china ai research papers

Is China Emerging as the Global Leader in AI?

by Daitian Li , Tony W. Tong and Yangao Xiao

Summary .   

China is quickly closing the once formidable lead the U.S. maintained on AI research. Chinese researchers now publish more papers on AI and secure more patents than U.S. researchers do. The country seems poised to become a leader in AI-empowered businesses, such as speech and image recognition applications. But while China has caught up with impressive speed, the conditions that have allowed it to do so — the open science nature of AI and the nature of the Chinese market, for instance — will likely also prevent it from taking a meaningful lead and leaving the U.S. in the dust.

Twenty years ago, there was a huge gulf between China and the United States on AI research. While the U.S. was witnessing sustained growth in research efforts by both public institutions and private sectors, China was still conducting low-value-added activities in global manufacturing. But in the intervening years, China has surged to rapidly catch up. From a research perspective, China has become a world leader in AI publications and patents. This trend suggests that China is also poised to become a leader in AI-empowered businesses, such as speech and image recognition applications.

Partner Center

The Promise and Perils of China's Regulation of Artificial Intelligence

Columbia Journal of Transnational Law (forthcoming)

40 Pages Posted: 12 Feb 2024 Last revised: 2 Sep 2024

Angela Huyue Zhang

University of Southern California Gould School of Law

Date Written: January 28, 2024

In recent years, China has emerged as a pioneer in formulating some of the world’s earliest and most comprehensive rules concerning algorithms, deepfakes, and generative artificial intelligence (AI) services. This proactive intervention has left the impression that China has stood at the forefront as a global leader in regulating AI. Yet this perception put too much emphasis on the law on paper while overlooking the country’s intricate institutional dynamics. The Chinese government simultaneously acts as a policymaker, an investor, a supplier, a customer and a regulator in the AI sector. Given its extensive and deep involvement in the AI ecosystem, the government lacks a strong commitment to regulate the industry. Factors such as the intense US-China tech rivalry and the escalating chip embargo on Chinese AI firms further diminish the government’s incentive to impose strict regulation. Meanwhile, the current downturn in the Chinese economy and low market confidence impose further constraints on the government’s actions. Consequently, despite maintaining strict information control over public-facing AI services, China’s overall approach to AI regulation has been markedly business-friendly. Recent legislative measures, such as the interim measures to regulate generative AI and several local AI legislations, offer little protective value to the Chinese public. Instead, these laws have primarily served as an enabler by sending a strong pro-growth signal to the industry while attempting to coordinate various stakeholders to accelerate technological progress. As evidenced by its permissive stance over the abusive use of facial recognition technology, Chinese regulators have favoured a light-touch approach to AI regulation in practice. Similarly, Chinese courts are trying to prop up the AI industry, as demonstrated by the Beijing Internet Court’s decision to grant copyrights in an AI-generated image. China’s strategic lenient approach to regulation may therefore offer its AI firms a short-term competitive advantage over their European and U.S. counterparts. However, this leniency risks creating potential regulatory lags that could escalate into AI-induced accidents and even disasters. The dynamic complexity of China’s regulatory tactics therefore underscores the urgent need for increased international dialogue and collaboration with the country to tackle the safety challenges in AI governance.

Keywords: artificial intelligence, AI, generative AI, China, regulation, governance, risks, facial recognition, copyright, AI safety, chips, semiconductors, tech war

JEL Classification: k23

Suggested Citation: Suggested Citation

Angela Huyue Zhang (Contact Author)

University of southern california gould school of law ( email ).

699 Exposition Blvd. Los Angeles, CA 90089 United States

Do you have a job opening that you would like to promote on SSRN?

Paper statistics, related ejournals, university of hong kong faculty of law legal studies research paper series.

Subscribe to this free journal for more curated articles on this topic

Law & Society: International & Comparative Law eJournal

Subscribe to this fee journal for more curated articles on this topic

Law & Society: The Legal Profession eJournal

Asian law ejournal, artificial intelligence ejournal, artificial intelligence - law, policy, & ethics ejournal, chinese law ejournal, international political economy: globalization ejournal, electronic, optical & magnetic materials ejournal, materials science education ejournal.

Featured Topics

Featured series.

A series of random questions answered by Harvard experts.

Explore the Gazette

Read the latest.

For this ring, I thee sue 

Speech is never totally free

EVs fight warming but are costly. So why aren’t we driving $10,000 Chinese imports?

In a recent U.S. government ranking of companies producing the most accurate facial recognition technology, the top five were all Chinese companies. China is currently installing facial recognition equipment across entrances to many of its subway stations.

Imaginechina via AP Images

Why China has edge on AI, what ancient emperors tell us about Xi Jinping

Christy DeSmith

Harvard Staff Writer

Recent event examines what social sciences can tell us about rising economic, geopolitical power

Dictatorships and authoritarian regimes tend to trail more democratic and inclusive nations in fostering cutting-edge, innovative technologies, such as robotics and clean energy.

Artificial intelligence may prove an exception, at least in China, owing to dovetailing interests.

Harvard Economics Professor David Yang  spoke to the outsized success of China’s AI sector at a recent dean’s symposium on insights gleaned from the social sciences about the ascendant global power. As evidence, he cited a recent U.S. government ranking of companies producing the most accurate facial recognition technology . The top five were all Chinese companies.

“Autocratic governments would like to be able to predict the whereabouts, thoughts, and behaviors of citizens,” Yang said. “And AI is fundamentally a technology for prediction.” This creates an alignment of purpose between AI technology and autocratic rulers, he argued.

Because AI heavily depends on data, and autocratic regimes are known to collect vast troves of it, this advantages companies with Chinese government contracts, which can turn around and use state data to bolster commercial projects, he added.

Yang’s research shows China exporting huge amounts of AI technology, dwarfing its contributions in other frontier technology sectors. Yang also demonstrated that autocratic regimes around the world have a particular interest in AI. “AI quite startlingly is the only sector out of the 16 frontier technologies where there’s disproportionately more buyers that are weak democracies and autocracies.”

“Autocratic governments would like to be able to predict the whereabouts, thoughts, and behaviors of citizens. And AI is fundamentally a technology for prediction.” David Yang, associate professor of economics

And just when are these countries most likely to buy the technology from China? Yang ended his symposium talk by mapping the uptick in purchases that follow political unrest and protest events. “To the extent that technology is exported,” Yang concluded, “it could generate a spreading of similar autocratic regimes to the rest of the world.”

Hosting Yang’s presentation was Lawrence D. Bobo, Dean of Social Science and W.E.B. Du Bois Professor of the Social Sciences. Launched in 2021, these virtual symposia gather scholars from across the division to trade research and thinking on topics of broad interest. “China in Focus: New Social Science Approaches,” which was held earlier this month, was moderated by Mark C. Elliott, the Mark Schwartz Professor of Chinese and Inner Asian History and vice provost for international affairs.

More bold predictions came from Professor of Government Yuhua Wang , whose current research relies not on contemporary economic data, but ancient indicators.

Drawing from his recent book “ The Rise and Fall of Imperial China: The Social Origins of State Development ,” Wang shared a chart of emperor assassinations across 2,000 years of Imperial China. Gathering this data meant analyzing the biographies of nearly 400 Chinese emperors, from the Qin Dynasty to the Qing Dynasty. Turns out, about a quarter were assassinated by members of their own government and most likely during economically strong governments, hitting their peak circa 900 A.D. during the late Tang Dynasty.

“Why do we see this contradiction between the strength of the ruler and the strength of the government?” Wang asked. “Chinese rulers — historically but also contemporarily — face a tradeoff that I call the Sovereign’s Dilemma.” That is, a coherent set of government elites is capable of strengthening the state but equally capable of overthrowing the boss.

On the other hand, fragmented elites spell longevity for rulers and decline for states. This is the very dynamic Wang sees playing out today under Chinese President Xi Jinping, whose anti-corruption campaign threatens government insiders with investigation and arrest.

As evidence of splintering elites, Wang cited the sudden pivot from China’s zero-COVID policy and the recent appearance of spy balloons in U.S. airspace. “It’s very clear the people sending balloons, maybe in the military, were not talking to the Foreign Ministry who were about to welcome [U.S. Secretary of State Antony] Blinken” for an official visit, Wang said.

“What happens is probably a very dramatic but also gradual decline of the capacity of the Chinese state.”

Also featured at the two-hour symposium was Victor Seow, assistant professor of the history of science, who covered 100 years of intensive energy extraction under multiple regimes in the country’s northeast. Ya-Wen Lei, associate professor of sociology , unpacked the human costs of China’s speedy transition from labor-intensive manufacturing to a science- and technology-driven economy.

Professors like these put the Division of Social Science in a strong position, Bobo noted at the end. “Harvard will be at the forefront of China scholarship for years to come.”

Share this article

You might like.

Unhappy suitor wants $70,000 engagement gift back. Now court must decide whether 

Cass Sunstein suggests universities look to First Amendment as they struggle to craft rules in wake of disruptive protests

Experts say tension between trade, green-tech policies hampers climate change advances; more targeted response needed

Harvard releases race data for Class of 2028

Cohort is first to be impacted by Supreme Court’s admissions ruling

Parkinson’s may take a ‘gut-first’ path

Damage to upper GI lining linked to future risk of Parkinson’s disease, says new study

Professor tailored AI tutor to physics course. Engagement doubled.

Preliminary findings inspire other large Harvard classes to test approach this fall

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• NATURE INDEX
• 09 August 2023

What China’s leading position in natural sciences means for global research

• Chris Woolston 0

Chris Woolston is a freelance journalist in Billings, Montana.

You can also search for this author in PubMed   Google Scholar

China is pursuing areas of green innovation, such as this smart greenhouse in Yongchuan that automates key processes in rice-plant cultivation. Credit: Si Chuan/Feature China/Future Publishing via Getty Images

Following an upward trajectory of scientific productivity that has been gathering steam for decades, China has reached a new milestone. In 2022, for the first time, the country had the highest Share score in the Nature Index for the natural sciences, surpassing the United States.

This has been on the cards since Nature Index was launched in 2014. China’s Share, which measures a country or institution’s contribution to publications in 82 high-quality natural-science journals, has rapidly grown since.

Nature Index 2023 China

Back in 2015, its adjusted Share , which takes account of yearly variations in article volume, was 8,430, almost one-third of the United States’ score. But year-on-year rises of between 8% and 21% (apart from 2020, the first year of the pandemic, when growth stalled) have seen it reach the top spot.

This is not the first time that China has been assessed as the leader on measures of scientific productivity. In 2017, it overtook the United States for the total number of scientific publications, according to the US National Science Foundation (NSF). And in 2022, Japan’s National Institute of Science and Technology Policy reported that China had surpassed the United States in a key metric that aims to estimate performance in high-quality science: the contribution to papers that rank in the top 1% most-cited publications.

Such results are significant for Chinese universities and policymakers, says Fei Shu, who studies bibliometrics and researcher assessment at the University of Montreal in Canada and at Hangzhou Dianzi University in China. “China is really rankings driven,” he says.

World-class science

The shift in country rankings in the Nature Index seemed inevitable given the data trend of the past few years, but the achievement is still noteworthy, says Caroline Wagner, a science policy and innovation researcher at Ohio State University in Columbus. “The Chinese have done something truly astonishing” in building a world-class science system in just four and a half decades, she says. China announced its ‘reform and opening-up’ policy in 1978, laying foundations for its higher-education and scientific system.

Hamish Coates, a higher-education researcher at Tsinghua University in Beijing, says the past seven years of China’s journey towards becoming a scientific superpower have highlighted the “strength of its innovation ecosystem”.

China has developed a reputation for relying on imitation to produce such large numbers of papers, says Wagner, but contrary to this common belief, the papers coming out of the country often show high levels of innovation. Wagner co-authored a 2020 study 1 that tracked the percentage of papers that included references to journals in other disciplines, a sign of more creative research that is attempting to cross disciplinary boundaries. The analysis found that papers with at least one China-based co-author were more likely than others to stretch these boundaries. “Not only were they doing good quality work, they were also doing novel work,” says Wagner.

Chinese research remains under-valued, says Coates. In Western universities, “there are plenty of people who have had passing or superficial engagements with higher education in China, or Asia more generally, and have yet to grasp the transformations in play”.

Coates’s home institution — which he says benefits from abundant resources, a concentration of talent and encouragement from academic leadership to publish in top academic journals — is one of many Chinese universities driving the surge in publications. The adjusted Share for Tsinghua University increased by 35.5% from 2020 to 2022, putting it in the top 10 rising Chinese institutions in that period. Others saw even bigger gains in adjusted Share. Sun Yat-sen University in Guangzhou, for instance, went up 52.4% and Shandong University in Jinan almost doubled (up 97.8%).

Immense resources, immense effort

A factor in China’s rising research productivity is its universities adopting a working culture similar to Chinese industry, says Miguel Lim, an education and international development researcher at the University of Manchester, UK. “They work very long hours and there’s a pressure to produce and a pressure to succeed,” says Lim. He adds that many researchers elsewhere work extremely hard for long hours, but that approach isn’t as widespread as it is in China.

The resources behind Chinese science are also immense. The NSF reports that China and the United States accounted for roughly half of global research and development investment in 2019, with the United States spending US$656 billion and China’s outlay being worth $526 billion. According to the National Bureau of Statistics of China, its spending on research and development reached 2.4% of its GDP in 2021, an all-time high. By comparison, China invested just 1.2% of its GDP on research and development in 2004.

Source: Nature Index

The United States’ expenditure on research and development in 2020 was equivalent to 3.4% of GDP. Much of that spending goes towards basic science and preliminary research that might or might not lead to new technologies or therapies. Although China still spends less than the United States on research and development, Lim says the money is focused on results. “There’s a whole state approach that is able to identify certain areas of national interest,” he says. “They’ve identified scientific, technological and engineering areas and concentrated their efforts. It’s not necessarily about blue-sky thinking. It’s a problem-solving kind of approach.”

China’s 14th Five-Year Plan, which sets specific development goals for the period 2021–25 and describes the country’s longer-term vision for 2035, puts a strong emphasis on technological innovation, highlighting recent successes in lunar exploration, supercomputing, quantum information and high-speed trains. It also calls for the creation of national laboratories to focus on network communications, modern energy systems, pharmaceuticals, biotechnology and artificial intelligence (AI) , among other fields.

China is already one of the world’s leading research nations in AI. Stanford University’s Artificial Intelligence Index Report 2023 found that China accounted for nearly 40% of all publications in AI in 2021, far exceeding the United Kingdom and Europe (15%) and the United States (10%). Papers from China accounted for 29% of all AI citations in 2021, which again puts it ahead of the United Kingdom and Europe (21.5%) and the United States (15%). China ranked second to the United States in a 2022 assessment of AI and robotics articles in the Nature Index, but its annual Share rose more than 1,100% between 2015 and 2021, significantly outpacing the United States, United Kingdom, France and Germany.

Environmental research, such as projects tackling green energy and pollution , have also rapidly progressed in China (see ‘A clearer view of progress’).

A clearer view of progress

China’s shift towards addressing urgent environmental challenges, such as air and water pollution, green-energy transition and biodiversity loss, has been a major win for researchers. In 2022, China surpassed the United States as the leading nation in the Earth and environmental sciences (E&E) in the Nature Index, owing, in no small part, to the funding and resources the country has poured into fields including the atmospheric sciences, geology and materials science.

Its efforts are paying off. Since China declared a ‘war on pollution’ in 2014, air quality in cities has steadily improved, thanks to restrictions on industrial emissions and other strategies. Upgrades to coal-fired power plants — retrofitting smokestacks with filters, for example — have had the biggest effect, according to atmospheric scientist Qiang Zhang, from Tsinghua University in Beijing. A 2019 study by Zhang and his colleagues analysed the main drivers of a recent decline in fine particulate matter in China, and is among the top-cited papers with Chinese authors in the Nature Index for that year ( Q. Zhang et al. Proc. Natl Acad. Sci. USA 116 , 24463–24469; 2019 ).

When it comes to China’s dominance in E&E research, increased funding is only part of the story. Greater numbers of Chinese scientists returning from training abroad have bolstered the country. In the Nature Index, China’s percentage increase in E&E between 2015 and 2022 was the highest among its rise in the four natural-sciences subject areas covered by the database. The country’s Share in E&E in 2022 (2,612) was more than six times that of the United Kingdom (393), which is ranked third in the Index in the subject after the United States, whose E&E Share was 2,352.

There’s still a long way to go. Water availability in Beijing is estimated to be ten times lower today than it was in 1949, and air pollution spikes were reported in 13 northern Chinese cities surrounding Beijing in March, highlighting the importance of the country’s continued investment in E&E research.

Shifting priorities

China’s current lead in some scientific publishing indicators is not guaranteed to continue, however. In February 2020, the Chinese Ministry of Education announced reforms in its researcher-evaluation system that could alter the publishing landscape. According to the new guidelines, researchers would no longer be evaluated for hiring and promotion decisions on the sheer number of papers they had contributed to. Instead, they would be judged on a limited selection of “representative” articles, preferably including papers published in journals with international influence. At least one-third of the representative papers must be published in Chinese journals . Coates says such policy changes could diminish the incentives for publishing large numbers of papers, potentially slowing the stream of publications from the country.

In a 2022 paper 2 , Shu and his co-authors questioned the real-world implications of the publishing reforms. They note that Chinese researchers had been bristling about the pressure and high expectations of the previous evaluation system, but the authors remain sceptical that the new guidelines will truly change the way in which hiring and promotion decisions are made at institutions.

Shu says that academic employment at Chinese universities remains closely tied to publications, putting intense pressure on researchers, despite China announcing in 2020 that it had banned the practice of scientists being offered cash rewards based on their publishing record. “Salary is based on publications, and you need a strong publication list to be promoted,” he says. Shu notes that many studies have compared the productivity of scholars in China with scholars in the United States or Europe, “but the comparison is unfair because they work in different environments”.

A cause for alarm?

At a time when some politicians in the United States, European Union and elsewhere are sounding alarms about China’s economic, military and industrial might, some may see the country’s ranking in the Nature Index as another cause for consternation. Wagner, who once advised the US government on science policy in her role as deputy to the director of the Science and Technology Policy Institute, a federally funded research and development centre located in Washington DC, says such rankings could add urgency to calls for greater investment in science in the West. As she explains, rankings lend themselves to a “horse race” mentality that disregards the nuances of international research and collaboration. “Legislators who don’t really understand science might say we need to spend more because we’re falling behind,” she says.

But Wagner stresses that China’s rise in scientific publishing shouldn’t cause panic in the West. For one thing, she says, China is still far behind much of the world in terms of scientific infrastructure, complex research networks and social support for innovation. “We would have to say that the United States, for example, is still vastly far ahead of China in terms of deep scientific roots and the ability to soak up new knowledge,” she says.

Any simplistic measure that compares one country to another also misses the bigger picture of interdependency and collaboration in global science, Wagner says. Here, she emphasizes how Chinese researchers are frequent collaborators on international studies. There are signs, though, that such partnerships are becoming less common. As reported in Nature 3 , the number of papers with authors from both the United States and China fell for the first time in 2021.

Wagner notes that since 2000, more than 5 million Chinese scholars and students have left China to work and study abroad. However, research conducted for Nature Index in 2021 by League of Scholars, an academic data and recruitment firm in Sydney, Australia, found the proportion of academics at Chinese universities who had arrived from abroad in the previous three years was almost triple the global average, suggesting many Chinese researchers are now returning. There are also indications of a slowdown in the number of students seeking to move abroad in the first place.

These changing patterns of interdependency may cause a more lasting effect than China reaching parity with the United States, or even overtaking it, on publication metrics. Denis Simon, who studies Chinese science and innovation at the University of North Carolina at Chapel Hill, warns that “the data should give us cause for concern”. “Not that China is catching up”, he says, but that the West might be missing out on “the expertise that could create very positive synergies in terms of addressing the world’s key problems in science and technology. It is the net addition of Chinese brain power to the solution of these problems that offers the world the greater hope.”

Nature 620 , S2-S5 (2023)

doi: https://doi.org/10.1038/d41586-023-02159-7

This article is part of Nature Index 2023 China , an editorially independent supplement. Advertisers have no influence over the content.

Wagner, C. S., Cai, X. & Mukherjee, S. Scientometrics 124 , 2457–2468 (2020).

Article   Google Scholar  

Shu, F., Liu, S. & Larivière, V. Minerva 60 , 329–347 (2022).

Article   PubMed   Google Scholar  

Van Noorden, R. Nature 606 , 235–236 (2022).

