arXiv is where most AI research appears first. Palanor counts it each month and, with OpenAlex, shows who writes it.
arXiv counts
New papers each month in machine learning (cs.LG), computation and language (cs.CL), computer vision (cs.CV), artificial intelligence (cs.AI), robotics (cs.RO) and statistical machine learning (stat.ML), from January 2010. A paper counts in a category when it is listed there, as its main category or as a cross-list, and in the month its first version was submitted. The combined AI and machine-learning series counts each paper once, and the AI share divides it by arXiv's own count of all new submissions that month. Only complete months are published. Conference deadlines make some months jump, so compare a month with the same month a year earlier.
Who writes it (OpenAlex)
arXiv records don't say where authors work. OpenAlex links each arXiv paper to its authors' institutions, with each institution's type and country. For AI papers on arXiv (OpenAlex's Artificial Intelligence and Computer Vision subfields), Palanor publishes:
- the industry share: papers with at least one author at a company, out of papers where OpenAlex identified any institution;
- country shares for the United States, China, the United Kingdom, Germany, India, Canada, France, Japan and South Korea, on the same base (a paper with authors in two countries counts in both);
- papers per company each quarter for large AI publishers, counting a company and its research units. These appear on the company's entity page.
OpenAlex dates a paper known only by its year as 1 January, so Palanor's windows skip 1 January. OpenAlex keeps adding affiliations, so the latest months are rewritten for half a year.
AI-written text share
Palanor scores arXiv abstracts and charities' Form 990 mission and program text with the same open detector it uses for earnings calls (Oxidane/tmr-ai-text-detector, MIT licence, pinned revision). Each corpus has its own cutoff, fixed by a rule written down before any text was scored: the score that 5% of pre-ChatGPT text exceeds (abstracts from 2019 to 2021, returns filed from 2020 to September 2022). So those years read about 5% by construction, and a rise means more text that reads like model output. It is not a measure of how much text a model wrote: the detector learned from earlier models, so newer models' text can read as human to it, and a falling share can mean it recognises less rather than that less text is written with AI. Palanor never publishes a score for one paper or one organization.
The arXiv share is no longer published (21 Oct 2026). It rose from about 5% to 27.5% in 2025 Q1, then fell back to 5.8% by 2026 Q3. That fall is most likely detector drift, newer models' text reading as human to a detector trained on older ones, not less AI-written text. A series whose recent moves mostly measure the detector is not worth showing, so the arXiv series and its chart are private. The charity series is still rising and stays public.
Source
arXiv (descriptive metadata under CC0 1.0; thank you to arXiv for use of its open access interoperability) and OpenAlex (CC0). Only counts are stored.