Progress: AI Research
How fast AI research is moving, measured week by week in new papers on arXiv.
1,907 new AI papers in a week (+8% YoY)
What’s happening right now
Weekly counts
| Week | Papers |
|---|---|
| 2026-09-17 → 2026-09-23 | 1,907 |
| 2026-09-10 → 2026-09-16 | 2,037 |
| 2026-09-03 → 2026-09-09 | 1,986 |
| 2026-08-27 → 2026-09-02 | 2,282 |
| 2026-08-20 → 2026-08-26 | 2,010 |
| 2026-08-13 → 2026-08-19 | 1,833 |
| 2026-08-06 → 2026-08-12 | 2,098 |
| 2026-07-30 → 2026-08-05 | 2,357 |
The trend
New AI papers on arXiv per week
arXiv API search totals for cs.AI, cs.LG and cs.CL (cross-lists once), 7-day windows ending 3 days ago. Missing bars mean the API didn't answer for that week.
Background: why this matters
This is one of two factors that push the needle toward 'prospering'. The idea is simple: humanity's odds improve when we get better at solving problems, and AI has become one of the most general problem-solving tools we have. Deep learning helped predict the 3D structures of proteins: AlphaFold earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry. AI is being used to discover materials, forecast weather, read medical scans and write software.
The modern boom has a short history. In 2012 a neural network called AlexNet won a major image-recognition competition by a wide margin, kicking off the deep-learning era. In 2017 Google researchers published 'Attention Is All You Need', the paper that introduced the transformer architecture behind today's large language models. ChatGPT's release in November 2022 brought these systems to hundreds of millions of people. Stanford's AI Index documents the growth in AI research, investment and capability each year.
arXiv, founded in 1991 and run by Cornell University, is where most AI research appears first, often months before formal peer review. We count new submissions in its three main AI categories (Artificial Intelligence, Machine Learning, and Computation and Language) over a seven-day window. For the score, we compare that count with the same week a year earlier. A rising count signals more people and resources working on the field, and more ideas in circulation.
Paper counts are a rough proxy. They measure activity, not quality or breakthroughs, and some growth reflects the pressure to publish. The same progress also feeds the risks tracked on the AI Risk page. We give this factor a small weight (3%) because it's a signal of momentum and capability, not a direct measure of human welfare.
What people watch: the share of research that is openly published versus kept private; applications in science and medicine; compute and energy costs of training; and whether safety research grows as fast as capabilities research.
What would move the score
R = new cs.AI/cs.LG/cs.CL papers in the 7 days ending 3 days ago ÷ the same window one year earlier (arXiv API). Score = 100 at R ≥ 1.5, 0 at R ≤ 0.8, linear. Each +0.1 in the ratio is worth about 14 points.
Right now: 40/100 · cs.AI/LG/CL papers, 2026-09-17 → 2026-09-23: 1,907 · Same week a year earlier (2025-09-17 → 2025-09-23): 1,764
↓ Toward extinction
- AI research activity falling compared with a year earlier (fewer submissions)
- Holiday or conference-deadline timing can temporarily depress one week versus the year before
↑ Toward prospering
- Faster year-on-year growth in openly published AI research
- More researchers and institutions entering the field
Exact scoring rule and calibration
Window = the 7 days ending 3 days ago (to let arXiv's search index catch up). R = papers in window ÷ papers in the same window one year earlier, both from the arXiv API (submittedDate search). Score = 100 at R ≥ 1.5 (50% year-on-year growth), 0 at R ≤ 0.8 (a 20% contraction), linear between. Note: API search totals are lower than arXiv's listing pages, but the method is identical for both years, so the ratio is comparable.
Sources
Score data as of Sep 23, 2026, fetched Sep 26, 2026, 23:51 UTC. Deep-dive data fetched Sep 26, 2026, 23:51 UTC. Pages refresh every 15 minutes; each source is cached for 15 minutes (GDELT) to 24 hours (long histories) to respect free-API limits. Nothing is estimated or filled in by hand.