Download references

Related Articles

Partner content: Turbocharging China’s Green Transition

Partner content: China’s Rising Star of medical innovation

Partner content: The hub of innovation helping China’s young scientists

Partner content: Zhengzhou University seeks global talent

• Environmental sciences

The human costs of the research-assessment culture

Career Feature 09 SEP 24

Publishing nightmare: a researcher’s quest to keep his own work from being plagiarized

News 04 SEP 24

Intellectual property and data privacy: the hidden risks of AI

Career Guide 04 SEP 24

The grassroots organizations continuing the fight for Ukrainian science

Career Feature 11 SEP 24

How a struggling biotech company became a university ‘spin-in’

Career Q&A 10 SEP 24

Japan moves to halt long-term postgraduate decline by tripling number of PhD graduates

Nature Index 29 AUG 24

Extreme solar storms and the quest for exact dating with radiocarbon

Perspective 11 SEP 24

Wildfires are spreading fast in Canada — we must strengthen forests for the future

Comment 09 SEP 24

Simple steps could shrink US beef industry’s carbon hoofprint

Research Highlight 05 SEP 24

Senior Scientific Officer

The Senior Scientific Officer manage HFSP awards, contribute to scientific outreach, and support program development for life sciences research.

Strasbourg-Ville, Bas-Rhin (FR)

HUMAN FRONTIER SCIENCE PROGRAM ORGANIZATION

Sydney Horizon Educators (Faculty of Science)

An opportunity for the best and brightest talented educators to be part of our future at the Faculty of Science.

Sydney, New South Wales (AU)

University of Sydney

Sydney Horizon Educators (Faculty of Engineering)

An opportunity for the best and brightest talented educators to be part of our future at the Faculty of Engineering.

Sydney, New South Wales

Sydney Horizon Educators/Teaching and Research (Faculty of Medicine and Health)

An opportunity for the best and brightest talented educators and researchers to be part of our future at the Faculty of Medicine and Health

Sydney (Region), New South Wales (AU)

Faculty Positions at SUSTech School of Medicine

SUSTech School of Medicine offers equal opportunities and welcome applicants from the world with all ethnic backgrounds.

Shenzhen, Guangdong, China

Southern University of Science and Technology, School of Medicine

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

• Account Details
• Newsletters
• Group Subscription

China trounces U.S. in AI research output and quality

Tencent, Alibaba and Huawei among the top 10 companies

TOKYO/BEIJING -- China is the undisputed champion in artificial intelligence research papers, a Nikkei study shows, far surpassing the U.S. in both quantity and quality.

Tencent Holdings, Alibaba Group Holding and Huawei Technologies are among the top 10 companies producing AI research, according to the study. The Chinese contingent is steadily gaining representation in an area dominated by U.S. players.

Japan, Europe lead in global hydrogen patent race: report

Alibaba, huawei race to gain edge in southeast asia's cloud market, just how over is china's tech crackdown it depends who you ask, china tops u.s. to take research crown at global chip conference, japan bets on telecom ace to counter china in global 6g standards, byd takes commanding ev patent lead among chinese rivals, china closes in on japan's hydrogen technology patent lead, u.s.'s frontier retains throne as world's fastest supercomputer, china and u.s. soar ahead in flying car patent race, latest on china tech, china's geely builds out satellite network with latest launch, alibaba accelerates u.s. push with ai sourcing tool, china's dji taps young farmers to cultivate drone market, sponsored content, about sponsored content this content was commissioned by nikkei's global business bureau..

Nikkei Asian Review, now known as Nikkei Asia, will be the voice of the Asian Century.

Celebrate our next chapter Free access for everyone - Sep. 30

To revisit this article, visit My Profile, then View saved stories .

• The Big Story
• Newsletters
• Steven Levy's Plaintext Column
• WIRED Classics from the Archive
• WIRED Insider
• WIRED Consulting

China Is Catching Up to the US in AI Research—Fast

Machine learning

Machine vision

At the world’s top computer-vision conference last June, Google and Apple sponsored an academic contest that challenged algorithms to make sense of images from twin cameras collected under varied conditions, such as sunny and poor weather. Artificial intelligence software proficient at that task could help the US tech giants with money-making projects such as autonomous cars or augmented reality. But the winner was an institution with very different interests and allegiances: China’s National University of Defense Technology, a top military academy of the People’s Liberation Army.

That anecdote helps illustrate China’s broad ambitions in AI and recent prominence on the field’s frontiers. In 2017 the country’s government announced a new artificial intelligence strategy that aims to rival the US in the crucial technology by 2020. The latest data on the output of US and Chinese AI researchers suggest China is on track.

Chinese researchers have published more AI research papers than the US for several years, but questions have lingered about the quality and influence of those publications. A new analysis by the Allen Institute for AI shows that China’s share of top AI publications is rapidly approaching that of the US. If current trends continue, the two nations will produce an equal share of top AI publications by 2020.

The Allen Institute analyzed data on more than 2 million AI research publications through the end of 2018 from its Semantic Scholar academic search engine. Comparing US and Chinese AI publications makes it clear that China was an emerging powerhouse of AI research well before the recent national strategy was launched. The country has published more AI papers than the US since 2005, according to Semantic Scholar data.

That trend has long been noted, including by a report on US competitiveness in AI research commissioned by the Obama White House. It has also met with skepticism, because Chinese research institutions have a reputation for low quality and even fraudulent publications.

Yet when the Allen Institute repeated the analysis to include only the research papers cited most often by other scholars, the US did not emerge very far ahead. Extrapolating from data through the end of 2018 suggests China will match the US in its share of the top 10 percent of AI research papers in 2020—the same year China’s government says it wants to draw level with America’s AI prowess.

Citation counts don’t perfectly reflect the quality and influence of ideas, and the Allen Institute is planning more analysis to check to what extent the effect is explained by Chinese authors being more likely to cite fellow nationals. Still, Oren Etzioni, the organization’s CEO, says the findings suggest the US government needs to better support AI research. President Trump recently signed an executive order asking government agencies to do more in AI, but many in the field are skeptical it will be very effective. “It was well intentioned but low on specifics, and it didn’t deliver the two most important things that we need,” Etzioni says— a more welcoming immigration policy to draw top research talent and significantly more research funding.

Greg Allen, an adjunct senior fellow at think tank the Center for a New American Security, says the Allen Institute analysis should drive home the message that China’s AI ambitions are serious. Lofty bureaucratic strategies and targets like those detailed in China’s AI plan can seem curious when viewed from the US, but they can be effective. “That’s what happens when you call something a national priority and you mean it,” Allen says.

Allen recently published a report on how China’s military and national security apparatus are central to the country’s evolving AI strategy—as evidenced by the military university winning the contest sponsored by Apple and Google. He found that the country’s defense ministry is investing deeply in new AI research, for example, by setting up two new research centers in Beijing dedicated to AI and unmanned systems. A paper released by one of them in December tried to explain the inner workings of Alphabet’s AlphaZero system that is capable of superhuman performance in both chess and Go .

Data from the Stanford-affiliated AI Index , which tracks the trajectory of AI development using dozens of measures, shows how China’s government is growing its already central role in the country’s research. Government-affiliated AI research papers increased 400 percent between 2007 and 2017, dwarfing the growth from Chinese corporate labs, although China’s state-funded academic institutions still produce most of the country’s research output.

In the US, by contrast, companies such as Alphabet play a much more significant role. The share of AI publications that come from corporations is seven times higher in the US than in China.

Updated, 3-14-19, 2:05 pm ET: Two of the charts in this story have been corrected using revised data from the Allen Institute for Artificial Intelligence.

• Ways to get text onto your screen— without a keyboard
• 23 amazing female photographers you should know
• Inside the “black box” of a neural network
• Quantum physics could (maybe) save the grid from hacks
• Will AI achieve consciousness? Wrong question
• 👀 Looking for the latest gadgets? Check out our latest buying guides and best deals all year round
• 📩 Want more? Sign up for our daily newsletter and never miss our latest and greatest stories

Advertisement

Supported by

China’s Rush to Dominate A.I. Comes With a Twist: It Depends on U.S. Technology

China’s tech firms were caught off guard by breakthroughs in generative artificial intelligence. Beijing’s regulations and a sagging economy aren’t helping.

• Share full article

By Paul Mozur John Liu and Cade Metz

The reporters interviewed more than a dozen A.I. experts about China’s competitiveness in the critical field.

In November, a year after ChatGPT’s release, a relatively unknown Chinese start-up leaped to the top of a leaderboard that judged the abilities of open-source artificial intelligence systems.

The Chinese firm, 01.AI, was only eight months old but had deep-pocketed backers and a $1 billion valuation and was founded by a well-known investor and technologist, Kai-Fu Lee. In interviews, Mr. Lee presented his A.I. system as an alternative to options like Meta’s generative A.I. model , called LLaMA.

There was just one twist: Some of the technology in 01.AI’s system came from LLaMA. Mr. Lee’s start-up then built on Meta’s technology, training its system with new data to make it more powerful.

The situation is emblematic of a reality that many in China openly admit. Even as the country races to build generative A.I., Chinese companies are relying almost entirely on underlying systems from the United States. China now lags the United States in generative A.I. by at least a year and may be falling further behind, according to more than a dozen tech industry insiders and leading engineers, setting the stage for a new phase in the cutthroat technological competition between the two nations that some have likened to a cold war .

“Chinese companies are under tremendous pressure to keep abreast of U.S. innovations,” said Chris Nicholson, an investor with the venture capital firm Page One Ventures who focuses on A.I. technologies. The release of ChatGPT was “yet another Sputnik moment that China felt it had to respond to.”

Jenny Xiao, a partner at Leonis Capital, an investment firm that focuses on A.I.-powered companies, said the A.I. models that Chinese companies build from scratch “aren’t very good,” leading to many Chinese firms often using “fine-tuned versions of Western models.” She estimated China was two to three years behind the United States in generative A.I. developments.

We are having trouble retrieving the article content.

Please enable JavaScript in your browser settings.

Thank you for your patience while we verify access. If you are in Reader mode please exit and  log into  your Times account, or  subscribe  for all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?  Log in .

Want all of The Times?  Subscribe .

The next frontier for AI in China could add $600 billion to its economy

In the past decade, China has built a solid foundation to support its AI economy and made significant contributions to AI globally. Stanford University’s AI Index, which assesses AI advancements worldwide across various metrics in research, development, and economy, ranks China among the top three countries for global AI vibrancy. 1 “Global AI Vibrancy Tool: Who’s leading the global AI race?” Artificial Intelligence Index, Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford University, 2021 ranking. On research, for example, China produced about one-third of both AI journal papers and AI citations worldwide in 2021. In economic investment, China accounted for nearly one-fifth of global private investment funding in 2021, attracting $17 billion for AI start-ups. 2 Daniel Zhang et al., Artificial Intelligence Index report 2022 , Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford University, March 2022, Figure 4.2.6, “Private investment in AI by geographic area, 2013–21.”

Five types of AI companies in China

In China, we find that AI companies typically fall into one of five main categories:

• Hyperscalers develop end-to-end AI technology capability and collaborate within the ecosystem to serve both business-to-business and business-to-consumer companies.
• Traditional industry companies serve customers directly by developing and adopting AI in internal transformation, new-product launch, and customer services.
• Vertical-specific AI companies develop software and solutions for specific domain use cases.
• AI core tech providers provide access to computer vision, natural-language processing, voice recognition, and machine learning capabilities to develop AI systems.
• Hardware companies provide the hardware infrastructure to support AI demand in computing power and storage.

Today, AI adoption is high in China in finance, retail, and high tech, which together account for more than one-third of the country’s AI market (see sidebar “Five types of AI companies in China”). 3 iResearch, iResearch serial market research on China’s AI industry III , December 2020. In tech, for example, leaders Alibaba and ByteDance, both household names in China, have become known for their highly personalized AI-driven consumer apps. In fact, most of the AI applications that have been widely adopted in China to date have been in consumer-facing industries, propelled by the world’s largest internet consumer base and the ability to engage with consumers in new ways to increase customer loyalty, revenue, and market valuations.

So what’s next for AI in China?

About the research

This research is based on field interviews with more than 50 experts within McKinsey and across industries, along with extensive analysis of McKinsey market assessments in Europe, the United States, Asia, and China specifically between October and November 2021. In performing our analysis, we looked outside of commercial sectors, such as finance and retail, where there are already mature AI use cases and clear adoption. In emerging sectors with the highest value-creation potential, we focused on the domains where AI applications are currently in market-entry stages and could have a disproportionate impact by 2030. Applications in these sectors that either are in the early-exploration stage or have mature industry adoption, such as manufacturing-operations optimization, were not the focus for the purpose of the study.

In the coming decade, our research indicates that there is tremendous opportunity for AI growth in new sectors in China, including some where innovation and R&D spending have traditionally lagged global counterparts: automotive, transportation, and logistics; manufacturing; enterprise software; and healthcare and life sciences. (See sidebar “About the research.”) In these sectors, we see clusters of use cases where AI can create upwards of $600 billion in economic value annually. (To provide a sense of scale, the 2021 gross domestic product in Shanghai, China’s most populous city of nearly 28 million, was roughly $680 billion.) In some cases, this value will come from revenue generated by AI-enabled offerings, while in other cases, it will be generated by cost savings through greater efficiency and productivity. These clusters are likely to become battlegrounds for companies in each sector that will help define the market leaders.

Unlocking the full potential of these AI opportunities typically requires significant investments—in some cases, much more than leaders might expect—on multiple fronts, including the data and technologies that will underpin AI systems, the right talent and organizational mindsets to build these systems, and new business models and partnerships to create data ecosystems, industry standards, and regulations. In our work and global research , we find many of these enablers are becoming standard practice among companies getting the most value from AI.

To help leaders and investors marshal their resources to accelerate, disrupt, and lead in AI, we dive into the research, first sharing where the biggest opportunities lie in each sector and then outlining the core enablers to be tackled first.

Following the money to the most promising sectors

We looked at the AI market in China to determine where AI could deliver the most value in the future. We studied market projections at length and dug deep into country and segment-level reports worldwide to see where AI was delivering the greatest value across the global landscape. We then spoke in depth with experts across sectors in China to understand where the greatest opportunities could emerge next. Our research led us to several sectors: automotive, transportation, and logistics, which are collectively expected to contribute the majority—around 64 percent—of the $600 billion opportunity; manufacturing, which will drive another 19 percent; enterprise software, contributing 13 percent; and healthcare and life sciences, at 4 percent of the opportunity.

Within each sector, our analysis shows the value-creation opportunity concentrated within only two to three domains. These are typically in areas where private-equity and venture-capital-firm investments have been high in the past five years and successful proof of concepts have been delivered.

Automotive, transportation, and logistics

China’s auto market stands as the largest in the world, with the number of vehicles in use surpassing that of the United States. The sheer size—which we estimate to grow to more than 300 million passenger vehicles on the road in China by 2030—provides a fertile landscape of AI opportunities. Indeed, our research finds that AI could have the greatest potential impact on this sector, delivering more than $380 billion in economic value. This value creation will likely be generated predominantly in three areas: autonomous vehicles, personalization for auto owners, and fleet asset management.

Autonomous, or self-driving, vehicles. Autonomous vehicles make up the largest portion of value creation in this sector ($335 billion). Some of this new value is expected to come from a reduction in financial losses, such as medical, first-responder, and vehicle costs. Roadway accidents stand to decrease an estimated 3 to 5 percent annually as autonomous vehicles actively navigate their surroundings and make real-time driving decisions without being subject to the many distractions, such as text messaging, that tempt humans. Value would also come from savings realized by drivers as cities and enterprises replace passenger vans and buses with shared autonomous vehicles. 4 Estimate based on McKinsey analysis. Key assumptions: 3 percent of light vehicles and 5 percent of heavy vehicles on the road in China to be replaced by shared autonomous vehicles; accidents to be reduced by 3 to 5 percent with adoption of autonomous vehicles.

Already, significant progress has been made by both traditional automotive OEMs and AI players to advance autonomous-driving capabilities to level 4 (where the driver doesn’t need to pay attention but can take over controls) and level 5 (fully autonomous capabilities in which inclusion of a steering wheel is optional). For instance, WeRide, which achieved level 4 autonomous-driving capabilities, 5 Based on WeRide’s own assessment/claim on its website. completed a pilot of its Robotaxi in Guangzhou, with nearly 150,000 trips in one year without any accidents with active liability. 6 The pilot was conducted between November 2019 and November 2020.

Personalized experiences for car owners. By using AI to analyze sensor and GPS data—including vehicle-parts conditions, fuel consumption, route selection, and steering habits—car manufacturers and AI players can increasingly tailor recommendations for hardware and software updates and personalize car owners’ driving experience. Automaker NIO’s advanced driver-assistance system and battery-management system, for instance, can track the health of electric-car batteries in real time, diagnose usage patterns, and optimize charging cadence to improve battery life span while drivers go about their day. Our research finds this could deliver $30 billion in economic value by reducing maintenance costs and unanticipated vehicle failures, as well as generating incremental revenue for companies that identify ways to monetize software updates and new capabilities. 7 Estimate based on McKinsey analysis. Key assumptions: AI will generate 5 to 10 percent savings in customer maintenance fee (hardware updates); car manufacturers and AI players will monetize software updates for 15 percent of fleet.

Fleet asset management. AI could also prove critical in helping fleet managers better navigate China’s immense network of railway, highway, inland waterway, and civil aviation routes, which are some of the longest in the world. Our research finds that $15 billion in value creation could emerge as OEMs and AI players specializing in logistics develop operations research optimizers that can analyze IoT data and identify more fuel-efficient routes and lower-cost maintenance stops for fleet operators. 8 Estimate based on McKinsey analysis. Key assumptions: 5 to 15 percent cost reduction in automotive fleet fuel consumption and maintenance; approximately 2 percent cost reduction for aircrafts, vessels, and trains. One automotive OEM in China now offers fleet owners and operators an AI-driven management system for monitoring fleet locations, tracking fleet conditions, and analyzing trips and routes. It is estimated to save up to 15 percent in fuel and maintenance costs.

Manufacturing

In manufacturing, China is evolving its reputation from a low-cost manufacturing hub for toys and clothes to a leader in precision manufacturing for processors, chips, engines, and other high-end components. Our findings show AI can help facilitate this shift from manufacturing execution to manufacturing innovation and create $115 billion in economic value.

The majority of this value creation ($100 billion) will likely come from innovations in process design through the use of various AI applications, such as collaborative robotics that create the next-generation assembly line, and digital twins that replicate real-world assets for use in simulation and optimization engines. 9 Estimate based on McKinsey analysis. Key assumptions: 40 to 50 percent cost reduction in manufacturing product R&D based on AI adoption rate in 2030 and improvement for manufacturing design by sub-industry (including chemicals, steel, electronics, automotive, and advanced industries). With digital twins, manufacturers, machinery and robotics providers, and system automation providers can simulate, test, and validate manufacturing-process outcomes, such as product yield or production-line productivity, before commencing large-scale production so they can identify costly process inefficiencies early. One local electronics manufacturer uses wearable sensors to capture and digitize hand and body movements of workers to model human performance on its production line. It then optimizes equipment parameters and setups—for example, by changing the angle of each workstation based on the worker’s height—to reduce the likelihood of worker injuries while improving worker comfort and productivity.

The remainder of value creation in this sector ($15 billion) is expected to come from AI-driven improvements in product development. 10 Estimate based on McKinsey analysis. Key assumptions: 10 percent cost reduction in manufacturing product R&D based on AI adoption rate in 2030 and improvement for product R&D by sub-industry (including electronics, machinery, automotive, and advanced industries). Companies could use digital twins to rapidly test and validate new product designs to reduce R&D costs, improve product quality, and drive new product innovation. On the global stage, Google has offered a glimpse of what’s possible: it has used AI to rapidly assess how different component layouts will alter a chip’s power consumption, performance metrics, and size. This approach can yield an optimal chip design in a fraction of the time design engineers would take alone.

Would you like to learn more about QuantumBlack, AI by McKinsey ?

Enterprise software.

As in other countries, companies based in China are undergoing digital and AI transformations, leading to the emergence of new local enterprise-software industries to support the necessary technological foundations.

Solutions delivered by these companies are estimated to deliver another $80 billion in economic value. Offerings for cloud and AI tooling are expected to provide more than half of this value creation ($45 billion). 11 Estimate based on McKinsey analysis. Key assumptions: 12 percent CAGR for cloud database in China; 20 to 30 percent CAGR for AI tooling. In one case, a local cloud provider serves more than 100 local banks and insurance companies in China with an integrated data platform that enables them to operate across both cloud and on-premises environments and reduces the cost of database development and storage. In another case, an AI tool provider in China has developed a shared AI algorithm platform that can help its data scientists automatically train, predict, and update the model for a given prediction problem. Using the shared platform has reduced model production time from three months to about two weeks.

AI-driven software-as-a-service (SaaS) applications are expected to contribute the remaining $35 billion in economic value in this category. 12 Estimate based on McKinsey analysis. Key assumptions: 17 percent CAGR for software market; 100 percent SaaS penetration rate in China by 2030; 90 percent of the use cases empowered by AI in enterprise SaaS applications. Local SaaS application developers can apply multiple AI techniques (for instance, computer vision, natural-language processing, machine learning) to help companies make predictions and decisions across enterprise functions in finance and tax, human resources, supply chain, and cybersecurity. A leading financial institution in China has deployed a local AI-driven SaaS solution that uses AI bots to offer personalized training recommendations to employees based on their career path.

