and scanned 1000 abstracts from each month using our AI detector to try and understand the role AI is playing in the boom.

Key Findings

• 75% more math papers posted in August 2026 than in August 2025: 5,882 versus 3,366.
• 68% relative growth in the share of Likely AI abstracts: 24.2%, up from 14.4% a year earlier.

August 2026 had the highest monthly posting count in the full January 2020–August 2026 dataset.

Each point represents one month, starting in December 2022. These are arXiv preprints, not peer-reviewed journal publication counts.

Likely AI Abstracts

Nearly one in four sampled abstracts was classified as Likely AI in August 2026. In the pre-ChatGPT baseline, one of 3,400 was flagged.

August 2025: 144 of 1,000 abstracts flagged. August 2026: 242 of 1,000. The increase is 9.8 percentage points, or 68.1% relative growth.

Why This Matters: The Burden of Checking

Mathematicians are raising concerns about the work needed to check AI-assisted research and understand its results.

Tamara G. Kolda · SIAM News essay

Daniel Litt · original X post

Terence Tao · commentary on the Alpöge–Buckmaster work, before OpenAI’s announcement

What mathematicians are doing about it

Published in June 2026, before these disputes, the Leiden Declaration calls for disclosure of AI use and keeps responsibility for correctness with human authors.

The Leiden Declaration turns these concerns into a shared set of recommendations. Its website listed 3,889 signatories when checked on September 11, 2026. The International Mathematical Union has endorsed it, alongside public endorsements from mathematicians including Terence Tao and Peter Scholze.

The Mathathon organizers say they have all signed the declaration.

Its main recommendations summarized:

• Disclose AI use: explain which tools helped produce the research.
• Own the results: human authors remain responsible for correctness and citations.
• Make work easier to check: provide clear arguments, precise references and formal proofs where appropriate.
• Protect peer review: press releases and announcements cannot replace scientific evaluation.
• Credit human contributions: acknowledge earlier discoveries and respect authors’ rights.
• Protect research independence: commercial incentives should not determine which mathematical questions matter.

Signups peaked around the June launch, with a smaller pickup in September.

Source: Leiden Declaration public signatory dates. Chart includes complete days only; the partial final day is excluded.

Study Methodology

• Counted 259,158 unique papers first posted to arXiv’s primary mathematics categories from January 2020 through August 2026; compared August 2025 with August 2026.
• Randomly selected first-version English abstracts with at least 100 prose words: 100 per month through November 2022, then 1,000 per month.
• Scanned 48,500 abstracts with Originality.ai AI Allowance 15. Scores of 0.50 or higher were classified as Likely AI; 48,484 scans succeeded and 16 remained missing.
• Calculated each month’s flagged share using successful scans.

Data available upon request