The United States and China, five times over
9,660 co-authored AI publications in 2021 — five times the 2010 figure, and 2.7 times the next pairing on the list.
AI research doubled in a decade — and in 2021 China published the most of it
Chapter 1 of the AI Index 2022 maps where AI knowledge came from in 2021: journal articles, conference papers, preprint repositories, patents, conferences and open-source libraries. Two things stand out. Output kept compounding — total AI publications doubled between 2010 and 2021, and patent filings grew far faster than that. And the map of who produces it split in two: China leads on volume, the United States leads on the citations. The numbers:
334,497AI publications worldwide in 2021 (162,444 in 2010)
63.2% by which China’s journal, conference and repository papers combined exceeded the US in 2021
76.9% compound annual growth rate of AI patent filings, 2015–21
9,660US–China co-authored AI publications in 2021, five times the 2010 level
88,760attendees across 16 tracked AI conferences in 2021, level with 2020
160,700cumulative GitHub stars for TensorFlow in 2021, about three times OpenCV
1.1 — The journal quietly became AI’s main publishing venue
Total AI publications doubled from 162,444 in 2010 to 334,497 in 2021. But the mix shifted underneath that growth: journal articles took 51.5% of everything published in 2021, conference papers 21.5% and preprint repositories 17.0% — and conference papers have been falling since 2018.
Over the past 12 years journal publications grew about 2.5 times and repository publications about 30 times, while conference papers peaked in 2019 and finished 2021 roughly 19.4% below that peak. Books, book chapters, theses and unknown document types made up the remaining 10.1%. The share statistics tell the same story from the other side: AI reached 2.5% of all journal publications in 2021 (1.5% in 2010), 17.8% of all conference publications, and 15.3% of everything posted to preprint repositories.
What the field is actually working on
- Pattern recognition was the single largest field of study in 2021 with 51,690 publications, followed by machine learning with 39,930. Both have more than doubled since 2015.
- Computer vision (24,800), algorithms (18,300), data mining (15,270) and natural language processing (13,430) all grew more slowly, despite being the areas most visibly reshaped by deep learning.
- Human–computer interaction (9,700), control theory (8,450) and linguistics (5,780) trail the pack.
- The counts exclude a large residual category of other AI work, so they describe the shape of the field rather than its full size.
Universities still do most of the work
Worldwide, 59.6% of AI publications in 2021 came from the education sector, against 11.3% from nonprofits, 5.2% from companies and 3.2% from government; a further 20.8% could not be classified at all. Company participation is highest in the United States at 9.8%, then the European Union plus the United Kingdom at 5.7%, and lowest in China at 3.9%. China is also the only area where the education share is still rising — it reached 60.2% in 2021, above the United States at 57.6% and the EU plus UK at 54.8%.
China wins on volume, the United States wins on citations
In 2021 China led the world in journal, conference and repository publications combined — 63.2% ahead of the United States. But on the papers people actually cite, the ranking flips in two of those three venues.
Journals: China’s strongest suit
- China produced 31.0% of the world’s AI journal publications in 2021, ahead of the European Union plus the United Kingdom at 19.1% and the United States at 13.7%.
- China leads journal citations too, with 27.8%, ahead of the EU plus UK at 21.1% and the United States at 17.5%. Those three areas together account for more than 66% of all AI journal citations in the world.
- By region, East Asia and the Pacific took 42.9% of journal publications, Europe and Central Asia 22.7% and North America 15.6%. South Asia and the Middle East and North Africa grew fastest, around 12 and 7 times respectively over the 12 years.
Conferences and repositories: the ranking inverts
- China produced 27.6% of AI conference publications in 2021, ahead of the EU plus UK at 19.0% and the United States at 16.9% — but the United States took 29.5% of conference citations, against 23.3% for the EU plus UK and 15.3% for China.
- On preprint repositories the United States leads outright: 32.5% of publications and 38.6% of citations, against 23.9% and 20.1% for the EU plus UK, and 16.6% and 16.4% for China.
- By region, North America still holds 35.9% of repository publications, ahead of Europe and Central Asia at 27.3% and East Asia and the Pacific at 26.4% — the one publication type where North America has kept the lead since 2014.
- South Asia is the fastest-moving region at conferences: its share of AI conference publications rose from 4.0% in 2010 to 10.4% in 2021.
Geopolitics rose, and so did US–China co-authorship
CSET counts a cross-country collaboration as a distinct pair of countries across the authors of one publication — four US and four Chinese authors on a single paper count once. From 2010 to 2021 the biggest pairing by a distance was the United States and China, and the biggest cross-sector pairing was not company and university. Tap a card for the detail.
The rest of the world’s pairings
Once the US–China pair is set aside, the United Kingdom sits in the middle of everything: UK–China 3,560 and US–UK 3,340 in 2021.
Universities and nonprofits, not universities and companies
29,839 education–nonprofit collaborations in 2021 — 2.5 times as many as between education and companies.
The pairings that barely happen
Companies and governments co-produced 660 AI publications in 2021, the smallest pairing the chapter tracks.
Filings grew more than 30-fold in six years — and China was granted almost none of them
141,240 AI patents were filed in 2021, more than 30 times the 2015 figure, a compound annual growth rate of 76.9%. China filed 51.7% of them. China was granted 5.9% of the world’s AI patents.
Filings: East Asia took off in 2014
- East Asia and the Pacific accounted for 62.1% of all AI patent applications in 2021, ahead of North America at 17.1% and Europe and Central Asia at 4.2%.
- By geographic area, China filed 51.7% of the world’s AI patents in 2021, the United States 16.9% and the European Union plus the United Kingdom 3.9%.