Healthcare and life sciences

In recent years, China has stepped up its investment in innovation in healthcare and life sciences with AI. China’s “14th Five-Year Plan” targets 7 percent annual growth by 2025 for R&D expenditure, of which at least 8 percent is devoted to basic research. 13 “‘14th Five-Year Plan’ Digital Economy Development Plan,” State Council of the People’s Republic of China, January 12, 2022.

One area of focus is accelerating drug discovery and increasing the odds of success, which is a significant global issue. In 2021, global pharma R&D spend reached $212 billion, compared with $137 billion in 2012, with an approximately 5 percent compound annual growth rate (CAGR). Drug discovery takes 5.5 years on average, which not only delays patients’ access to innovative therapeutics but also shortens the patent protection period that rewards innovation. Despite improved success rates for new-drug development, only the top 20 percent of pharmaceutical companies worldwide realized a breakeven on their R&D investments after seven years.

Another top priority is improving patient care, and Chinese AI start-ups today are working to build the country’s reputation for providing more accurate and reliable healthcare in terms of diagnostic outcomes and clinical decisions.

Our research suggests that AI in R&D could add more than $25 billion in economic value in three specific areas: faster drug discovery, clinical-trial optimization, and clinical-decision support.

Rapid drug discovery. Novel drugs (patented prescription drugs) currently account for less than 30 percent of the total market size in China (compared with more than 70 percent globally), indicating a significant opportunity from introducing novel drugs empowered by AI in discovery. We estimate that using AI to accelerate target identification and novel molecules design could contribute up to $10 billion in value. 14 Estimate based on McKinsey analysis. Key assumptions: 35 percent of AI enablement on novel drug discovery; 10 percent revenue from novel drug development through AI empowerment. Already more than 20 AI start-ups in China funded by private-equity firms or local hyperscalers are collaborating with traditional pharmaceutical companies or independently working to develop novel therapeutics. Insilico Medicine, by using an end-to-end generative AI engine for target identification, molecule design, and lead optimization, discovered a preclinical candidate for pulmonary fibrosis in less than 18 months at a cost of under $3 million. This represented a significant reduction from the average timeline of six years and an average cost of more than $18 million from target discovery to preclinical candidate. This antifibrotic drug candidate has now successfully completed a Phase 0 clinical study and entered a Phase I clinical trial.

Clinical-trial optimization. Our research suggests that another $10 billion in economic value could result from optimizing clinical-study designs (process, protocols, sites), optimizing trial delivery and execution (hybrid trial-delivery model), and generating real-world evidence. 15 Estimate based on McKinsey analysis. Key assumptions: 30 percent AI utilization in clinical trials; 30 percent time savings from real-world-evidence expedited approval. These AI use cases can reduce the time and cost of clinical-trial development, provide a better experience for patients and healthcare professionals, and enable higher quality and compliance. For instance, a global top 20 pharmaceutical company leveraged AI in combination with process improvements to reduce the clinical-trial enrollment timeline by 13 percent and save 10 to 15 percent in external costs. The global pharmaceutical company prioritized three areas for its tech-enabled clinical-trial development. To accelerate trial design and operational planning, it utilized the power of both internal and external data for optimizing protocol design and site selection. For streamlining site and patient engagement, it established an ecosystem with API standards to leverage internal and external innovations. To establish a clinical-trial development cockpit, it aggregated and visualized operational trial data to enable end-to-end clinical-trial operations with full transparency so it could predict potential risks and trial delays and proactively take action.

Clinical-decision support. Our findings indicate that the use of machine learning algorithms on medical images and data (including examination results and symptom reports) to predict diagnostic outcomes and support clinical decisions could generate around $5 billion in economic value. 16 Estimate based on McKinsey analysis. Key assumptions: 10 percent higher early-stage cancer diagnosis rate through more accurate AI diagnosis; 10 percent increase in efficiency enabled by AI. A leading AI start-up in medical imaging now applies computer vision and machine learning algorithms on optical coherence tomography results from retinal images. It automatically searches and identifies the signs of dozens of chronic illnesses and conditions, such as diabetes, hypertension, and arteriosclerosis, expediting the diagnosis process and increasing early detection of disease.

How to unlock these opportunities

During our research, we found that realizing the value from AI would require every sector to drive significant investment and innovation across six key enabling areas (exhibit). The first four areas are data, talent, technology, and significant work to shift mindsets as part of adoption and scaling efforts. The remaining two, ecosystem orchestration and navigating regulations, can be considered collectively as market collaboration and should be addressed as part of strategy efforts.

Some specific challenges in these areas are unique to each sector. For example, in automotive, transportation, and logistics, keeping pace with the latest advances in 5G and connected-vehicle technologies (commonly referred to as V2X) is crucial to unlocking the value in that sector. Those in healthcare will want to stay current on advances in AI explainability; for providers and patients to trust the AI, they must be able to understand why an algorithm made the decision or recommendation it did.

Broadly speaking, four of these areas—data, talent, technology, and market collaboration—stood out as common challenges that we believe will have an outsized impact on the economic value achieved. Without them, tackling the others will be much harder.

For AI systems to work properly, they need access to high-quality data, meaning the data must be available, usable, reliable, relevant, and secure. This can be challenging without the right foundations for storing, processing, and managing the vast volumes of data being generated today. In the automotive sector, for instance, the ability to process and support up to two terabytes of data per car and road data daily is necessary for enabling autonomous vehicles to understand what’s ahead and delivering personalized experiences to human drivers. In healthcare, AI models need to take in vast amounts of omics 17 “Omics” includes genomics, epigenomics, transcriptomics, proteomics, metabolomics, interactomics, pharmacogenomics, and diseasomics. data to understand diseases, identify new targets, and design new molecules.

Companies seeing the highest returns from AI—more than 20 percent of earnings before interest and taxes (EBIT) contributed by AI—offer some insights into what it takes to achieve this. McKinsey’s 2021 Global AI Survey shows that these high performers are much more likely to invest in core data practices, such as rapidly integrating internal structured data for use in AI systems (51 percent of high performers versus 32 percent of other companies), establishing a data dictionary that is accessible across their enterprise (53 percent versus 29 percent), and developing well-defined processes for data governance (45 percent versus 37 percent).

Participation in data sharing and data ecosystems is also crucial, as these partnerships can lead to insights that would not be possible otherwise. For instance, medical big data and AI companies are now partnering with a wide range of hospitals and research institutes, integrating their electronic medical records (EMR) with publicly available medical-research data and clinical-trial data from pharmaceutical companies or contract research organizations. The goal is to facilitate drug discovery, clinical trials, and decision making at the point of care so providers can better identify the right treatment procedures and plan for each patient, thus increasing treatment effectiveness and reducing chances of adverse side effects. One such company, Yidu Cloud, has provided big data platforms and solutions to more than 500 hospitals in China and has, upon authorization, analyzed more than 1.3 billion healthcare records since 2017 for use in real-world disease models to support a variety of use cases including clinical research, hospital management, and policy making.

The state of AI in 2021

In our experience, we find it nearly impossible for businesses to deliver impact with AI without business domain knowledge. Knowing what questions to ask in each domain can determine the success or failure of a given AI effort. As a result, organizations in all four sectors (automotive, transportation, and logistics; manufacturing; enterprise software; and healthcare and life sciences) can benefit from systematically upskilling existing AI experts and knowledge workers to become AI translators—individuals who know what business questions to ask and can translate business problems into AI solutions. We like to think of their skills as resembling the Greek letter pi ( π ). This group has not only a broad mastery of general management skills (the horizontal bar) but also spikes of deep functional knowledge in AI and domain expertise (the vertical bars).

To build this talent profile, some companies upskill technical talent with the requisite skills. One AI start-up in drug discovery, for instance, has created a program to train newly hired data scientists and AI engineers in pharmaceutical domain knowledge such as molecule structure and characteristics. Company executives credit this deep domain knowledge among its AI experts with enabling the discovery of nearly 30 molecules for clinical trials. Other companies seek to arm existing domain talent with the AI skills they need. An electronics manufacturer has built a digital and AI academy to provide on-the-job training to more than 400 employees across different functional areas so that they can lead various digital and AI projects across the enterprise.

Technology maturity

McKinsey has found through past research that having the right technology foundation is a critical driver for AI success. For business leaders in China, our findings highlight four priorities in this area:

Increasing digital adoption. There is room across industries to increase digital adoption. In hospitals and other care providers, many workflows related to patients, personnel, and equipment have yet to be digitized. Further digital adoption is required to provide healthcare organizations with the necessary data for predicting a patient’s eligibility for a clinical trial or providing a physician with intelligent clinical-decision-support tools.

The same holds true in manufacturing, where digitization of factories is low. Implementing IoT sensors across manufacturing equipment and production lines can enable companies to accumulate the data necessary for powering digital twins.

Implementing data science tooling and platforms. The cost of algorithmic development can be high, and companies can benefit greatly from using technology platforms and tooling that streamline model deployment and maintenance, just as they benefit from investments in technologies to improve the efficiency of a factory production line. Some essential capabilities we recommend companies consider include reusable data structures, scalable computation power, and automated MLOps capabilities . All of these contribute to ensuring AI teams can work efficiently and productively.

Advancing cloud infrastructures. Our research finds that while the percent of IT workloads on cloud in China is almost on par with global survey numbers, the share on private cloud is much bigger due to security and data compliance concerns. As SaaS vendors and other enterprise-software providers enter this market, we advise that they continue to advance their infrastructures to address these concerns and provide enterprises with a clear value proposition. This will require further advances in virtualization, data-storage capacity, performance, elasticity and resilience, and technological agility to customize business capabilities, which enterprises have come to expect from their vendors.

Investments in AI research and advanced AI techniques. Many of the use cases described here will require fundamental advances in the underlying technologies and techniques. For instance, in manufacturing, additional research is needed to improve the performance of camera sensors and computer vision algorithms to detect and recognize objects in dimly lit environments, which can be common on factory floors. In life sciences, further innovation in wearable devices and AI algorithms is necessary to enable the collection, processing, and integration of real-world data in drug discovery, clinical trials, and clinical-decision-support processes. In automotive, advances for improving self-driving model accuracy and reducing modeling complexity are required to enhance how autonomous vehicles perceive objects and perform in complex scenarios.

For conducting such research, academic collaborations between enterprises and universities can advance what’s possible.

Market collaboration

AI can present challenges that transcend the capabilities of any one company, which often gives rise to regulations and partnerships that can further AI innovation. In many markets globally, we’ve seen new regulations, such as Global Data Protection Regulation (GDPR) in Europe and the California Consumer Privacy Act in the United States, begin to address emerging issues such as data privacy, which is considered a top AI relevant risk in our 2021 Global AI Survey . And proposed European Union regulations designed to address the development and use of AI more broadly will have implications globally.

Our research points to three areas where additional efforts could help China unlock the full economic value of AI:

Data privacy and sharing. For individuals to share their data, whether it’s healthcare or driving data, they need to have an easy way to give permission to use their data and have trust that it will be used appropriately by authorized entities and safely shared and stored. Guidelines related to privacy and sharing can create more confidence and thus enable greater AI adoption. A 2019 law enacted in China to improve citizen health, for instance, promotes the use of big data and AI by developing technical standards on the collection, storage, analysis, and application of medical and health data. 18 Law of the People’s Republic of China on Basic Medical and Health Care and the Promotion of Health, Article 49, 2019.

Meanwhile, there has been significant momentum in industry and academia to build methods and frameworks to help mitigate privacy concerns. For example, the number of papers mentioning “privacy” accepted by the Neural Information Processing Systems, a leading machine learning conference, has increased sixfold in the past five years. 19 Artificial Intelligence Index report 2022 , March 2022, Figure 3.3.6.

Market alignment. In some cases, new business models enabled by AI will raise fundamental questions around the use and delivery of AI among the various stakeholders. In healthcare, for instance, as companies develop new AI systems for clinical-decision support, debate will likely emerge among government and healthcare providers and payers as to when AI is effective in improving diagnosis and treatment recommendations and how providers will be reimbursed when using such systems. In transportation and logistics, issues around how government and insurers determine culpability have already arisen in China following accidents involving both autonomous vehicles and vehicles operated by humans. Settlements in these accidents have created precedents to guide future decisions, but further codification can help ensure consistency and clarity.

Standard processes and protocols. Standards enable the sharing of data within and across ecosystems. In the healthcare and life sciences sectors, academic medical research, clinical-trial data, and patient medical data need to be well structured and documented in a uniform manner to accelerate drug discovery and clinical trials. A push by the National Health Commission in China to build a data foundation for EMRs and disease databases in 2018 has led to some movement here with the creation of a standardized disease database and EMRs for use in AI. However, standards and protocols around how the data are structured, processed, and connected can be beneficial for further use of the raw-data records.

Likewise, standards can also eliminate process delays that can derail innovation and scare off investors and talent. An example involves the acceleration of drug discovery using real-world evidence in Hainan’s medical tourism zone; translating that success into transparent approval protocols can help ensure consistent licensing across the country and ultimately would build trust in new discoveries. On the manufacturing side, standards for how organizations label the various features of an object (such as the size and shape of a part or the end product) on the production line can make it easier for companies to leverage algorithms from one factory to another, without having to undergo costly retraining efforts.

Patent protections. Traditionally, in China, new innovations are rapidly folded into the public domain, making it difficult for enterprise-software and AI players to realize a return on their sizable investment. In our experience, patent laws that protect intellectual property can increase investors’ confidence and attract more investment in this area.

AI has the potential to reshape key sectors in China. However, among business domains in these sectors with the most valuable use cases, there is no low-hanging fruit where AI can be implemented with little additional investment. Rather, our research finds that unlocking maximum potential of this opportunity will be possible only with strategic investments and innovations across several dimensions—with data, talent, technology, and market collaboration being foremost. Working together, enterprises, AI players, and government can address these conditions and enable China to capture the full value at stake.

Kai Shen and Ting Wu are partners in McKinsey’s Shenzhen office, and Xiaoxiao Tong is a consultant in the Shanghai office, where Fangning Zhang is a partner.

The authors wish to thank Forest Hou, Joanna Mak, Tamim Saleh, Christoph Sandler, Alex Sawaya, Florian Then, Joanna Wu, Xiaolu Xu, and Jeff Yang for their contributions to this article.

Explore a career with us

Related articles.

Reimagining your business for AI

Scaling AI like a tech native: The CEO’s role

The Future of Digital Innovation in China: Megatrends Shaping One of the World’s Fastest Evolving Digital Ecosystems

Center for Security and Emerging Technology

Comparing U.S. and Chinese Contributions to High-Impact AI Research

Ashwin Acharya

and Brian Dunn

In the past decade, Chinese researchers have become increasingly prolific authors of highly cited AI publications, approaching the global research share of their U.S. counterparts. However, some analysts question the impact of Chinese publications; are they well respected internationally, and do they cover important topics? In this data brief, the authors build on prior analyses of top AI publications to provide a richer understanding of the two countries’ contributions to high-impact AI research.

Executive Summary

While the large and growing number of Chinese artificial intelligence publications is well known, the quality of this research is debated. Some observers claim that China is capable of producing a high quantity of AI publications, but lags in original ideas and impactful research. 1 Even Chinese researchers occasionally criticize their country’s academic system for its lack of innovation in AI. 2 In recent years, however, quantitative analyses have found that Chinese AI publications are increasingly influential. 3

AI is an economically and strategically important emerging technology, and the Chinese government has promoted domestic AI progress for years. Chinese and U.S. strengths in AI development will have ramifications for the two countries’ relative capabilities in areas ranging from science and medicine to battlefield applications. Further, Chinese researchers’ ability to produce impactful AI advances reflects on the more general question of whether Beijing can foster impactful innovation—a capability sometimes called into question by U.S. and European observers. 4

This brief provides a data-driven comparison of U.S. and Chinese AI research, examining both publications that are highly cited and those published in top AI conferences. 5

We find that:

• Chinese researchers’ output of highly cited AI publications is increasingly competitive with the work of their U.S. counterparts. Over the past decade, Chinese researchers have published a growing share of the world’s top-5-percent AI publications, rising from half of U.S. output in 2010 to parity in 2019.
• Top Chinese publications are often cited outside of China, although China still lags behind the United States in international citations. Highly cited Chinese publications receive 35 percent of their citations from non-Chinese sources, and their citation count from international sources has steadily increased over time. However, U.S. publications maintain a lead over Chinese ones in international citations, reflecting the United States’ closer ties to other leading AI producers.
• China contributes an increasing share of publications at 13 top AI conferences, while the U.S. share of publications at these conferences is stagnant. Between 2010 and 2019, China’s share of these publications grew from 13 percent to 31 percent, while the U.S. share fell from 55 percent to 51 percent.
• A notable share of both U.S. and Chinese researchers’ high-impact AI publications were U.S.-Chinese collaborations. For example, such collaborations accounted for 24 percent of both countries’ highly cited AI publications in 2019.
• Clusters with a disproportionate share of China’s highly cited and top-venue publications include publications on general-purpose computer vision research, as well as applications of AI to surveillance and industry.  
• Clusters with a disproportionate share of the United States’ highly cited and top-venue publications cover algorithmic innovations in deep learning, such as transformers and deep reinforcement learning, as well as AI ethics and safety research.
• The United States and China combined publish about 65 percent of highly cited AI research. U.S. allies, particularly the European Union and the Five Eyes countries, also make significant contributions to AI research. 6

Download Full Report

• For example, a McKinsey report claims that “China lags behind the United States and the United Kingdom in terms of fundamental research that advances the field of AI.” Dominic Barton, Jonathan Woetzel, Jeongmin Seong, and Qinzheng Tian, “Artificial Intelligence: Implications for China” (McKinsey & Company, April 2017), https://www.mckinsey.com/~/media/mckinsey/featured percent20insights/China/Artificial percent20intelligence percent20Implications percent20for percent20China/MGI-Artificial-intelligence-implications-for-China.ashx
• For example, several Chinese researchers recently wrote that “although aggregate AI research outputs (e.g., scientific publications, patents) are rising rapidly in China, truly original ideas and breakthrough technologies are lacking.” Daitian Li, Tony W. Wong, and Yangao Xiao, “Is China Emerging as the Global Leader in AI?,” Harvard Business Review, February 18, 2021, https://hbr.org/2021/02/is-china-emerging-as-the-global-leader-in-ai .
• Jiangjiang Yang and Oren Etzioni, “China is closing in on the US in AI research,” Allen Institute for AI (Medium), May 11, 2021 , https://medium.com/ai2-blog/china-is-closing-in-on-the-us-in-ai-research-ea5213ae80df ; Dewey Murdick, James Dunham, and Jennifer Melot, “AI Definitions Affect Policymaking” (Center for Security and Emerging Technology, June 2020), https://cset.georgetown.edu/wp-content/uploads/CSET-AI-Definitions-Affect-Policymaking.pdf .
• Robert D. Atkinson and Caleb Foote, “Is China Catching Up to the United States in Innovation?” (Information Technology & Innovation Foundation, April 2019), https://projects.iq.harvard.edu/files/innovation/files/2019-china-catching-up-innovation.pdf .
• Our analysis is not limited to publications in academic journals and conferences; it also includes preprints on the ArXiv repository, which private AI labs often use to report their latest innovations. For example, most of the publications linked on OpenAI’s publications page are ArXiv preprints. Such preprints can still appear in our subset of highly cited AI publications. For example, OpenAI’s ArXiv preprint “Deep Double Descent: Where Bigger Models and More Data Hurt” appears in the CSET merged corpus as one of the most highly cited AI publications of 2019, placing in the highest percentile for computer science publications in that year. Preetum Nakkiran, “Deep Double Descent: Where Bigger Models and More Data Hurt,” arXiv preprint arXiv:1912.02292 (2019), https://arxiv.org/abs/1912.02292 .
• We refer to the Five Eyes countries, excluding the United States, as CANZUK. This group includes Canada, the United Kingdom, Australia, and New Zealand. In this brief, European Union refers to the 27 member states of the after the departure of the United Kingdom: Austria, Belgium, Bulgaria, Croatia, Cyprus, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, the Netherlands, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, and Sweden.

This website uses cookies.

Privacy overview.

China now publishes more high-quality science than any other nation – should the US be worried?

Milton & Roslyn Wolf Chair in International Affairs, The Ohio State University

Disclosure statement

Caroline Wagner does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

The Ohio State University provides funding as a founding partner of The Conversation US.

View all partners

By at least one measure, China now leads the world in producing high-quality science . My research shows that Chinese scholars now publish a larger fraction of the top 1% most cited scientific papers globally than scientists from any other country.

I am a policy expert and analyst who studies how governmental investment in science, technology and innovation improves social welfare. While a country’s scientific prowess is somewhat difficult to quantify, I’d argue that the amount of money spent on scientific research, the number of scholarly papers published and the quality of those papers are good stand-in measures.

China is not the only nation to drastically improve its science capacity in recent years, but China’s rise has been particularly dramatic. This has left U.S. policy experts and government officials worried about how China’s scientific supremacy will shift the global balance of power . China’s recent ascendancy results from years of governmental policy aiming to be tops in science and technology. The country has taken explicit steps to get where it is today, and the U.S. now has a choice to make about how to respond to a scientifically competitive China.

Growth across decades

In 1977, Chinese leader Deng Xiaoping introduced the Four Modernizations , one of which was strengthening China’s science sector and technological progress. As recently as 2000, the U.S. produced many times the number of scientific papers as China annually. However, over the past three decades or so, China has invested funds to grow domestic research capabilities, to send students and researchers abroad to study, and to encourage Chinese businesses to shift to manufacturing high-tech products.

Since 2000, China has sent an estimated 5.2 million students and scholars to study abroad . The majority of them studied science or engineering. Many of these students remained where they studied, but an increasing number return to China to work in well-resourced laboratories and high-tech companies.

Today, China is second only to the U.S. in how much it spends on science and technology . Chinese universities now produce the largest number of engineering Ph.D.s in the world, and the quality of Chinese universities has dramatically improved in recent years .

Producing more and better science

Thanks to all this investment and a growing, capable workforce, China’s scientific output – as measured by the number of total published papers – has increased steadily over the years. In 2017, Chinese scholars published more scientific papers than U.S. researchers for the first time.

Quantity does not necessarily mean quality though. For many years, researchers in the West wrote off Chinese research as low quality and often as simply imitating research from the U.S. and Europe . During the 2000s and 2010s, much of the work coming from China did not receive significant attention from the global scientific community.

But as China has continued to invest in science, I began to wonder whether the explosion in the quantity of research was accompanied by improving quality.

To quantify China’s scientific strength, my colleagues and I looked at citations. A citation is when an academic paper is referenced – or cited – by another paper. We considered that the more times a paper has been cited, the higher quality and more influential the work. Given that logic, the top 1% most cited papers should represent the upper echelon of high-quality science.

My colleagues and I counted how many papers published by a country were in the top 1% of science as measured by the number of citations in various disciplines. Going year by year from 2015 to 2019, we then compared different countries. We were surprised to find that in 2019, Chinese authors published a greater percentage of the most influential papers , with China claiming 8,422 articles in the top category, while the U.S had 7,959 and the European Union had 6,074. In just one recent example, we found that in 2022, Chinese researchers published three times as many papers on artificial intelligence as U.S. researchers; in the top 1% most cited AI research, Chinese papers outnumbered U.S. papers by a 2-to-1 ratio. Similar patterns can be seen with China leading in the top 1% most cited papers in nanoscience, chemistry and transportation.

Our research also found that Chinese research was surprisingly novel and creative – and not simply copying western researchers. To measure this, we looked at the mix of disciplines referenced in scientific papers. The more diverse and varied the referenced research was in a single paper, the more interdisciplinary and novel we considered the work. We found Chinese research to be as innovative as other top performing countries.

Taken together, these measures suggest that China is now no longer an imitator nor producer of only low-quality science. China is now a scientific power on par with the U.S. and Europe, both in quantity and in quality.

Fear or collaboration?

Scientific capability is intricately tied to both military and economic power. Because of this relationship, many in the U.S. – from politicians to policy experts – have expressed concern that China’s scientific rise is a threat to the U.S., and the government has taken steps to slow China’s growth. The recent Chips and Science Act of 2022 explicitly limits cooperation with China in some areas of research and manufacturing. In October 2022, the Biden administration put restrictions in place to limit China’s access to key technologies with military applications .

A number of scholars, including me, see these fears and policy responses as rooted in a nationalistic view that doesn’t wholly map onto the global endeavor of science.

Academic research in the modern world is in large part driven by the exchange of ideas and information. The results are published in publicly available journals that anyone can read. Science is also becoming ever more international and collaborative , with researchers around the world depending on each other to push their fields forward. Recent collaborative research on cancer , COVID-19 and agriculture are just a few of many examples. My own work has also shown that when researchers from China and the U.S. collaborate, they produce higher quality science than either one alone.

China has joined the ranks of top scientific and technological nations, and some of the concerns over shifts of power are reasonable in my view. But the U.S. can also benefit from China’s scientific rise. With many global issues facing the planet – like climate change , to name just one – there may be wisdom in looking at this new situation as not only a threat, but also an opportunity.

• Geopolitics
• Science policy
• Globalization
• Science and technology
• Chinese science
• CHIPS and Science Act

Communications Director

Associate Director, Post-Award, RGCF

University Relations Manager

2024 Vice-Chancellor's Research Fellowships

Community member RANZCO Education Committee (Volunteer)

• Open the Search Form

Publications

• Congressional Testimony

Research Areas

• Hard Choices in Defense
• The Future of Warfare
• Strengthening Deterrence
• The Gaming Lab
• Defense Discussions
• The China Challenge
• Regional Alliances and Partnerships
• The India Opportunity
• The North Korea Threat
• Confronting Threats to Democracy
• NATO and European Security
• Responding to Strategic Challenges
• Israel-Hamas War
• Evolving the Mission–Iraq/Syria/ISIS
• Iran Futures
• Constructing Regional Partnerships and Seizing Emerging Opportunities
• Security Assistance
• Targeted Sanctions: Russia and Iran
• Economic Statecraft
• Energy & Geopolitics
• Sanctions by the Numbers
• Artificial Intelligence
• Technology Strategy
• Critical Digital Infrastructure
• Biotechnology
• Enhancing DHS Oversight & Accountability
• Congress and National Security
• Renewing the National Security Consensus

Resident Experts

• All Resident Experts
• Molly Campbell
• Vivek Chilukuri
• Carrie Cordero
• Lisa Curtis
• Michael Depp
• Bill Drexel
• Joshua Fitt
• Richard Fontaine
• Geoffrey Gertz
• Noah Greene
• Kate Johnston
• Andrea Kendall-Taylor
• Emily Kilcrease
• Katherine L. Kuzminski
• Nicholas Lokker
• Jonathan Lord
• Gibbs McKinley
• Andrew Metrick
• Carisa Nietsche
• Stacie Pettyjohn
• Ruby Scanlon
• Paul Scharre
• Philip Sheers
• Jacob Stokes
• Taren Sylvester
• Adam H. Tong
• Constanza M. Vidal Bustamante
• Josh Wallin
• Becca Wasser
• Caleb Withers
• Robert O. Work

Adjunct Experts

• All Adjunct Experts
• Manpreet Singh Anand
• Brandon J. Archuleta
• Lt. Gen. Robert Ashley, Jr. (Ret.)
• Christian Beckner
• Samuel Bendett
• Paul Benfield
• Rachel Brandenburg
• Dr. Jonathan Brewer
• Josh Campbell
• Pablo Chavez
• Richard Connolly
• John Costello
• Richard J. Danzig
• Anthony DeMartino
• Jason Dempsey
• Billy Fabian
• Yaya J. Fanusie
• Ryan Fedasiuk
• David Feith
• Edward Fishman
• Ben FitzGerald
• Dr. Erik Lin-Greenberg
• Michael Greenwald
• Mikhail Grinberg
• Hamzeh Hadad
• Hon. Robert F. Hale
• Heli Hautala
• Franz-Stefan Gady
• Dr. Jeannette Gaudry Haynie
• Jacob Helberg
• Michelle Holko, PhD, PMP
• General Mike Holmes, U.S. Air Force (Ret.)
• John Hughes
• Dr. Kyleanne Hunter
• LTG Anthony R. "Tony" Ierardi, USA (Ret.)
• Akira Igata
• Elsa B. Kania
• Robert D. Kaplan
• Dr. Duyeon Kim
• Christopher D. Kolenda
• Margarita "Rita" Konaev
• Steven Kosiak
• Dr. Andrew F. Krepinevich, Jr.
• Thomas Krueger
• Peter L. Levin
• Jennifer McArdle
• Brendan McCord
• Dr. ED McGrady
• Jack Midgley
• J Travis Mosier
• Dr. Go Myong-Hyun
• Catherine A. Novelli
• Dr. John Park
• Dr. Lynne E. Parker
• Diem Salmon
• Jordan Schneider
• Peter Schroeder
• Michael Sellitto
• Vance Serchuk
• John (Jack) N.T. Shanahan
• Tom Shugart
• Daniel Silverberg
• Alexander Sullivan
• Tobias Switzer
• Rachel Tecott Metz
• Alanna C. Torres-Van Antwerp
• Jim Townsend
• Richard R. Verma
• Anthony Vinci
• Jon B. Wolfsthal
• Alex Zerden
• Rachel Ziemba
• Full-Time Staff
• Executive Team
• Board of Directors
• Board of Advisors
• Directors Emeriti
• Distinguished Senior Fellows
• Adjunct Fellows
• Senior Military Fellows
• Joseph S. Nye Interns
• CNAS Supporters
• Next Generation National Security Fellows
• Visiting Fellows
• Writer in Residence
• In Memoriam
• Internships

CNAS Programs

• The Shawn Brimley Next Generation National Security Leaders Fellowship
• Robert M. Gates Fellowship
• 1LT Andrew J. Bacevich Jr., USA Award
• Joseph S. Nye, Jr. National Security Internship and Mentoring Program
• The Pitch: A Competition of New Ideas
• Women in National Security
• Writers in Residence
• Senior Military Fellows Program
• Join the CNAS Council
• CNAS Corporate Partnership Program
• Frequently Asked Questions
• CNAS 2024 National Security Conference: High Stakes

February 06, 2019

Understanding China's AI Strategy

Clues to Chinese Strategic Thinking on Artificial Intelligence and National Security

By: Gregory C. Allen

Introduction

In the second half of 2018, I traveled to China on four separate trips to attend major diplomatic, military, and private-sector conferences focusing on Artificial Intelligence (AI). During these trips, I participated in a series of meetings with high-ranking Chinese officials in China’s Ministry of Foreign Affairs, leaders of China’s military AI research organizations, government think tank experts, and corporate executives at Chinese AI companies. From these discussions – as well as my ongoing work analyzing China’s AI industry, policies, reports, and programs – I have arrived at a number of key judgments about Chinese leadership’s views, strategies, and prospects for AI as it applies to China’s economy and national security. Of course, China’s leadership in this area is a large population with diversity in its views, and any effort to generalize is inherently presumptuous and essentially guaranteed to oversimplify. However, the distance is large between prevailing views in American commentary on China’s AI efforts and what I have come to believe are the facts. I hope by stating my takeaways directly, this report will advance the assessment of this issue and be of benefit to the wider U.S. policymaking community.

Chinese Views on the Importance of AI

1. china’s leadership – including president xi jinping – believes that being at the forefront in ai technology is critical to the future of global military and economic power competition..

In July 2017, China’s State Council issued the New Generation Artificial Intelligence Development Plan (AIDP). 1 This document – along with Made in China 2025 , 2 released in May 2015 – form the core of China’s AI strategy. Both documents, as well as the issue of AI more generally, have received significant and sustained attention from the highest levels of China’s leadership, including Xi Jinping. Total Chinese national and local government spending on AI to implement these plans is not publicly disclosed, but it is clearly in the tens of billions of dollars. At least two 3 Chinese regional governments have each committed to investing 100 billion yuan (~$14.7 billion USD). 4 The opening paragraphs of the AIDP exemplify mainstream Chinese views regarding AI:

AI has become a new focus of international competition. AI is a strategic technology that will lead in the future; the world’s major developed countries are taking the development of AI as a major strategy to enhance national competitiveness and protect national security. 5

The above quote also reflects how China’s AI policy community 6 is paying close attention to the AI industries and policies of other countries, particularly the United States. Chinese government organizations routinely translate, disseminate, and analyze U.S. government and think tank reports about AI. In my conversations with Chinese officials and my reading of Chinese government AI reports, they demonstrated substantive and timely knowledge of AI developments in the United States and elsewhere. Chinese government AI reports frequently cite U.S. national security think tank publications. 7 The U.S. policymaking community ought to make it a priority to be equally effective at translating, analyzing, and disseminating Chinese publications on AI for the insights they provide into Chinese thinking. 8

2. China’s leadership – including Xi Jinping – believes that China should pursue global leadership in AI technology and reduce its vulnerable dependence on imports of international technology.

In October 2018, Xi Jinping led a Politburo study session on AI. Such sessions are reserved for the high-priority policy issues where leaders need the benefit of outside expertise. Xi’s publicly reported comments during and after the study session reiterated the main conclusions of both the AIDP and Made in China 2025 , which were that China should “achieve world-leading levels” 9 in AI technology and reduce its vulnerable “external [foreign] dependence for key technologies and advanced equipment.” 10

In his speech during the study session, Xi said that China must “ensure that our country marches in the front ranks where it comes to theoretical research in this important area of AI, and occupies the high ground in critical and AI core technologies.” 11 Xi further said that China must “pay firm attention to the structure of our shortcomings, ensure that critical and core AI technologies are firmly grasped in our own hands.” Xi’s speech demonstrates that China’s leadership continues to subscribe to AIDP’s and Made in China 2025 ’s two major conclusions that China should pursue both world leadership and self-reliance in AI technology. The Chinese AI sector’s dependence on foreign technology is discussed further in point nine.

Chinese Views on AI's International Security Implications

3. recently, chinese officials and government reports have begun to express concern in multiple diplomatic forums about arms race dynamics associated with ai and the need for international cooperation on new norms and potentially arms control..

In a keynote speech during China’s largest international relations conference on July 15, 2018, Fu Ying, 12 the Vice-Chair of the Foreign Affairs Committee of the National People’s Congress, said that Chinese technologists and policymakers agree regarding the “threat of the new [AI] technology to mankind.” She further stated that “We believe that we should cooperate to preemptively prevent the threat of AI.”

Madam Fu’s depiction of AI as posing a shared threat to international security was echoed by many other Chinese diplomats and PLA think tank scholars in my private meetings with them. For instance, one official told me he was concerned that AI “will lower the threshold of military action,” because states may be more willing to attack each other with AI military systems due to the lack of casualty risk. Chinese officials also expressed concern that increased used of AI systems would make misperceptions and unintentional conflict escalation more likely due to the lack of well-defined norms regarding the use of such systems. Additionally, Chinese officials displayed substantive knowledge of the cybersecurity risks associated with AI sytems, as well as their implications for Chinese and international security.

Madam Fu said that China was interested in playing a leading role in creating norms to mitigate these risks. At the World Peace Forum private roundtable on AI, one senior PLA think tank scholar privately expressed support for “mechanisms that are similar to arms control” for AI systems in cybersecurity and military robotics. However, he also said that AI-related arms control would be uniquely difficult since “AI is low-cost and can be disseminated easily and cannot be monitored easily.”

Notably, the recent “Artificial Intelligence Security White Paper,” published in September 2018 by the China Academy of Information and Communications Technology (CAICT), an influential Chinese government think tank, calls upon the Chinese government to “avoid Artificial Intelligence arms races among countries.” 13 The AIDP does not address arms races but does state that China will “deepen international cooperation on AI laws and regulations, international rules and so on, and jointly cope with global challenges.”

Such concerns extend to the China’s private sector. Jack Ma, the chairman of Alibaba, said explicitly in a speech at the 2019 Davos World Economic Forum that he was concerned that global competition over AI could lead to war. 14

4. Despite expressing concern on AI arms races, most of China’s leadership sees increased military usage of AI as inevitable and is aggressively pursuing it. China already exports armed autonomous platforms and surveillance AI.

At the Beijing Xiangshan Forum on October 24, 2018, Major General Ding Xiangrong, Deputy Director of the General Office of China’s Central Military Commission, gave a major speech in which he defined China’s military goals to “narrow the gap between the Chinese military and global advanced powers” by taking advantage of the “ongoing military revolution . . . centered on information technology and intelligent technology.” Chinese military leaders increasingly refer to intelligent or “intelligentized” (智能化) military technology as their confident expectation for the future basis of warfare. Use of the term “intelligentized” is meant to signify a new stage of military technology beyond the current stage based on information technology. 15 China’s AIDP strategy document states that China will “Promote all kinds of AI technology to become quickly embedded in the field of national defense innovation.”

The next day at the Xiangshan Forum, Zeng Yi, a senior executive at China’s third largest defense company, 16 gave a speech in which he described his company’s (and China’s) expectations for the future implementation of AI weapons: “In future battlegrounds, there will be no people fighting.” Zeng predicted that by 2025 lethal autonomous weapons would be commonplace and said that his company believes ever-increasing military use of AI is “inevitable […] We are sure about the direction and that this is the future.”

Zeng’s comments are consistent with ongoing Chinese autonomous military vehicle development programs and China’s current approach to exports of military unmanned systems. China’s government already is exporting many of its most advanced military aerial drones to Middle Eastern countries such as Saudi Arabia and the UAE. China’s government has stated that it also will export its next generation stealth drones when those are available. 17 Though many current generation drones are primarily remotely operated, Chinese officials generally expect drones and military robotics to feature ever more extensive AI and autonomous capabilities in the future. Chinese weapons manufacturers already are selling armed drones with significant amounts of combat autonomy. Ziyan, a Chinese military drone manufacturer, has sold its Blowfish A2 model to the UAE and in November 2019 reportedly was in negotiations with Saudi Arabia and Pakistan for Blowfish A2 sales. 18 Ziyan’s website states that the 38kg Blowfish A2 “autonomously performs more complex combat missions, including fixed-point timing detection, fixed-range reconnaissance, and targeted precision strikes.” 19 Depending on customer preferences, Ziyan offers to equip Blowfish A2 with either missiles or machine guns.

Beyond using AI for autonomous military robotics, China is also interested in AI capabilities for military command decisionmaking. Zeng Yi expressed some remarkable opinions on this subject, stating that today “mechanized equipment is just like the hand of the human body. In future intelligent wars, AI systems will be just like the brain of the human body.” Zeng also said that “Intelligence supremacy will be the core of future warfare” and that “AI may completely change the current command structure, which is dominated by humans” to one that is dominated by an “AI cluster.” Zeng did not elaborate on his claims, but they are consistent with broader thinking in Chinese military circles. Several months after AlphaGo’s momentous March 2016 victory over Lee Sedol, a publication by China’s Central Military Commission Joint Operations Command Center argued that AlphaGo’s victory “demonstrated the enormous potential of artificial intelligence in combat command, program deduction, and decisionmaking.” 20

China is currently making extensive use of AI in domestic surveillance applications. General Wang Ning of the Chinese People’s Armed Police Force recently boasted about China’s success in using AI in Xinjiang province:

In Xinjiang, we use big data AI to fight terrorists. We have intercepted 1200 terror organizations when still planning an attack. We use technology to identify and locate activities of terrorists, including the smart city system. We have a face recognition system, and for all terrorists there is a database. 21

Xinjiang is home to millions of China’s Uighur ethnic minority, which has been subject to extraordinary persecution aided by AI surveillance technology. 22 China’s SenseTime corporation, a national champion in computer vision AI, is a major provider of surveillance technology to China’s government, including for Xinjiang. SenseTime’s security and surveillance products often are described using the “smart city” euphemism. However, SenseTime also has many non-security products, such as computer vision machine learning related to autonomous vehicles.

SenseTime is a major exporter of surveillance technology in government and commercial markets across Latin America, Africa, and Asia. China’s government and leadership is enthusiastic about using AI for surveillance. One scholar at a Chinese think tank told me that he looks forward to a world in AI will make it “impossible” to “commit a crime without being caught,” a sentiment that echoes the marketing materials put out by Chinese AI surveillance companies.

China’s behavior of aggressively developing, utilizing, and exporting increasingly autonomous robotic weapons and surveillance AI technology runs counter to China’s stated goals of avoiding an AI arms race. However, this by itself does not necessarily mean that Chinese officials are being insincere in their expressions of concern about such arms races. Lamenting arms race dynamics while aggressively participating in them is a common story in the history of international relations. The strongest behavioral indication that China might be insincere comes from China’s April 2018 United Nations position paper, 23 in which China’s government supported a worldwide ban on “lethal autonomous weapons” but used such a bizarrely narrow definition of lethal autonomous weapons that such a ban would appear to be both unnecessary and useless. This rhetorical gambit allowed China to reap positive media attention for their support of global restrictions while avoiding hypocrisy over Chinese development of more advanced military AI and autonomy. 24 More broadly there seems to be less grassroots concern of the issue among Chinese AI researchers than their counterparts in the West, though evidence on this point is limited. 25

5. China’s Ministry of National Defense has established two major new research organizations focused on AI and unmanned systems under the National University of Defense Technology (NUDT).

The National Innovation Institute of Defense Technology (NIIDT, an NUDT subsidiary), has established and is rapidly growing two Beijing-based research organizations focusing on the military use of AI and related tech. These are the Unmanned Systems Research Center (USRC), led by Yan Ye, and the Artificial Intelligence Research Center (AIRC), led by Dai Huadong. 26 Each organization was created in early 2018, and each now has a research staff of over 100 (more than 200 total), which makes it one of the largest and fastest growing government AI research organizations in the world. However, there are larger private sector AI research organizations in both China and the United States. SenseTime, for example, has roughly 600 full-time research staff. DeepMind – a Google subsidiary focused on AI research – has around 700 total staff and annual expenditures of over $400 million. 27 Salaries of Chinese AI PhD’s educated in China are generally much lower than salaries of Western AI PhD’s, or Western-educated Chinese, which makes estimating the AIRC’s budget based on staff difficult. AIRC staff are engaged in basic research into dual-use AI technology, including applying machine learning to robotics, swarm networking, wireless communications, and cybersecurity. The AIRC also likely does classified work for the Chinese Military and Intelligence Community.