- The United States files almost all of North America’s AI patents — and does so at one-third the rate of China.
Grants: the map reverses
- On granted patents North America leads with 57.0%, ahead of East Asia and the Pacific at 31.1% and Europe and Central Asia at 11.3%. Every other region combined accounts for roughly 1% of the world’s granted AI patents.
- By geographic area the United States holds 39.6% of grants, the EU plus UK 7.6% and China 5.9%.
- The raw counts make the gap concrete. In 2021 China filed 87,343 AI patent applications and was granted 1,407. The United States filed 19,610 and was granted 9,450; the EU plus UK filed 4,880 and was granted 1,810.
The chapter reports the two series side by side without reconciling them, and it is worth keeping them apart when reading any AI-patent league table. Counted by filings, China holds more than half the world. Counted by grants, it holds less than the European Union plus the United Kingdom. The same underlying dataset supports two opposite headlines.
Five things the chapter answers about 2021
Conferences, the Women in Machine Learning workshop, open-source libraries, and what the publication numbers do and do not prove.
Did a second virtual year change who shows up?
Attendance at top AI conferences in 2021 was relatively consistent with 2020, at more than 88,000 participants across the 16 conferences the Index tracks — more conferences than any previous edition covered. Almost all were virtual; only ICRA and EMNLP ran in a hybrid format. Organizers told the AI Index that measuring exact attendance at a virtual conference is difficult, and that virtual formats allow far higher attendance from researchers around the world. The individual counts come with footnotes: ICML used session visitors as a proxy for attendees, which explains its 29,540; IROS let users watch events for up to three months in 2020; and AAMAS reported on-site users in 2020 but total registrants in 2021. Among the smaller conferences, UAI drew 2,100, IJCAI 1,900, FAccT 1,350, AAMAS 1,080 and ICRA 1,000.
How large is the Women in Machine Learning workshop?
The 2021 WiML workshop colocated with NeurIPS drew an estimated 1,486 attendees, counted as the number of unique individuals who accessed the virtual platform at neurips.cc. The figure has risen steadily since the workshop was first held in 2006. It ran as multiple sessions over three days, a change of format from 2020, and was again held virtually because of the pandemic. Among survey respondents who consented to have their information aggregated, 53.4% lived in North America, 19.9% in Europe, 16.2% in Asia and 7.3% in Africa. PhD students made up 49.4%, MSc students 16.5%, research scientists and engineers 14.2%, and undergraduates 10.2%, while university faculty accounted for around 1.2%.
Which open-source AI libraries do developers actually use?
Measured by cumulative GitHub stars in 2021, TensorFlow is still far ahead with around 160,700 — about three times the next library, OpenCV at 58,600, and only a slight increase over its own 2020 figure. Keras (53,200), PyTorch (52,700) and Scikit-learn (48,000) sit close together behind it, then DeepLearning-500-questions (46,700) and TensorFlow-Examples (41,500). Below 40,000 stars the list is not really about libraries at all: faceswap leads with 39,880, followed by 100-Days-Of-ML-Code (33,580), AiLearning (32,270), BVLC/caffe (32,140), Real-Time-Voice-Cloning (32,020), deeplearningbook-chinese (32,000), Deep Learning Papers Reading Roadmap (31,280) and DeepFaceLab (30,260) — mostly tutorials, reading lists and face-manipulation projects.
Does leading on publications mean leading on AI?
The chapter’s own numbers argue against reading it that way. China led the world in 2021 on journal, conference and repository publications combined, 63.2% ahead of the United States, and led journal citations with 27.8%. But the United States held a dominant lead in conference citations at 29.5% against China’s 15.3%, and in repository citations at 38.6% against 16.4% — and it produced 32.5% of the world’s repository publications against China’s 16.6%. Volume and influence point in different directions depending on which venue you look at, and the chapter reports both rather than resolving them.
How reliable are these publication counts?
The chapter is explicit about its limits. Publication data comes from CSET’s merged corpus of scholarly literature — Dimensions, Web of Science, Microsoft Academic Graph, China National Knowledge Infrastructure, arXiv and Papers with Code — with a classifier applied to identify English-language publications related to the development or application of AI and ML since 2010. Non-English work is therefore out of scope. The 2021 count may also be lower than the true figure because of lag in how those databases collect publication metadata. And because this edition changed both data provider and classification method, the trend is not directly comparable with earlier AI Index reports. Even inside the data, 20.8% of 2021 publications could not be assigned to a sector at all.
The chapter in four lines
Headline findings from Chapter 1 · Research and Development.
Despite rising geopolitical tensions, the United States and China had the greatest number of cross-country collaborations in AI publications from 2010 to 2021, increasing five times since 2010. The collaboration between the two countries produced 2.7 times more publications than between the United Kingdom and China — the second highest on the list.
In 2021, China continued to lead the world in the number of AI journal, conference, and repository publications — 63.2% higher than the United States with all three publication types combined. In the meantime, the United States held a dominant lead among major AI powers in the number of AI conference and repository citations.
From 2010 to 2021, the collaboration between educational and nonprofit organizations produced the highest number of AI publications, followed by the collaboration between private companies and educational institutions and between educational and government institutions.
The number of AI patents filed in 2021 is more than 30 times higher than in 2015, showing a compound annual growth rate of 76.9%.
Read Chapter 1 in full
Chapter 1 (sections 1.1–1.3) — publications, patents, conferences and open-source software libraries — with every figure, footnote and citation, is free from Stanford HAI.
Open the AI Index 2022 →