6. China’s government sees AI as a promising military “leapfrog development” opportunity, meaning that it offers military advantages over the US and will be easier to implement in China than the United States.

The term “leapfrog development” describes a technology for which laggard countries can skip a development stage, or one for which being behind on the current generation of technology actually offers an advantage in adopting the next generation. A commonly cited example is the rapid and widespread adoption of cellular phone technology in countries that had only minimal landline phone adoption. Kai-Fu Lee, one of the leading venture capitalists in China’s AI sector, argues that the absence of many developed-economy capabilities, such as easy credit checks, have led to a flood of Chinese entrepreneurs making innovative use of AI capabilities to fill those gaps. 28 Plastic credit cards are nearly nonexistent in China, but mobile phone payments secured by facial recognition are ubiquitous.

China’s emphasis on AI as a leapfrog technology enabler extends to national security applications. China’s 2017 National AI Development Plan identifies AI as a “historic opportunity” for national security leapfrog technologies. 29 Chinese Defense executive Zeng Yi echoed that claim, saying that AI will “bring about a leapfrog development” in military technology and presents a critical opportunity for China.

If China is correct that AI presents a leapfrog opportunity, it would mean that China is better positioned to adopt military AI than the United States. In this theory, the United States’ current advantages in stealth aircraft, aircraft carriers, and precision munitions actually would be long-term disadvantages because the entrenched business and political interests that support military dominance today will hamper the United States in transitioning to an AI-enabled military technology paradigm in the future. 30 As one Chinese think tank scholar explained to me, China believes that the United States is likely to spend too much to maintain and upgrade mature systems and underinvest in disruptive new systems that make America’s existing sources of advantage vulnerable and obsolete. China’s military also faces perverse incentives to protect legacy systems, but to a far lesser extent: Military spending tripled over the 2007–2017 period, 31 modernization is a top priority, and there is a general understanding that many of its current platforms and approaches are obsolete and must be replaced regardless.

Just one of many examples of China’s AI leapfrog strategy is its prioritized investment 32 and technology espionage 33 for low-cost, long-range, autonomous, and unmanned submarines. China believes these systems will be a cheap and effective means of threatening U.S. aircraft carrier battlegroups and an alternative path to projecting Chinese power at range. In general, China sees military AI R&D as a cheaper and easier path to threatening America’s sources of military power than developing Chinese equivalents of American systems.

Chinese Views on the Strengths of China's AI Ecosystem

7. china’s government and industry believe that they have largely closed the gap with the united states in both ai r&d and commercial ai products. china now sees ai as “a race of two giants,” between itself and the united states..

China’s July 2017 national AI strategy set a 2020 goal for China’s “AI industry’s competitiveness [to] have entered the first echelon internationally.” In truth, China’s leadership already assesses China as having achieved this objective as of mid-2018. At the World Peace Forum, Tsinghua University’s Xue Lan delivered a briefing on Tsinghua University’s major report on the state of the AI sector in China. 34 This study found that “China has secured a leading position in the top [AI] echelon in both technology development and market applications and is in a race of ‘two giants’ with the U.S.” It also finds that China is:

• #1 in both total AI research papers and highly cited AI papers worldwide
• #1 in AI patents
• #1 in AI venture capital investment
• #2 in the number of AI companies
• #2 in the largest AI talent pool.

China’s assessment of being in the first echelon is correct, though there are important caveats that will be discussed more below. Not only is China advancing the state of the art in AI research, its companies are very successfully developing genuinely innovative and market-competitive products and services around AI applications. SenseTime, for example, is undisputedly one of the world leaders in computer vision AI and claims to have achieved annual revenue growth of 400 percent for three consecutive years. DJI offers another example. DJI leads the world in consumer drones with 74 percent market share. 35 DJI has innovatively incorporated machine learning technology into its most recent products.

In many cases the products and underlying technologies between commercial AI and military/security AI products are identical or nearly so. DJI recently was selected as the sole drone provider to the New York Police Department, which will use DJI’s consumer model drones. Similarly, SenseTime’s consumer facial recognition systems share infrastructure and technology with its security systems, used by both Chinese law enforcement and intelligence organizations.

8. China’s strong current position in AI R&D and commercial applications has been enabled by access to international markets, technology, and research collaboration.

China’s success has been enabled by its access to global technology research and markets. Many seemingly “Chinese” AI achievements are actually achievements of multinational research teams and companies, and such international collaboration has been critical to China’s research progress. 36 According to the Tsinghua University study of China’s AI ecosystem, “More than half of China’s AI papers were international joint publications,” meaning that Chinese AI researchers – the top tier of whom often received their degrees abroad – were coauthoring with non-Chinese individuals. Even purely Chinese successes often build upon open source technologies developed most often by international groups.

Partly as a result of this, leading Chinese technology companies have significant and under-reported dependencies on the United States. For example, DJI, the Shenzhen-headquartered, world-leading drone manufacturer, is vertically integrated with nearly all design, manufacturing, and marketing done in-house. However, all of DJI’s drone flight software development is performed at DJI’s American office in Palo Alto, which predominantly employs U.S. citizens as staff. Additionally, nearly 35 percent of the bill of materials in each of DJI’s products are from the United States, mostly reflecting semiconductor content.

Chinese View on the Weaknesses of China's AI Ecosystem

9. despite china’s strength in ai r&d and commercial applications, china’s leadership perceives major weaknesses relative to the united states in top talent, technical standards, software platforms, and semiconductors..

Though most in China’s leadership agree that China is one of two “giants” in AI, there is a similarly widespread understanding that China is not strong in all areas. China’s January 2018 “White Paper on Artificial Intelligence Standardization” points out that the China’s AI ecosystem lags in several key areas:

Although China has a good foundation in the field of AI, even as core technologies such as speech recognition, visual recognition, and Chinese-language information processing have achieved breakthroughs and possess huge market environments for applications, the overall level of development still lags behind that of developed countries. 37

Similarly, the Tsinghua University China AI Development report finds:

China’s strengths are mainly shown in AI applications and it is still weak on the front of core technologies of AI, such as hardware and algorithm development, China’s AI development lacks top-tier talent and has a significant gap with developed countries, especially the U.S., in this regard. 38

There are additional comparative weaknesses in China’s AI ecosystem worth discussing, but I will focus on the four that most often came up in my meetings in China: top talent, technical standards, software platforms, and semiconductors.

Weaknesses in Top Talent

The Tsinghua University China AI report did a remarkable study of the global AI talent distribution, concluding that by the end of 2017, the international AI talent pool comprised 204,575 individuals, with the United States having 28,536 such individuals and China in second place with 18,232. However, China’s ranks eighth in the world in terms of Top AI talent, with only 977 individuals compared to the United States’ 5,518. Though acknowledging the disparity, venture capitalist Kai-Fu Lee argues that this is not a major barrier because “the current age of implementation [AI application commercialization] appears well-suited to China’s strengths in research: large quantities of highly skilled, though not necessarily best-of-best, AI researchers and practitioners.” 39 Some researchers at leading Western AI research insitutions have told me they agree with this conclusion, noting that AI breakthroughs by leading institutions are quickly replicated by other institutions worldwide.

Lee is influential among China’s technology industry, but not everyone agrees with his theory. Many that I spoke with said that China’s shortage of top talent will be a handicap in the future development of China’s AI sector, and China’s government is taking aggressive action to improve the size and quality of China’s AI talent pool. 40 In April 2018, China’s Ministry of Education (MOE) launched its AI Innovation Action Plan for Colleges and Universities . Among other elements, the plan:

• Will create “50 world-class teaching materials for undergraduate and graduate studies” related to AI applications for specific industries;
• Will create “50 national-level high-quality online open courses”;
• Will establish “50 artificial intelligence faculties, research institutions, or interdisciplinary research centers.” 41

In a separate initiative, the MOE also plans to launch a new five-year AI talent training program to train 500 more AI instructors and 5,000 more top students at top Chinese universities. 42

Weaknesses in Technical Standards

The determination and common adoption of international technical standards is a key enabler of technology interoperability and market growth. Common adoption of Wi-Fi standard, for example, is what allowed such a wide diversity of modems, routers, mobile phones, and computers to all effectively connect to each other over Wi-Fi networks. Companies that create intellectual property related to such standards often reap significant rewards, especially when their patents, such as the design of a specific semiconductor chip, are declared essential to effective operation of any device using the standard. 43 For example, Qualcomm’s intellectual property was critical to development of the Code-Division Multiple Access (CDMA) cellular standard. It is essentially impossible for a device to access CDMA cellular networks unless the device uses Qualcomm semiconductor patents, hence why they are an example of so-called “Standard Essential Patents” (SEPs). Historically, Chinese companies and government organizations produced very few SEPs, but China has made rapid progress on this front. Huawei, ZTE, and the China Academy of Telecommunications Technology have produced hundreds of SEPs related to fifth generation (5G) cellular standards. 44

AI technical standards are far less mature than those in cellular networking, but China’s government strategy for pursuing leadership in AI technical standards is informed by its experience in the cellular networking. China’s government and Chinese corporations want to ensure that their intellectual property and products are critical features of the future of AI. Because of China’s experience with ZTE export restrictions, Chinese leadership perceives its success in technical standards as critical to both economic growth and national security.

Weaknesses in Software Frameworks & Platforms

Developers of AI systems rarely start from scratch. More often, they leverage pre-written programs developed by others and shared into code libraries. This allows developers to focus on the unique specifics of their application usage requirements, rather than solving generic problems faced by all AI developers. Some organizations have combined machine learning code libraries with other AI software development tools into mature machine learning software frameworks, many of which are open source. Popular machine learning frameworks include, but are not limited to, TensorFlow (Google), Spark (Apache), CNTK (Microsoft), and PyTorch (Facebook).

Notably, none of the most popular machine learning software frameworks have been developed in China. The importance of leadership in software frameworks is debated even among America’s leading technology companies. Companies that do prioritize framework development claim that it offers opportunities to attract top talent, influence technical standards, and guide the overall ecosystem toward increased usage of their products and services. The absence of Chinese AI companies among the major AI framework developers and open source AI software communities was identified as a noteworthy weakness of China’s AI ecosystem in several of my conversations with executives in China’s technology industry. Additionally, China’s CAICT AI and Security White Paper lamented the fact that “At present, the research and development of domestic artificial intelligence products and applications is mainly based on Google and Microsoft.” 45 SenseTime has devoted extensive resources its own machine learning framework, Parrots, which is intended to be superior for computer vision AI applications. So far, the company appears to have had limited success in promoting adoption: No Chinese computer scientists I met with outside of SenseTime had even heard of Parrots, even though it was announced more than two years ago.

Weaknesses in Semiconductors

Most of the world’s consumer electronics products bear a “Made in China” label. Sixty-five percent of the world’s personal computers, notebooks, and tablets as well as nearly 85 percent of the world’s mobile phones reportedly are made in China. 46 However, many of these products are assembled with high-value semiconductor chips that are designed in the United States, manufactured in Taiwan or Korea, and running software developed by American firms such as Google, Microsoft, and Apple. The iPhone, for example, bears a “Made in China” label, but only low-skill assembly and commodity component production takes place in China. A study found that Chinese contributions account for less than 2 percent of the overall cost of the iPhone, even though 100 percent of the cost of the device is counted in the United States’ trade deficit with China. 47

Even in the consumer drones market, where the leading Chinese company (DJI) enjoys 74 percent global market share, 35 percent of the bill of materials in each drone is actually U.S. content, primarily semiconductors. China brings extraordinary scale, skills, and infrastructure to bear in electronics manufacturing, which accounts for its central role in the global electronics supply chain. However, recent developments suggest that this centrality may be less irreplaceable than is often claimed. In the face of increasing Chinese wages and U.S. tariffs, many international electronics manufacturers, such as Samsung, 48 Apple, and Foxconn, 49 are relocating ever more of their Chinese operations to lower-cost countries such as Vietnam and India. China’s 85 percent share of global mobile phone manufacturing in 2017 is actually down from 90 percent in 2016. 50 In other words, electronics is following other rapidly relocating industries such as textiles. 51 China is attempting to forestall these movements by massively increasing its use of robotics and automation in manufacturing, 52 with unclear prospects.

By contrast, U.S. and international products and services are sometimes irreplaceable, such as when Chinese electronics manufacturer ZTE faced a quick turn from profitability to imminent bankruptcy in the wake of U.S. export restrictions on critical input products such as semiconductors. 53

Why China's Tech Sector is Unlikely to Face Soviet-Style Stagnation

Like the Soviet Union during the Cold War, China today is engaged in an extensive campaign to harvest technological and scientific information from the rest of the world, using both legal and illegal means. Unlike the Soviet Union, China’s efforts have prioritized using such access to build industries that are competitive in global markets and research institutions that lead the world in strategic fields. For example, the Soviet Union gave an overwhelming priority to the military application of illegally imported semiconductor manufacturing equipment well into the 1980s, which essentially guaranteed that the Soviet Union’s industry would remain dependent on Western technology and never reach internationally competitive economies of scale. 54

By contrast, China’s strategy for making effective use of foreign technology is to use it to support domestic commercial industry. When a state-owned Chinese company recently sought to steal U.S. memory chip semiconductor manufacturing technology, the primary motive was to raise the technological competitiveness of China’s domestic semiconductor industry in global markets. 55 China’s leadership has concluded that possessing commercially competitive industries often is of greater long-term benefit to China’s national security sector than short-term military utilization of any stolen technology. For example, China’s approach for AI, as outlined in its national AIDP strategy document, is to:

Follow the rules of the market . . . accelerate the commercialization of AI technology and results, and create a competitive advantage. Grasp well the division of labor between government and the market . 56

The Soviet Union had a large community of brilliant scientists and technologists, but this community spent a disproportionate amount of its creative and intellectual potential on compensating for the shortcomings of the Soviet system. On top of perverse institutional incentives divorced from economic reality, the Soviet economy was deliberately self-isolated from global trade. 57 Compared with the Soviet Union’s non-market communist economy, China’s policies promoting market-oriented entrepreneurship have made them far superior consumers of international and especially U.S. technology, whether gathered by legal or illegal means. Despite sensational successes in the Space Race and some key military technologies, overall, the Soviet Union fell further and further behind each year that the Cold War dragged on. China, by contrast, has gone from a scientific backwater to a leading player in a long list of scientific fields and technology industries in just two decades.

China's Near-Term Goal: Maintain Access to Foreign Technology But Reduce Dependence

10. china's leaders seek to preserve access to foreign technology in the short term but believe that they must promote domestic independence in the longterm. this has long been china's goal, but it has taken on new urgency..

In November 2018, Dr. Tan Tieniu, Deputy Secretary-General of the Chinese Academy of Sciences, gave a wide-ranging speech before many of China’s most senior leadership at the 13th National People’s Congress Standing Committee. In the speech, he argued that China’s lagging status in technical standards, software frameworks, and semiconductors left China vulnerable and in dire need of domestic alternatives. Due to the frankness and insightfulness of Dr. Tan’s comments, they are worth quoting at length:

[China should] construct an independent and controllable innovation ecosystem. American companies such as Google, IBM, Microsoft, and Facebook have actively built innovation ecosystems, seized the innovative high ground, and already in the international AI industry hold the upper hand in AI chips, servers, operating systems, open source algorithms, cloud services, and autonomous driving, among others. China’s AI open source community and technological innovation ecosystem are comparatively lagging, the strength of technology platform construction needs to be reinforced, and [China’s] international influence remains to be improved.

The U.S. ban on ZTE fully demonstrates the importance of independent, controllable core-, high-, and foundational technologies. In order to avoid repeating this disaster, China should learn its lesson about importing core electronic components, high-end general-purpose chips, and foundational software. 58

Though expressed in a more urgent tone, Tan’s comments are in line with China’s preexisting technology policy. The Tsinghua University AI Report conducted a comprehensive quantitative analysis of Chinese technology policy documents and found that Made in China 2025 is the single most important policy underpinning Chinese regional governments’ development of AI policies. 59 The regional governments bear primary responsibility for implementing the strategic objectives laid out by the central government. Made in China 2025 notably outlines policies across various industries for China to reduce dependency on foreign technology, either by developing it indigenously or acquiring it from foreign sources, and thereafter capture global market share.

Tan Tieniu also argued that China can leverage its existing strength in AI applications to improve its position in other parts of the AI value chain, such as international standards. “As China is at the global forefront of AI technology applications, it should seize its right to speak in the formulation of international AI standards,” he said.

11. China's pursuit of reducing foreign dependence is bearing fruit, as show by increasing value capture share by Chinese suppliers in the global smartphone market supply chain and China's success in advanced semiconductor design.

A 2011 study 60 of which countries capture what share of revenue from each sale of the iPhone found that the factories assembling the iPhone in China captured less than 2 percent of the value 61 of each iPhone sold and that there were no Chinese suppliers to the iPhone other than assembly laborers. 62 By contrast, nearly half of the value of each device was captured by Chinese companies in the case of Huawei’s 2017 flagship P9 smartphone, a direct iPhone competitor. 63 For Huawei, these value capture share gains are not limited to low-skill tasks. Huawei’s HiSilicon subsidiary designed the main semiconductor processor of the P9, including its AI deep learning accelerator element, in-house. 64 Indeed, the study arguably understates China’s value capture in smartphones because it undercounts China’s software gains. Though Chinese firms are not major competitors in the smartphone operating system market, Tencent’s WeChat app fulfills many of the functions of an operating system and is ubiquitous among Chinese smartphone owners.

There are three major segments of the semiconductor value chain: design, manufacturing, and assembly. 65 China historically has only been a major player in assembly, which is relatively low skill. Recently, Chinese companies have demonstrated remarkably high quality and competitive semiconductor design, exemplified by Huawei’s Kirin 980. The Kirin 980 is one of only two smartphone processors in the world to use 7 nanometer (nm) process technology, the other being the Apple-designed A12 Bionic. Both Apple and Huawei rely upon Taiwain’s TSMC for outsourced 7nm manufacturing. Even the most advanced Chinese semiconductor manufacturers are only in 2019 introducing 14nm technology, which international firms such as Intel and Samsung achieved in 2014. SMIC, China’s most advanced semiconductor manufacturer, hopes to reach 7nm manufacturing in the early 2020s, 66 which would still be significanty behind the most advanced global competitors, though possibly by a smaller margin. 67

The Importance of Semiconductors to Future AI Competition

12. other than military ai applications, the future focus of strategic national ai competition is likely to be the semiconductor industry, 68 because the cutting edge of ai technology increasingly depends on custom computer chips..

Historically, AI companies have been able to build competitive advantages based on possessing more and higher quality data to use for training purposes. Data quality, diversity, and especially quantity all remain key sources of competitive advantage for many AI applications, but there are two caveats to this. First, much of the training data for machine learning is application-specific. This means, for example, that having a large quantity of health care data does nothing if one’s goal is to develop a driverless car. Second, some applications of AI can use so-called “synthetic data,” 69 created through computational simulation or self-play, to reduce or eliminate the performance advantage from very large quantities of real-world data.

Training machine learning algorithms on large data sets is very computationally intensive. Running simulations to generate synthetic data is, for many applications, even more computationally intensive. For the large and growing set of AI applications where massive data sets are needed or where synthetic data is viable, AI performance is often limited by computing power. 70 This is especially true for the state-of-the-art AI research. 71 As a result, leading technology companies and AI research institutions are investing vast sums of money in acquiring high performance computing systems.

Chinese companies and government laboratories are strong in high performance computing and specifically on efficient high performance AI computing. China’s SenseTime, for example, revealed in December 2018 that its aggregate computing power is more than 160 petaflops, more than the world’s top-ranked supercomputer at Oak Ridge National Laboratory. 72 SenseTime’s computing infrastructure includes more than 54,000,000 Graphical Processing Unit (GPU) cores across 15,000 GPUs within 12 GPU clusters. Such numbers indicate that SenseTime has spent hundreds of millions of $USD on computing infrastructure. SenseTime’s computer network spans multiple countries but is not connected to the Internet, using a so-called “under the top” setup. At the JP Morgan Asia TMT conference on November 14, 2018, where SenseTime was presenting to potential investors, cofounder Bing Xu said that SenseTime’s willingness to invest in supercomputing infrastructure was critical to its overall ability to generate IP and sustainable competitive advantages. He further said that “30–40 percent” of SenseTime’s research team is devoted to improving SenseTime’s internal machine learning framework, Parrots, and improving SenseTime’s computing infrastructure. Several Chinese researchers told me that they consider China’s expertise in designing and integrating high-performance computing systems to be one of China’s strongest advantages in AI.

Most of the world’s GPUs are designed by NVIDIA in the United States and manufactured by TSMC in Taiwan. At the moment, China does not have a major manufacturer or designer of advanced GPUs. However, the GPU’s current position as the most commonly used AI computing accelerator chip is under increased competition from chips custom-designed to run AI applications. 73 Many traditionally software-focused U.S. technology companies, such as Google and Amazon, have created and acquired semiconductor design divisions specifically to work on AI accelerator chips. These chips can offer dramatically superior performance over GPUs for AI applications even when manufactured using older processes and equipment. The first generation of Google’s primary AI chip, called a Tensor Processing Unit (TPU), for example, is manufactured using 28 nanometer process technology, which is already widely available in China. Google claimed in 2017 that its first generation TPU was 15–30 times faster and 30–80 times more power efficient for AI workloads than contemporary GPUs. 74

Chinese firms Baidu (in partnership with Intel), 75 Alibaba (via a new subsidiary, Pingtouge), 76 and Huawei (via its HiSilicon subsidiary) have all established semiconductor design divisions focused on developing AI accelerator chips. Chinese AI chip startups Horizon Robotics and Cambricon have raised hundreds of millions of $USD in venture capital funding at multibillion-dollar valuations. 77

China's Prospects for AI and Semiconductors

13. china's prospects in the ai chip semiconductor market are strong, likely stronger than they are in the overall semiconductor industry..

China’s goal as outlined in Made in China 2025 is to increase domestic semiconductor manufacturing as a share of domestic consumption to 80 percent by 2030 and to reduce all external dependences, including reliance on Taiwanese firms such as TSMC. According to China’s Semiconductor Industry Association (CSIA), Chinese producers are on track to increase their share of domestic consumption from 29 percent in 2014 (the year before Made in China 2025 was announced) to 49 percent by the end of 2019. 78 However, most of these gains have been in product segments that do not require the most advanced semiconductors, which remain a large share of the market. 79 In its Q4 2018 financial disclosures, TSMC (which has roughly half of the global semiconductor foundry market share) 80 revealed that nearly 17 percent of its revenue came from eight-year old 28nm processes, and that 37 percent came from even older processes. 81 Chinese manufacturers plan to prioritize those market segments where older processes can be competitive.

AI chips offer Chinese manufacturers a uniquely attractive opening for their older process technology. As mentioned above, AI chips can offer potentially superior performance and cost than state-of-the-art GPUs even while using less advanced manufacturing processes. 82 The rise of AI chips therefore offers China the chance to combine its highly advanced semiconductor design and AI software sectors to expand market share and competitiveness in the broader semiconductor industry. Though flagship mobile phones likely will always demand the most advanced generation of semiconductor manufacturing processes, many applications can be addressed with older technology nodes. With low-cost AI chips, this could be a uniquely attractive, diverse, and rapidly growing set of applications. One Chinese industry observer has openly promoted this exact strategy. 83 Understanding of the importance of AI chips appears to be increasingly widespread in China. The recent Tsinghua University “White Paper on AI Chip Technologies” demonstrates a deep understanding of all the relevant technology and market dynamics. That report strongly emphasizes the strategic importance of AI chips:

Whether it is the realization of algorithms, the acquisition and a massive database, or the computing capability, the secret behind the rapid development of the AI industry lies in the one and only physical basis, that is, the chips. Therefore, it is no exaggeration to say “No chip, no AI” given the irreplaceable role of AI chip as the cornerstone for AI development and its strategic significance. 84

At the same time, China hopes to use success in AI chips to build an enduring competitive advantage in the overall AI industry, underpinned by superior computing capacity, larger datasets, and a more favorable regulatory environment. This is a high priority area for China’s AI companies and government. Yu Kai, the CEO of Chinese AI Chip startup Horizon Robotics, is an influential member of China’s Ministry of Science and Technology (MOST) 85 AI Strategic Advisory Committee. 86

14. Where China is behind in AI and semiconductors, present trends suggest that the gap will narrow. This is a key government priority, receiving enormous attention and investment.

In both AI and semiconductors, China has dramatically shrunk the gap between its domestic firms and leading international ones. Absent some kind of major change in U.S. policy to increase competitiveness, or a major Chinese economic crisis, China’s policies likely will be sufficient to ensure that over the next 5 years China secures a defensible competitive advantage across many AI application markets and at least narrows the gap between Chinese and non-Chinese firms in many semiconductor market segments.

In 2014, China’s government established a national integrated circuit industry investment fund 87 to reduce China’s dependence on foreign semiconductors. The first fund ultimately invested 138.7 billion RMB ($20.5 billion) and was followed in 2018 by a second government fund that will reportedly invest 300 billion RMB ($44.5 billion). Recent moves by the United States – including the Obama administration’s April 2015 decision to restrict semiconductor exports to Chinese supercomputing centers and the Trump administration’s previously mentioned semiconductor export restrictions on ZTE – have strengthened the conclusion of China’s leadership that increasing “self-reliance” is more important than ever. Dr. Tan Tieniu stated this explicitly in his November Party Congress speech before China’s leadership, and Alibaba cofounder Jack Ma publicly announced similar conclusions in April 2018: “the market for chips is controlled by Americans,” Ma said. “And suddenly if they stop selling – what that means, you understand. And that’s why China, Japan, and any country, you need core technologies.” 88

As demonstrated by Huawei, the top tier of China’s semiconductor design segment is already competitive at the global state of the art. Chinese design firms benefit from access to world-leading Taiwanese semiconductor foundry companies that manufacture semiconductors but do not design them.

The primary barriers to additional Chinese semiconductor manufacturing progress are access to the most advanced semiconductor manufacturing equipment and access to skilled workers with the knowledge of and training in how to effectively implement the most advanced manufacturing processes. China is making significant progress on both points, but the gap in the number of skilled workers is notably large given the scale of China’s semiconductor industry growth ambitions. 89

While it does not possess any of the world’s most advanced equipment manufacturing companies, China has strong negotiating leverage with foreign companies due to the size and growth of its domestic market. Semiconductor manufacturing equipment sales in China represented 11.8 percent ($6.5B) of the global market in 2017 but are expected to grow in 2019 to 25.6 percent ($17.3B). 90 Recently, semiconductor equipment manufacturers in Europe have signed deals with Chinese companies to export critical 7nm manufacturing equipment. 91 China also has successfully recruited many workers and executives from leading Taiwanese semiconductor companies, 92 including SMIC’s new co-CEO, who has a documented history of stealing intellectual property. 93 When I toured a Samsung semiconductor lab, they noted that all of the printer paper in the building was laced with a metallic thread to set off the exit door metal detectors, a potent illustration of Samsung’s view that intellectual property theft is a significant threat.

15. Adverse macroeconomic factors and a potential financial bubble could slow China's AI sector growth.

China’s venture capital and technology entrepreneurial ecosystem is one of the country’s major strengths. Chinese AI startups increased their share of global AI equity investment to 48 percent in 2017, while U.S. startups attracted 38 percent. 94 However, China’s investment is concentrated on far fewer firms, most of which have extraordinarily high valuations relative to their current profitability. Several leading Chinese investors have hypothesized that this represents a financial bubble in China’s technology sector, where growth is fueled primarily by the sector’s easy access to investment capital rather than prospects for profitable revenue growth. 95 If true, such a bubble would not call into question the existence of China’s strong AI sector but rather its financial sustainability. Additionally, in the second half of 2018, China’s tech sector saw reports of sufficiently widespread layoffs that office real estate prices fell in the major technology districts of Beijing. 96 The broader macroeconomic climate in China also worsened in 2018, partly as a result of China’s trade dispute with the United States. It is difficult to determine what extent this reflects a tech sector slowdown, a change in the financial environment, or merely the tech sector’s share of macroeconomic headwinds. However, a major technology sector downturn or economic recession would make it difficult for China’s government and companies to afford the R&D investments necessary to improve competitiveness.

The Importance of Commercial AI Success to Chinese Power

16. china's success in commercial ai and semiconductor markets has direct relevance to china's geopolitical power as well as its military and espionage ai capabilities..

China’s commercial market success has direct relevance to China’s national security, both because it reduces the ability of the United States government to put diplomatic and economic pressure on China and because it increases the technological capabilities available to China’s military and intelligence community. Regarding the latter, essentially all major technology firms in China cooperate extensively with China’s military and state security services and are legally required to do so. Article 7 of China’s National Intelligence Law gives the government legal authority to compel such assistance, though the government also has powerful non-coercive tools to incentivize cooperation. 97 “Military-Civil Integration” is one of the cornerstones of China’s national AI strategy. Several Chinese executives that I spoke with reported that they are feeling significantly more oversight and pressure from China’s central government, a finding consistent with recent media reports. 98

In 2018, China’s government took the remarkable step of announcing that Baidu, Alibaba, Tencent, iFlytek, and SenseTime were officially the country’s “AI Champions.” SenseTime executives told me that this position gave the companies privileged positions for national technical standards setting and also was intended to give the companies confidence that they would not be threatened with competition from state-owned enterprises. In December, SenseTime cofounder Bing Xu said, “We are very lucky to be a private company working at a technology that will be critical for the next two decades. Historically, governments would dominate nuclear, rocket, and comparable technologies and not trust private companies.” In explicitly comparing AI to nuclear and rocket technology, Xu appears to be referencing the critical role of AI to the future of national security. The price of SenseTime and the other AI Champions being allowed to dominate these technologies is the Champions’ extensive cooperation with China’s national security community. Even beyond direct cooperation, China’s success in commercial AI and semiconductor markets brings funding, talent, and economies of scale that both reduce China’s vulnerability from losing access to international markets and offer useful technology for the development of weaponry and espionage capabilities.

In my interactions with Chinese government officials, they demonstrated remarkably keen understanding of the issues surrounding AI and international security. It is clear that China’s government views AI as a high strategic priority and is devoting the required resources to cultivate AI expertise and strategic thinking among its national security community. This includes knowledge of U.S. AI policy discussions. I believe it is vital that the U.S. policymaking community similarly prioritize cultivating expertise and understanding of AI developments in China. I hope this report has helped contribute to that objective.

Still, no amount of information on China’s AI strategy will be sufficient by itself to meet the competitive challenge posed by China. If the United States wants to lead the world in AI, it will require funding, focus, and a willingness among U.S. policymakers to drive large-scale necessary change. U.S. leaders have more powerful tools to influence the technological and economic competitiveness of the United States than they have tools to influence China’s competitiveness. They should prioritize accordingly.

Acknowledgments

I would like to thank Graham Allison, Jason Matheny, Paul Scharre, Richard Danzig, Matt Daniels, Joseph Nye, Helen Toner, Carrick Flynn, Lora Saalman, Elsa Kania, Ben Chang, and Tim Hwang for for their valuable input, feedback, and suggestions for this report.

I would like to thank Jeffrey Ding, Elsa Kania, Rogier Creemers, Graham Webster, Lorand Laskai, Mingli Shi, Dahlia Peterson, Samm Sacks, Cameron Hickert, Paul Triolo, and others for the extremely valuable work they do translating Chinese government and corporate publications on Artificial Intelligence into English.

I would like to thank the following institutions for their invitations to participate in conferences and meetings in China and Hong Kong on Artificial Intelligence:

• Tsinghua University and the Chinese People’s Institute of Foreign Affairs: World Peace Forum, Beijing, July 13-15, 2018
• China Academy of Military Science and the China Institute of International Strategic Studies: 8th Beijing Xiangshan Forum, Beijing, October 24-26, 2018
• JP Morgan: Global Technology, Media, and Telecom Conference in Asia, Hong Kong, November 14-15, 2018
• Atlantic Council Scowcroft Center and The Swedish Ministry of Foreign Affairs: Fact Finding Trip on Chinese Technological Innovation, Shenzhen, December 3-7, 2018.

Finally, I would like to thank the dozens of individuals with whom I met on trips to China. Their willingness to engage was critical to this report. Where they engaged with an expectation of anonymity or non-public disclosure, I have made sure to respect that here.

Views expressed in this report are the author’s alone. CNAS does not take institutional positions.

Read the full white paper.

Download PDF

• The China State Council’s AIDP is available in English at Graham Webster, Rogier Creemers, Paul Triolo, and Elsa Kania (translators), “Full Translation: China’s ‘New Generation Artificial Intelligence Development Plan,’” New America , August 1, 2017. https://www.newamerica.org/cybersecurity-initiative/digichina/blog/full-translation-chinas-new-generation-artificial-intelligence-development-plan-2017/ . ↩
• China State Council, “Made in China 2025,” July 7, 2015; English translation available at http://www.cittadellascienza.it/cina/wp-content/uploads/2017/02/IoT-ONE-Made-in-China-2025.pdf . ↩
• Xinhua, "Shanghai to Set up Multi-billion-dollar Fund to Develop AI," China Daily, September 18, 2018, http://www.chinadaily.com.cn/a/201809/18/WS5ba0ade9a31033b4f4656be2.html . ↩
• Meng Jing, “This Chinese City Plans a US$16 Billion Fund for AI Development,” South China Morning Post, May 16, 2018, https://www.scmp.com/tech/innovation/article/2146428/tianjin-city-china-eyes-us16-billion-fund-ai-work-dwarfing-eus-plan . ↩
• Webster et al. (transl.), “Full Translation: China’s ‘New Generation Artificial Intelligence Development Plan.’” ↩
• The AIDP was officially released by the Chinese State Council, but the advisory committees and authoring individuals included representation from China’s national security, diplomatic, academic, and private sectors. ↩
• For example, the recent “Artificial Intelligence Security White Paper,” published in September 2018 by the China Academy of Information and Communications Technology, includes a section summarizing my own report. Greg Allen and Taniel Chan, “Artificial Intelligence and National Security,” Harvard Belfer Center for Science and International Affairs, July 2017, https://www.belfercenter.org/publication/artificial-intelligence-and-national-security . ↩
• See the acknowledgments section for a list of some of those engaged in this critical work. ↩
• China State Council. “Made in China 2025.” ↩
• Xi, Jinping’s speech is available at Elsa Kaniia and Rogier Creemers (translators), “Xi Jinping Calls for ‘Healthy Development’ of AI (Translation),” New America. November 5, 2018, https://www.newamerica.org/cybersecurity-initiative/digichina/blog/xi-jinping-calls-for-healthy-development-of-ai-translation/ . ↩
• A former Vice Minister of Foreign Affairs and UK ambassador, Fu Ying plays an important role in advancing Chinese interests before American think tank audiences. See Larry Diamond, and Orville Schell, “Chinese Influence & American Interests: Promoting Constructive Vigilance” (Hoover Institution, 2018), 64. https://www.hoover.org/sites/default/files/research/docs/00_diamond-schell_fullreport_2ndprinting_web-compressed.pdf . ↩
• China Academy for Information and Communications Technology (CAICT) & China Institute of Information and Communications Security, “Artificial Intelligence and Security,” September 2018, http://www.caict.ac.cn/kxyj/qwfb/bps/201809/P020180918473525332978.pdf . ↩
• Ma stated: “The First World War was because of the first technology revolution. The second technology revolution caused the Second World War. This is the third technology revolution – we’re coming.” See Ryan Browne, “Alibaba’s Jack Ma Suggests Technology Could Result in a New World War,” CNBC, January 25, 2019, https://www.cnbc.com/2019/01/23/alibaba-jack-ma-suggests-technology-could-result-in-a-new-world-war.html . ↩
• Elsa Kania, "AlphaGo and Beyond: The Chinese Military Looks to Future ‘Intelligentized’ Warfare." Lawfare. June 07, 2017. https://www.lawfareblog.com/alphago-and-beyond-chinese-military-looks-future-intelligentized-warfare . ↩
• By revenue, NORINCO is the third largest defense company in China and the ninth largest worldwide. ↩
• Dake Kang and Christopher Bodeen, “China Unveils Stealth Combat Drone in Development,” Associated Press, November 07, 2018, https://www.apnews.com/6b2d2857f73c4fa387379c16b0dc60b9 . ↩
• Ludovic Ehret, “China Steps up Drone Race with Stealth Aircraft,” Phys.org, November 9, 2018, https://phys.org/news/2018-11-china-drone-stealth-aircraft.html . ↩
• “Blowfish A2 Product Overview,” Ziyan, January 2019, http://ziyanuav.com/blowfish2.html . ↩
• Central Military Commission Joint Staff, “Accelerate the Construction of a Joint Operations Command System with Our Nation’s Characteristics CMC Joint Operations Command Center,” Seeking Truth, August 15, 2016, http://www.qstheory.cn/dukan/qs/2016-08/15/c_1119374690.htm . More analysis of China’s views regarding AI and command decisionmaking can be found in Elsa Kania, "AI Titans, Entangled?" Hoover Institution, October 29, 2019, https://www.hoover.org/research/ai-titans-entangled . ↩
• General Wang Ning, “Global Terrorism: Threats and Countermeasures” (8 th Beijing Xiangshan Forum, Beijing, October 25, 2018). ↩
• Maya Wang, "Interview: China's Crackdown on Turkic Muslims," Human Rights Watch, September 10, 2018. https://www.hrw.org/news/2018/09/10/interview-chinas-crackdown-turkic-muslims . ↩
• People’s Republic of China, “Position Paper of China,” submitted at the United National “Group of Governmental Experts of the High Contracting Parties to the Convention on Prohibitions or Restrictions on the Use of Certain Conventional Weapons Which May Be Deemed to Be Excessively Injurious or to Have Indiscriminate Effects,” April 11, 2018, https://www.unog.ch/80256EDD006B8954/(httpAssets)/E42AE83BDB3525D0C125826C0040B262/$file/CCW_GGE.1_2018_WP.7.pdf . ↩
• Further analysis of China’s U.N. position paper is available in Kania, Elsa. "China's Strategic Ambiguity and Shifting Approach to Lethal Autonomous Weapons Systems," Lawfare, April 20, 2018, https://www.lawfareblog.com/chinas-strategic-ambiguity-and-shifting-approach-lethal-autonomous-weapons-systems . ↩
• Jeff Ding has pointed out, “of the 3462 AI/robotics researchers who signed a Future of Life Institute open letter to ban autonomous weapons, only three were based at Chinese institutions (all were affiliated with the Chinese University of Hong Kong).” See Jeffrey Ding, “Deciphering China’s AI Dream,” Future of Humanity Institute, Oxford University, March 2018, https://www.fhi.ox.ac.uk/wp-content/uploads/Deciphering_Chinas_AI-Dream.pdf . ↩
• Dr. Dai is also a professor of computer science at the NUDT College of Computer Science. ↩
• Sam Shead, “DeepMind Losses Grew To $368 Million In 2017,” Forbes , October 5, 2018, https://www.forbes.com/sites/samshead/2018/10/05/deepmind-losses-grew-to-302-million-in-2017/#7a137da2490e . ↩
• Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Boston: Houghton Mifflin Harcourt Trade & Reference Publishers, 2018), 112. ↩
• Specifically, the report says that China should “firmly seize the major historic opportunity for the development of AI . . . and support national security, promoting the overall elevation of the nation’s competitiveness and leapfrog development.” ↩
• See, for example, Leo Blanken, Jason Lepore, and Stephen Rodriguez, “America's Military Is Choking on Old Technology,” Foreign Policy , January 29, 2018, https://foreignpolicy.com/2018/01/29/americas-military-is-choking-on-old-technology . ↩
• In nominal RMB terms. Source: Nan Tian, Aude Fleurant, Alexandra Kuimova, Peter D. Wezeman, and Siemon T. Wezeman, “Trends in World Military Expenditure, 2017,” SIPRI, May 2018, https://www.sipri.org/publications/2018/sipri-fact-sheets/trends-world-military-expenditure-2017 . ↩
• Stephen Chen, “China Developing Robotic Subs to Launch New Era of Sea Power,” South China Morning Post , July 23, 2018, https://www.scmp.com/news/china/society/article/2156361/china-developing-unmanned-ai-submarines-launch-new-era-sea-power . ↩
• James Eng, “Woods Hole Oceanographic Institution Says Hack Linked to China,” NBCNews.com, October 16, 2016, https://www.nbcnews.com/tech/security/woods-hole-oceanographic-institution-says-hack-linked-china-n446226 . ↩
• China Institute for Science and Technology Policy at Tsinghua University, “China AI Development Report 2018,” Tsinghua University, July 2018, http://www.sppm.tsinghua.edu.cn/eWebEditor/UploadFile/China_AI_development_report_2018.pdf . ↩
• Skylogic Research, “2018 Drone Market Sector Report,” January 2019, https://droneanalyst.com/research/research-studies/2018-drone-market-sector-report-purchase . ↩
• Of course, this is also true of many “U.S.” AI research achievements, though to a lesser extent. ↩
• An English translation of the Chinese Electronics Standards Institute paper is available at Jeffrey Ding and Paul Triolo (translators), “Translation: Excerpts from China's ‘White Paper on Artificial Intelligence Standardization.’” New America, June 20, 2018, https://www.newamerica.org/cybersecurity-initiative/digichina/blog/translation-excerpts-chinas-white-paper-artificial-intelligence-standardization/ . ↩
• China Institute for Science and Technology Policy, “China AI Development Report 2018.” ↩
• Kai-fu Lee and Matt Sheehan, “China's Rise in Artificial Intelligence: Ingredients and Economic Implications,” Hoover Institution, October 29, 2018, https://www.hoover.org/research/chinas-rise-artificial-intelligence-ingredients-and-economic-implications . ↩
• Elsa Kania, "China's AI Talent 'arms Race,'" The Strategist. April 22, 2018. https://www.aspistrategist.org.au/chinas-ai-talent-arms-race/ . ↩
• Mos Zhang, “China Puts Education Focus on AI; Plans 50 AI Research Centres By 2020,” Synced , April 10, 2018, https://syncedreview.com/2018/04/10/china-puts-education-focus-on-ai-plans-50-ai-research-centres-by-2020/ . ↩
• Andy Chun, "China's AI Dream Is Well on Its Way to Becoming a Reality," South China Morning Post . April 22, 2018. https://www.scmp.com/comment/insight-opinion/article/2142641/chinas-ai-dream-well-its-way-becoming-reality . ↩
• Gregory K. Leonard and Mario A. Lopez, “Determining RAND Royalty Rates for Standard-Essential Patents,” Antitrust , 29 no. 1, 86–94, https://www.edgewortheconomics.com/files/documents/Determining_RAND_Royalty_Rates_for_Standard-Essential_Patents.pdf . ↩
• Tim Pohlmann, “Who Is Leading the 5G Patent Race?” Lexology , December 12, 2018, https://www.lexology.com/library/detail.aspx?g=64ea84d0-f9ce-4c2b-939b-dec5c2560e06 . ↩
• China Academy for Information and Communications Technology (CAICT) & China Institute of Information and Communications Security. “Artificial Intelligence and Security” September 2018. http://www.caict.ac.cn/kxyj/qwfb/bps/201809/P020180918473525332978.pdf . ↩
• Shanshan Du, “China Integrated Circuit Ecosystem Report,” SEMI Industry Research and Statistics October 2018, page 5, http://www1.semi.org/en/china-ic-ecosystem-report . ↩
• Jason Dedrick and Kenneth L. Kraemer, “Intangible assets and value capture in global value chains: the smartphone industry,” World Intellectual Property Organization Working Paper, November 2017, https://www.wipo.int/edocs/pubdocs/en/wipo_pub_econstat_wp_41.pdf . ↩
• Ju-min Park and Heekyong Yang, “Samsung to Shut Mobile Phone Plant in China's Tianjin” Reuters, December 12, 2018, https://www.reuters.com/article/us-samsung-elec-smartphones-china/samsung-to-shut-mobile-phone-plant-in-chinas-tianjin-idUSKBN1OB0YP . ↩
• Mai Nguyuen and Jessica Macy Yu, “Apple Assembler Foxconn considering IPhone Factory in Vietnam. . .” Reuters, December 5, 2018, https://www.reuters.com/article/us-foxconn-iphone-vietnam/apple-assembler-foxconn-considering-iphone-factory-in-vietnam-state-media-idUSKBN1O3128 . ↩
• Shanshan Du, “China Integrated Circuit Ecosystem Report.” ↩
• Sidney Leng, “China's Once-booming Textile and Clothing Industry Faces Tough times” South China Morning Post , April 30, 2018, https://www.scmp.com/news/china/economy/article/2143938/chinas-once-booming-textile-and-clothing-industry-faces-tough . ↩
• He Huifeng and Celia Chen, “A Robot Revolution Is Under Way at the ‘World’s Factory.’ Here’s Why,” South China Morning Post , October 25, 2018, https://www.scmp.com/economy/china-economy/article/2164103/made-china-2025-peek-robot-revolution-under-way-hub-worlds . ↩
• Aruna Viswanatha, Eva Dou, and Kate O’Keeffe, “ZTE to Pay $892 Million to U.S., Plead Guilty in Iran Sanctions Probe” The Wall Street Journal , March 8, 2017, https://www.wsj.com/articles/zte-to-pay-892-million-to-u-s-plead-guilty-in-iran-sanctions-probe-1488902019 . ↩
• See for example, the declassified 1986 CIA report “Soviet Microelectronics: Impact of Western Technology Acquisitions,” available at https://www.cia.gov/library/readingroom/docs/DOC_0000499603.pdf . ↩
• “PRC State-Owned Company, Taiwan Company, and Three Individuals Charged With Economic Espionage,” Department of Justice, press release, November, 1 2018, https://www.justice.gov/opa/pr/prc-state-owned-company-taiwan-company-and-three-individuals-charged-economic-espionage . ↩
• Chi Ling Chan, “Fallen Behind: Science, Technology, and Soviet Statism,” Intersect: The Stanford Journal of Science, Technology and Society , vol. 8, no. 3, (July 2015). ↩
• English translation of Tan Tieniu’s speech is available via Cameron Hickert and Jeffrey Ding (translators), “Read What Top Chinese Officials Are Hearing About AI Competition and Policy” New America , November 29, 2018, https://www.newamerica.org/cybersecurity-initiative/digichina/blog/read-what-top-chinese-officials-are-hearing-about-ai-competition-and-policy/. https://www.newamerica.org/cybersecurity-initiative/digichina/blog/read-what-top-chinese-officials-are-hearing-about-ai-competition-and-policy/ . ↩
• Specifically, the report claimed, “Made in China 2025 is at the center of the China AI policy citation network and has served as a programmatic document for local governments’ AI policymaking as they respond to the national AI development strategy.” ↩
• Kenneth Kraemer, Greg Linden, and Jason Dedrick, “Capturing Value in Global Networks: Apple's iPad and iPhone,” ResearchGate, July 2011, www.researchgate.net/publication/265187229_Capturing_Value_in_Global_Networks_Apple's_iPad_and_iPhone . ↩
• Value capture is defined in the report as follows: “Within a value chain, each producer purchases inputs and then adds value, which then becomes part of the cost of the next stage of production. The sum of the value added by everyone in the chain equals the final product price.” ↩
• As of 2017, 9 percent of Apple’s iPhone suppliers were Chinese companies. Debby Wu and Cheng Ting-Fang, “How the IPhone Reshaped Asian Tech,” Nikkei Asian Review , December 20, 2017, https://asia.nikkei.com/Spotlight/Cover-Story/How-the-iPhone-reshaped-Asian-tech2 .However, these companies and Chinese laborers at international companies still only captured 3 percent of the value of each iPhone. See Jason Dedrick and Kenneth L. Kraemer, “Intangible assets and value capture in global value chains: the smartphone industry,” World Intellectual Property Organization Working Paper, November 2017, https://www.wipo.int/edocs/pubdocs/en/wipo_pub_econstat_wp_41.pdf . ↩
• Ibid. ↩
• Though HiSilicon led the design effort, it licensed important intellectual property from international companies such as ARM. Additionally, Chinese AI chip startup Cambricon reportedly helped with the design of the deep learning accelerator element. ↩
• The production of semiconductor manufacturing equipment and semiconductor design software are two other critical areas. ↩
• Edward White, “China’s Ability to Make Computer Chips Still ‘Years Behind’ Industry Leaders,” Financial Times , January 22, 2019, https://www.ft.com/content/a002a9e4-1a42-11e9-b93e-f4351a53f1c3 . ↩
• China has no companies capable of producing the equipment required to manufacture at 7nm and other advanced process nodes. The top global equipment manufacturers are all based in the United States, Japan, South Korea, and Europe. ↩
• To my knowledge, Tim Hwang, Director, Harvard-MIT Ethics and Governance of AI Initiative and a member of the CNAS AI Task Force, is the first person to make this claim in publication.Tim Hwang, “Computational Power and the Social Impact of Artificial Intelligence,” March 24, 2018, updated January 2019, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3147971 . ↩
• The difference between 2015’s AlphaGo – which was trained in part upon a data corpus of historical human vs. human go matches – and 2017’s vastly superior AlphaGo Zero – which was trained entirely upon synthetic data generated from matches in which the AI played against itself – illustrates this issue. Synthetic data is not viable for all AI applications, since not all simulators are a perfect proxy for the real world they attempt to model. Nevertheless, synthetic data has proven to be increasingly important in cutting edge AI research and marketable AI applications. For example, U.S. self-driving car company Waymo (formerly Google) announced that in one year cars had driven 2.5 billion miles in virtual simulators compared with only 3 million miles of real-world roads. In 2017, one of the company’s executives said that “The vast majority of work done – new feature work – is motivated by stuff seen in simulation.” Alexis C. Madrigal, “Waymo Built a Secret World for Self-Driving Cars,” The Atlantic , August 23, 2017, https://www.theatlantic.com/technology/archive/2017/08/inside-waymos-secret-testing-and-simulation-facilities/537648/ . ↩
• Tim Hwang, “Computational Power and the Social Impact of Artificial Intelligence.” ↩
• Dario Amodei and Danny Hernandez, “AI and Compute,” OpenAI Blog, June 20, 2018, https://blog.openai.com/ai-and-compute/ . ↩
• In June 2018, Oak Ridge announced that its Summit supercomputer had achieved 122 petaflops in the Linpack benchmark test. SenseTime’s aggregate computer network is not capable of utilizing all of its computing power to work simultaneously on a single software problem such as Linpack, so this is not an apples to apples comparison, though it remains informative. “ORNL’s Summit Supercomputer Named World’s Fastest,” Oak Ridge National Laboratory, June 25, 2018, https://www.ornl.gov/news/ornl-s-summit-supercomputer-named-world-s-fastest. ↩
• “Hitting the Accelerator: The Next Generation of Machine-learning Chips,” Deloitte Global, 2017, https://www2.deloitte.com/content/dam/Deloitte/global/Images/infographics/technologymediatelecommunications/gx-deloitte-tmt-2018-nextgen-machine-learning-report.pdf ↩
• Stephanie Condon, “TPU Is 15x to 30x Faster than GPUs and CPUs, Google Says,” ZDNet, April 5, 2017, https://www.zdnet.com/article/tpu-is-15x-to-30x-faster-than-gpus-and-cpus-google-says/ . ↩
• Andy Patrizio, “Baidu Takes a Major Leap as an AI Player with New Chip, Intel Alliance,” Network World , July 11, 2018, https://www.networkworld.com/article/3289387/data-center/baidu-takes-a-major-leap-as-an-ai-player-with-new-chip-intel-alliance.html . ↩
• Cate Cadell and Adam Jourdan, “China's Alibaba Doubles Down on Chips amid Cloud Computing Push,” Reuters, September 19, 2018, https://uk.reuters.com/article/us-china-alibaba/chinas-alibaba-doubles-down-on-chips-amid-cloud-computing-push-idUKKCN1LZ0PR . ↩
• Jill Shen, “Briefing: AI Startup Horizon Robotics to Raise $600 Million in Series B Funding,” TechNode, January 15, 2019, https://technode.com/2019/01/15/ai-startup-horizon-robotics-funding/ . ↩
• Jianbin Gao and Conall Dempsey, “China’s Semiconductor Industry,” PwC, January 2018, https://www.pwc.com/gx/en/industries/technology/chinas-impact-on-semiconductor-industry/chinas-semiconductor-industry.html . ↩
• Dieter Ernst, “From Catching Up to Forging Ahead: China’s Policies for Semiconductors,” East-West Institute, 2015, https://www.ssrn.com/abstract=2744974 . ↩
• “TSMC Continues to Dominate the Worldwide Foundry Market,” IC Insights, April 24, 2018, http://www.icinsights.com/news/bulletins/tsmc-continues-to-dominate-the-worldwide-foundry-market/ . ↩
• “Quarterly Financial Results,” Taiwan Semiconductor Manufacturing Company, January 2019, https://www.tsmc.com/english/investorRelations/quarterly_results.htm . ↩
• For a helpful overview of how AI chips are more specialized than GPUs for machine learning, see Kaz Sato, “What Makes TPUs Fine-tuned for Deep Learning?” Google Cloud blog,August 30, 2018, https://cloud.google.com/blog/products/ai-machine-learning/what-makes-tpus-fine-tuned-for-deep-learning . ↩
• Saidong (塞冬), Jeffrey Ding (translator), “Breaking Open New Prospects for ‘China Chips,’” April 21, 2018, https://docs.google.com/document/d/1uqwII-c8shKGlQSy2aqKs8EIT9Fa_322xr_up6zKRzQ/edit . Original Mandarin: https://mp.weixin.qq.com/s/YDPSA5BdPPzpaMnQ65XjtA . ↩
• You Zheng and Wei Shaojun, eds., “White Paper on AI Chip Technologies,” Beijing Innovation Center for Future Chips, Tsinghua University, 2018, https://www.080910t.com/downloads/AI%20Chip%202018%20EN.pdf . ↩
• MOST is responsible for coordinating all Chinese government agencies to implement the AIDP. ↩
• Paul Triolo and Jimmy Goodrich, “As China’s Government Mobilizes for AI Leadership, Some Challenges Will Be Tougher Than Others,” New America , February 28, 2018, https://www.newamerica.org/cybersecurity-initiative/digichina/blog/riding-wave-full-steam-ahead/ . ↩
• “Guideline for the Promotion of the Development of the National Integrated Circuit Industry,” China State Council, 2014, https://members.wto.org/CRNAttachments/2014/SCMQ2/law47.pdf . ↩
• Yuki Nakamura and Yuki Furukawa, “Jack Ma Says Nations Need Tech to Sidestep U.S. Grip,” Bloomberg.com, April 25, 2018, https://www.bloomberg.com/news/articles/2018-04-25/jack-ma-says-nations-need-own-tech-to-sidestep-u-s-control . ↩
• Cate Cadell, “Chips Down: China Aims to Boost Semiconductors as Trade War Looms,” Reuters, April 20, 2018, https://www.reuters.com/article/us-usa-trade-china-chips/chips-down-china-aims-to-boost-semiconductors-as-trade-war-looms-idUSKBN1HR1DF . ↩
• Michael Hall, “$62.7 Billion Semiconductor Equipment Forecast – Tops Previous Record, Korea at Top but China Closes the Gap,” SEMI, July 9, 2018, https://globenewswire.com/news-release/2018/07/09/1534721/0/en/62-7-Billion-Semiconductor-Equipment-Forecast-Tops-Previous-Record-Korea-at-Top-but-China-Closes-the-Gap.html . ↩
• Cheng Ting-Fang, “Chinese Chipmaker Takes on TSMC and Intel with Cutting-edge Tool,” Nikkei Asian Review , May 15, 2018, https://asia.nikkei.com/Business/Companies/Chinese-chip-maker-invests-in-next-gen-tool-to-close-gaps-with-Intel-TSMC-Samsung . ↩
• See Ming-Chin Monique Chu, “Controlling the Uncontrollable: The Migration of the Taiwanese Semiconductor Industry to China and Its Security Ramifications” China Perspectives , no. 1 (2008), 54–68, https://journals.openedition.org/chinaperspectives/3343 . ↩
• Alan Patterson, “Leaker of TSMC Secrets Joins SMIC as Co-CEO,” EETimes , October 17, 2017, https://www.eetimes.com/document.asp?doc_id=1332462 . ↩
• Sarah Dai, “Funding Squeeze Seen Coming for Nine in 10 AI Start-ups in China,” South China Morning Post , August 26, 2018, https://www.scmp.com/tech/article/2161387/investor-warns-day-reckoning-90-pc-chinese-ai-start-ups-funding-dries . ↩
• “Internet industry encounters cold wave: Beijing office rents fell in the fourth quarter,” Caixin, December 26, 2018, http://companies.caixin.com/2018-12-26/101363762.html . ↩
• The relevant passage states: “Any organization and citizen shall, in accordance with the law, support, provide assistance, and cooperate in national intelligence work, and guard the secrecy of any national intelligence work that they are aware of. The state shall protect individuals and organizations that support, cooperate with, and collaborate in national intelligence work.” ↩
• Yoko Kubota and Liza Lin, “A Chill From Beijing Buffets China’s Tech Sector,” The Wall Street Journal , December 27, 2018, https://www.wsj.com/articles/a-chill-from-beijing-buffets-chinas-tech-sector-11545920268 . ↩

Gregory C. Allen

Former Adjunct Senior Fellow, Technology and National Security Program

Gregory C. Allen is a former Adjunct Senior Fellow at the Center for a New American Security (CNAS) Technology and National Security Program. Mr. Allen focuses on the intersec...

More from CNAS

• September 11, 2024

From the battlefields of Libya to Nagorno-Karabakh to Ukraine, the deployment of drones has become a critical element of modern warfare. Will the explosion of unmanned aerial ...

By Anna Pederson & Molly Campbell

• September 5, 2024

The PRC today operates the largest and most sophisticated propaganda apparatus in human history....

By Bill Drexel & Grace Gao

• The MIT Technology Review
• September 3, 2024

Far from being dominated by AI-enabled catastrophes, this election “super year” at that point was pretty much like every other election year....

By Keegan McBride, Felix M. Simon & Sacha Altay

• Tech Policy Press
• August 21, 2024

A strong NIST will continue to help build standards that are adopted globally and lay the foundation for further American AI innovation and dissemination....

By Keegan McBride

View All Reports View All Articles & Multimedia

Get the Latest from CNAS

Sign up for weekly updates and analysis on the most important issues in u.s. national security..

Thanks for subscribing!

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

There was a problem submitting the form. Please double check your information and try again.

• Work & Careers
• Life & Arts

China’s AI ambitions revealed by most cited research papers

Limited time offer, save 50% on standard digital, explore more offers..

Then $75 per month. Complete digital access to quality FT journalism. Cancel anytime during your trial.

Premium Digital

Complete digital access to quality FT journalism with expert analysis from industry leaders. Pay a year upfront and save 20%.

• Global news & analysis
• Expert opinion
• FT App on Android & iOS
• FT Edit app
• FirstFT: the day's biggest stories
• 20+ curated newsletters
• Follow topics & set alerts with myFT
• FT Videos & Podcasts
• 20 monthly gift articles to share
• Lex: FT's flagship investment column
• 15+ Premium newsletters by leading experts
• FT Digital Edition: our digitised print edition

FT Digital Edition

10% off your first year. The new FT Digital Edition: today’s FT, cover to cover on any device. This subscription does not include access to ft.com or the FT App.

Terms & Conditions apply

Explore our full range of subscriptions.

Why the ft.

See why over a million readers pay to read the Financial Times.

Can democracies cooperate with China on AI research?

• Download the full report

Subscribe to Governance Weekly

Rebalancing ai research networks, cameron f. kerry , cameron f. kerry ann r. and andrew h. tisch distinguished visiting fellow - governance studies , center for technology innovation joshua p. meltzer , and joshua p. meltzer senior fellow - global economy and development matt sheehan matt sheehan fellow - carnegie endowment for international peace.

January 9, 2023

China looms large in the global landscape of artificial intelligence (AI) research, development, and policymaking. Its talent, growing technological skill and innovation, and national investment in science and technology have made it a leader in AI.

Over more than two decades, China has become deeply enmeshed in the international network of AI research and development (R&D): co-authoring papers with peers abroad, hosting American corporate AI labs, and helping expand the frontiers of global AI research. During most of that period, these links and their implications went largely unexamined in the policy world. Instead, the nature of these connections was dictated by the researchers, universities, and corporations who were forging them.

But in the past five years, these ties between China and global networks for R&D have come under increasing scrutiny by governments as well as universities, companies, and civil society. Four factors worked together to drive this reassessment: (1) the growing capabilities of AI itself and its impacts on both economic competitiveness and national security; (2) China’s unethical use of AI, including its deployment of AI tools for mass surveillance of its citizens, most notably the Uyghur ethnic group in Xinjiang but increasingly more widespread; (3) the rise in Chinese capabilities and ambitions in AI, making it a genuine competitor with the U.S. in the field; and (4) the policies by which the Chinese state bolstered those capabilities, including state directed investments and illicit knowledge transfers from abroad.

Taken together, these concerns led to intense scrutiny and new questions about these long-standing ties. Is cooperation helping China overtake democratic nations in AI? To what extent are technologists and companies in democratic nations contributing to China’s deployment of repressive AI tools?

This working paper considers whether and to what extent international collaboration with China on AI can endure. China has been a subject of discussions among the government officials and experts participating in the Forum for Cooperation on AI (FCAI) over the past two years. The 2021 FCAI progress report identified the implications of China’s development and use of AI for international cooperation. 1 The report touched on China in connection with several of the recommendations regarding regulatory alignment, standards development, trade agreements, and R&D projects but also focused on Chinese policies and applications of AI that present a range of challenges in the context of that nation’s broader geopolitical, economic, and authoritarian policies. A roundtable discussion on December 8, 2021 presented these issues to FCAI participants more fully and elicited their views.

This paper expands and distills this work with a focus on the scope, benefits, and prospective limits of China’s involvement in international AI R&D networks. In Part I, it presents the history of China’s AI development and extraordinarily successful engagement with international R&D and explains how this history has helped China become a global leader in the field. Part II shows how China has become embedded in international AI R&D networks, with China and the United States becoming each other’s largest collaborator and China also a major collaborator with each of the other six countries participating in FCAI. This collaboration takes place through multiple pathways: enrollment at universities, conferences, joint publications, and work in research labs that all operate in various ways to develop, disseminate, and deploy AI.

Part III then provides an overview of the economic, ethical, and strategic issues that call into question whether such levels of collaboration on AI can continue, as well as the challenges and disadvantages of disconnecting the channels of collaboration. The analysis then looks at how engagement with China on AI R&D might evolve. It does so primarily through a U.S.-focused lens because the U.S., as by far China’s largest competitor and collaborator in AI, provides an umbrella and a template for countries and FCAI participants that also collaborate with China on AI R&D and face many of the same issues. Moreover, measures to respond to the challenges China presents are more likely to be effective in coordination than in isolation. Recent U.S. export controls on semiconductors and the technologies used to manufacture them have laid bare the critical role of countries such as Japan and Korea. For now, the U.S. government is able to force foreign compliance through administrative measures, such as the foreign direct product rule, but these mechanisms may be made moot if foreign manufacturers engineer U.S. technology out of their supply chain. This paper deals with cooperative research rather than hardware supply chains, but similar dynamics exist across these domains. Accordingly, this paper is not just about collaboration with China but also about collaboration in relation to China.

Measures to respond to the challenges China presents are more likely to be effective in coordination than in isolation.

The U.S., other governments participating in FCAI, and their partners are not the only actors in this drama. What AI R&D with China looks like going forward will also be determined by what China does. China’s intensifying push for technological self-reliance has accelerated China’s disengagement from the international technology ecosystem in certain respects, while so far keeping it deeply enmeshed in other international research networks. The future trajectory of this engagement will depend heavily on actions taken by the Chinese government and the Chinese Communist Party.

In light of the issues presented by these changes, the paper proposes rebalancing AI R&D with Chinese researchers and institutions through a risk-based approach. Going forward, such collaboration will require a clear assessment of the costs and benefits, aiming to maximize the benefits of an open research environment and strong international links with the risks presented by AI R&D with China. Adopting an appropriately risk-based approach often will not counsel complete disengagement with China on AI R&D and instead require a rebalancing that takes into account the various vectors for knowledge transfer. Crucially, governments need to work collaboratively with each other and with companies, universities, and research labs to inform the assessment of the risks and understand the benefits of AI R&D with China. A failure to build these partnerships into the risk-assessment process could lead to bad outcomes that mismeasure risks and benefits, leaving the U.S. worse off.

Download the full report.

Related Content

Cameron F. Kerry, Joshua P. Meltzer, Andrea Renda

November 4, 2022

Cameron F. Kerry, Joshua P. Meltzer, Andrea Renda, Alex Engler, Rosanna Fanni

October 25, 2021

October 3, 2022

• Cameron F Kerry et al., “Strengthening International Cooperation on AI,” n.d., 123.

Artificial Intelligence

Global Economy and Development Governance Studies

The Forum for Cooperation on Artificial Intelligence

The Brookings Institution, Washington DC

10:00 am - 11:00 am EDT

Darrell M. West, Nicol Turner Lee

August 19, 2024

Darrell M. West, Roxana Muenster

August 5, 2024

• Topics ›
• Artificial intelligence (AI) worldwide ›

China's Race Towards AI Research Dominance

Artificial intelligence.

Since taking its first steps teaching computers board game strategies in the 1950s, research on artificial intelligence has come a long way. In the 21st century in particular, machine learning and its promise for real-time improvements of algorithms through experience and providing access to more data has become the single biggest research focus in the field. As our chart based on data provided by the OECD.AI project shows, China is well on its way to surpassing traditional artificial intelligence research powerhouses in the upcoming years.

While the U.S. still leads the world with about 150.000 research papers on AI published in 2021, the People's Republic's output isn't that far off thanks to an astronomical increase over the last two decades. The Eastern Asian country passed the number of AI research papers published in every single one of the 27 EU countries combined in 2008 and as of now sits in second place with roughly 138.000 papers pushed to publication in 2021. Overall, it increased its research output by 3,350 percent over the last two decades.

Even though AI research has led to improvements in terms of productivity in almost every sector imaginable, it's not without its downsides if left unchecked. For example, a Gizmodo investigation published in December 2021 revealed that PredPol, a predictive policing software based on AI, allegedly reproduced bias instead of giving neutral judgments due to the biased nature of the data it was trained on, mostly leaving predominantly white neighborhoods out of its equations.

Description

This chart shows the countries and regions with highest number of AI research papers submitted per year.

Can I integrate infographics into my blog or website?

Yes, Statista allows the easy integration of many infographics on other websites. Simply copy the HTML code that is shown for the relevant statistic in order to integrate it. Our standard is 660 pixels, but you can customize how the statistic is displayed to suit your site by setting the width and the display size. Please note that the code must be integrated into the HTML code (not only the text) for WordPress pages and other CMS sites.

Infographic Newsletter

Statista offers daily infographics about trending topics, covering: Economy & Finance , Politics & Society , Tech & Media , Health & Environment , Consumer , Sports and many more.

Related Infographics

Antitrust in the technology industry, what does the u.s. think about tech regulation, stock market, summer of 2024 has been a rollercoaster ride for investors, research & developemt spending, the most emission-intensive industries rarely focus on r&d, who’s got a samsung phone, semiconductor manufacturing, how the ai boom influenced tsmc’s bottom line, semiconductor industry, who leads the semiconductor foundry market, travel inspo: insta & tiktok beat ai, ai boom: nvidia becomes most valuable company in the world, online news, social media is the main gateway to online news, americans and their passwords: it's complicated, generative ai, generative ai: a jack of all trades, how widespread is chatgpt usage.

• Who may use the "Chart of the Day"? The Statista "Chart of the Day", made available under the Creative Commons License CC BY-ND 3.0, may be used and displayed without charge by all commercial and non-commercial websites. Use is, however, only permitted with proper attribution to Statista. When publishing one of these graphics, please include a backlink to the respective infographic URL. More Information
• Which topics are covered by the "Chart of the Day"? The Statista "Chart of the Day" currently focuses on two sectors: "Media and Technology", updated daily and featuring the latest statistics from the media, internet, telecommunications and consumer electronics industries; and "Economy and Society", which current data from the United States and around the world relating to economic and political issues as well as sports and entertainment.
• Does Statista also create infographics in a customized design? For individual content and infographics in your Corporate Design, please visit our agency website www.statista.design

Any more questions?

Get in touch with us quickly and easily. we are happy to help.

Feel free to contact us anytime using our contact form or visit our FAQ page .

Statista Content & Design

Need infographics, animated videos, presentations, data research or social media charts?

More Information

The Statista Infographic Newsletter

Receive a new up-to-date issue every day for free.

• Our infographics team prepares current information in a clear and understandable format
• Relevant facts covering media, economy, e-commerce, and FMCG topics
• Use our newsletter overview to manage the topics that you have subscribed to

Try AI-powered search

• China has become a scientific superpower

From plant biology to superconductor physics the country is at the cutting edge

Your browser does not support the <audio> element.

I n the atrium of a research building at the Chinese Academy of Sciences ( CAS ) in Beijing is a wall of patents. Around five metres wide and two storeys high, the wall displays 192 certificates, positioned in neat rows and tastefully lit from behind. At ground level, behind a velvet rope, an array of glass jars contain the innovations that the patents protect: seeds.

CAS —the world’s largest research organisation—and institutions around China produce a huge amount of research into the biology of food crops. In the past few years Chinese scientists have discovered a gene that, when removed, boosts the length and weight of wheat grains, another that improves the ability of crops like sorghum and millet to grow in salty soils and one that can increase the yield of maize by around 10%. In autumn last year, farmers in Guizhou completed the second harvest of genetically modified giant rice that was developed by scientists at CAS .

The Chinese Communist Party ( CCP ) has made agricultural research—which it sees as key to ensuring the country’s food security —a priority for scientists. Over the past decade the quality and the quantity of crop research that China produces has grown immensely, and now the country is widely regarded as a leader in the field. According to an editor of a prestigious European plant-sciences journal, there are some months when half of the submissions can come from China.

A journey of a thousand miles

The rise of plant-science research is not unique in China. In 2019 The Economist surveyed the research landscape in the country and asked whether China could one day become a scientific superpower. Today, that question has been unequivocally answered: “yes”. Chinese scientists recently gained the edge in two closely watched measures of high-quality science, and the country’s growth in top-notch research shows no sign of slowing. The old science world order, dominated by America, Europe and Japan, is coming to an end.

One way to measure the quality of a country’s scientific research is to tally the number of high-impact papers produced each year—that is, publications that are cited most often by other scientists in their own, later work. In 2003 America produced 20 times more of these high-impact papers than China, according to data from Clarivate, a science analytics company (see chart 1). By 2013 America produced about four times the number of top papers and, in the most recent release of data, which examines papers from 2022, China had surpassed both America and the entire European Union ( EU ).

Metrics based on citations can be gamed, of course. Scientists can, and do, find ways to boost the number of times their paper is mentioned in other studies, and a recent working paper, by Qui Shumin, Claudia Steinwender and Pierre Azoulay, three economists, argues that Chinese researchers cite their compatriots far more than Western researchers do theirs. But China now leads the world on other benchmarks that are less prone to being gamed. It tops the Nature Index, created by the publisher of the same name, which counts the contributions to articles that appear in a set of prestigious journals. To be selected for publication, papers must be approved by a panel of peer reviewers who assess the study’s quality, novelty and potential for impact. When the index was first launched, in 2014, China came second, but its contribution to eligible papers was less than a third of America’s. By 2023 China had reached the top spot.

According to the Leiden Ranking of the volume of scientific research output, there are now six Chinese universities or institutions in the world top ten, and seven according to the Nature Index. They may not be household names in the West yet, but get used to hearing about Shanghai Jiao Tong, Zhejiang and Peking (Beida) Universities in the same breath as Cambridge, Harvard and ETH Zurich. “Tsinghua is now the number one science and technology university in the world,” says Simon Marginson, a professor of higher education at Oxford University. “That’s amazing. They’ve done that in a generation.”

Today China leads the world in the physical sciences, chemistry and Earth and environmental sciences, according to both the Nature Index and citation measures (see chart 2). But America and Europe still have substantial leads in both general biology and medical sciences. “Engineering is the ultimate Chinese discipline in the modern period,” says Professor Marginson, “I think that’s partly about military technology and partly because that’s what you need to develop a nation.”

Applied research is a Chinese strength. The country dominates publications on perovskite solar panels, for example, which offer the possibility of being far more efficient than conventional silicon cells at converting sunlight into electricity. Chinese chemists have developed a new way to extract hydrogen from seawater using a specialised membrane to separate out pure water, which can then be split by electrolysis. In May 2023 it was announced that the scientists, in collaboration with a state-owned Chinese energy company, had developed a pilot floating hydrogen farm off the country’s south-eastern coast.

China also now produces more patents than any other country, although many are for incremental tweaks to designs, as opposed to truly original inventions. New developments tend to spread and be adopted more slowly in China than in the West. But its strong industrial base, combined with cheap energy, means that it can quickly spin up large-scale production of physical innovations like materials. “That’s where China really has an advantage on Western countries,” says Jonathan Bean, CEO of Materials Nexus, a British firm that uses AI to discover new materials.

The country is also signalling its scientific prowess in more conspicuous ways. Earlier this month, China’s Chang’e-6 robotic spacecraft touched down in a gigantic crater on the far side of the Moon, scooped up some samples of rock, planted a Chinese flag and set off back towards Earth. If it successfully returns to Earth at the end of the month, it will be the first mission to bring back samples from this hard-to-reach side of the Moon.

First, sharpen your tools

The reshaping of Chinese science has been achieved by focusing on three areas: money, equipment and people. In real terms, China’s spending on research and development ( R & D ) has grown 16-fold since 2000. According to the most recent data from the OECD , from 2021, China still lagged behind America on overall R & D spending, dishing out $668bn, compared with $806bn for America at purchasing-power parity. But in terms of spending by universities and government institutions only, China has nudged ahead. In these places America still spends around 50% more on basic research, accounting for costs, but China is splashing the cash on applied research and experimental development (see chart 3).

Money is meticulously directed into strategic areas. In 2006 the CCP published its vision for how science should develop over the next 15 years. Blueprints for science have since been included in the CCP ’s five-year development plans. The current plan, published in 2021, aims to boost research in quantum technologies, AI , semiconductors, neuroscience, genetics and biotechnology, regenerative medicine, and exploration of “frontier areas” like deep space, deep oceans and Earth’s poles.

Creating world-class universities and government institutions has also been a part of China’s scientific development plan. Initiatives like “Project 211”, the “985 programme” and the “China Nine League” gave money to selected labs to develop their research capabilities. Universities paid staff bonuses—estimated at an average of $44,000 each, and up to a whopping $165,000—if they published in high-impact international journals.

Building the workforce has been a priority. Between 2000 and 2019, more than 6m Chinese students left the country to study abroad, according to China’s education ministry. In recent years they have flooded back, bringing their newly acquired skills and knowledge with them. Data from the OECD suggest that, since the late 2000s, more scientists have been returning to the country than leaving. China now employs more researchers than both America and the entire EU .

Many of China’s returning scientists, often referred to as “sea turtles” (a play on the Chinese homonym haigui , meaning “to return from abroad”) have been drawn home by incentives. One such programme launched in 2010, the “Youth Thousand Talents”, offered researchers under 40 one-off bonuses of up to 500,000 yuan (equivalent to roughly $150,000 at purchasing-power parity) and grants of up to 3m yuan to get labs up and running back home. And it worked. A study published in Science last year found that the scheme brought back high-calibre young researchers—they were, on average, in the most productive 15% of their peers (although the real superstar class tended to turn down offers). Within a few years, thanks to access to more resources and academic manpower, these returnees were lead scientists on 2.5 times more papers than equivalent researchers who had remained in America.

As well as pull, there has been a degree of push. Chinese scientists working abroad have been subject to increased suspicion in recent years. In 2018 America launched the China Initiative, a largely unsuccessful attempt to root out Chinese spies from industry and academia. There have also been reports of students being deported because of their association with China’s “military-civilian fusion strategy”. A recent survey of current and former Chinese students studying in America found that the share who had experienced racial abuse or discrimination was rising.

The availability of scientists in China means that, for example in quantum computing, some of the country’s academic labs are more like commercial labs in the West, in terms of scale. “They have research teams of 20, 30, even 40 people working on the same experiments, and they make really good progress,” says Christian Andersen, a quantum researcher at Delft University. In 2023 researchers working in China broke the record for the number of quantum bits, or qubits, entangled inside a quantum computer.

China has also splurged on scientific kit. In 2019, when The Economist last surveyed the state of the country’s scientific research, it already had an enviable inventory of flashy hardware including supercomputers, the world’s largest filled-aperture radio telescope and an underground dark-matter detector. The list has only grown since then. The country is now home to the world’s most sensitive ultra-high-energy cosmic-ray detector (which has recently been used to test aspects of Albert Einstein’s special theory of relativity), the world’s strongest steady-state magnetic field (which can probe the properties of materials) and soon will have one of the world’s most sensitive neutrino detectors (which will be used to work out which type of these fundamental subatomic particles has the highest mass). Europe and America have plenty of cool kit of their own, but China is rapidly adding hardware.

Individual labs in China’s top institutions are also well equipped. Niko McCarty, a journalist and former researcher at the Massachusetts Institute of Technology who was recently given a tour of synthetic biology labs in China, was struck by how, in academic institutions, “the machines are just more impressive and more expansive” than in America. At the Advanced Biofoundry at the Shenzhen Institute of Advanced Technology, which the country hopes will be the centre of China’s answer to Silicon Valley, Mr McCarty described an “amazing building with four floors of robots”. As Chinese universities fill with state-of-the-art equipment and elite researchers, and salaries become increasingly competitive, Western institutions look less appealing to young and ambitious Chinese scientists. “Students in China don’t think about America as some “scientific Mecca” in the same way their advisers might have done,” said Mr McCarty.

Take AI , for example. In 2019 just 34% of Chinese students working in the field stayed in the country for graduate school or work. By 2022 that number was 58%, according to data from the AI talent tracker by MacroPolo, an American think-tank (in America the figure for 2022 was around 98%). China now contributes to around 40% of the world’s research papers on AI , compared with around 10% for America and 15% for the EU and Britain combined. One of the most highly cited research papers of all time, demonstrating how deep neural networks could be trained on image recognition, was written by AI researchers working in China, albeit for Microsoft, an American company. “China’s AI research is world-class,” said Zachary Arnold, an AI analyst at the Georgetown Centre for Emerging Security and Technology. “In areas like computer vision and robotics, they have a significant lead in research publications.”

Growth in the quality and quantity of Chinese science looks unlikely to stop anytime soon. Spending on science and technology research is still increasing—the government has announced a 10% increase in funding in 2024. And the country is training an enormous number of young scientists. In 2020 Chinese universities awarded 1.4m engineering degrees, seven times more than America did. China has now educated, at undergraduate level, 2.5 times more of the top-tier AI researchers than America has. And by 2025, Chinese universities are expected to produce nearly twice as many P h D graduates in science and technology as America.

To see further, ascend another floor

Although China is producing more top-tier work, it still produces a vast amount of lower-quality science too. On average, papers from China tend to have lower impact, as measured by citations, than those from America, Britain or the EU . And while the chosen few universities have advanced, mid-level universities have been left behind. China’s second-tier institutions still produce work that is of relatively poor quality compared with their equivalents in Europe or America. “While China has fantastic quality at the top level, it’s on a weak base,” explains Caroline Wagner, professor of science policy at Ohio State University.

When it comes to basic, curiosity-driven research (rather than applied) China is still playing catch-up—the country publishes far fewer papers than America in the two most prestigious science journals, Nature and Science . This may partly explain why China seems to punch below its weight in the discovery of completely new technologies. Basic research is particularly scant within Chinese companies, creating a gap between the scientists making discoveries and the industries that could end up using them. “For more original innovation, that might be a minus,” says Xu Xixiang, chief scientist at LONG i Green Energy Technology, a Chinese solar company.

Incentives to publish papers have created a market for fake scientific publications. A study published earlier this year in the journal Research Ethics , featured anonymous interviews from Chinese academics, one of whom said he had “no choice but to commit [research] misconduct”, to keep up with pressures to publish and retain his job. “Citation cartels” have emerged, where groups of researchers band together to write low-quality papers that cite each other’s work in an effort to drive up their metrics. In 2020 China’s science agencies announced that such cash-for-publication schemes should end and, in 2021, the country announced a nationwide review of research misconduct. That has led to improvements—the rate at which Chinese researchers cite themselves, for example, is falling, according to research published in 2023. And China’s middle-ranking universities are slowly catching up with their Western equivalents, too.

The areas where America and Europe still hold the lead are, therefore, unlikely to be safe for long. Biological and health sciences rely more heavily on deep subject-specific knowledge and have historically been harder for China to “bring back and accelerate”, says Tim Dafforn, a professor of biotechnology at University of Birmingham and former adviser to Britain’s department for business. But China’s profile is growing in these fields. Although America currently produces roughly four times more highly influential papers in clinical medicine, in many areas China is producing the most papers that cite this core research, a sign of developing interest that presages future expansion. “On the biology side, China is growing remarkably quickly,” says Jonathan Adams, chief scientist at the Institute for Scientific Information at Clarivate. “Its ability to switch focus into a new area is quite remarkable.”

The rise of Chinese science is a double-edged sword for Western governments. China’s science system is inextricably linked with its state and armed forces—many Chinese universities have labs explicitly working on defence and several have been accused of engaging in espionage or cyber-attacks. China has also been accused of intellectual-property theft and increasingly stringent regulations have made it more difficult for international collaborators to take data out of the country; notoriously, in 2019, the country cut off access to American-funded work on coronaviruses at the Wuhan Institute of Virology. There are also cases of Chinese researchers failing to adhere to the ethical standards expected by Western scientists.

Despite the concerns, Chinese collaborations are common for Western researchers. Roughly a third of papers on telecommunications by American authors involve Chinese collaborators. In imaging science, remote sensing, applied chemistry and geological engineering, the figures are between 25% and 30%. In Europe the numbers are lower, around 10%, but still significant. These partnerships are beneficial for both countries. China tends to collaborate more in areas where it is already strong like materials and physics. A preprint study, released last year, found that for AI research, having a co-author from America or China was equally beneficial to authors from the other country, conferring on average 75% more citations.

Several notable successes have come from working together, too. During the covid-19 pandemic a joint venture between Oxford University’s Engineering Department and the Oxford Suzhou Centre for Advanced Research developed a rapid covid test that was used across British airports. In 2015 researchers at University of Cardiff and South China Agricultural University identified a gene that made bacteria resistant to the antibiotic colistin. Following this, China, the biggest consumer of the drug, banned its use in animal feed, and levels of colistin resistance in both animals and humans declined.

In America and Europe, political pressure is limiting collaborations with China. In March, America’s Science and Technology Agreement with China, which states that scientists from both countries can collaborate on topics of mutual benefit, was quietly renewed for a further six months. Although Beijing appears keen to renew the 45-year-old agreement, many Republicans fear that collaboration with China is helping the country achieve its national-security goals. In Europe, with the exception of environmental and climate projects, Chinese universities have been effectively barred from accessing funding through the Horizon programme, a huge European research initiative.

There are also concerns among scientists that China is turning inwards. The country has explicit aims to become self-reliant in many areas of science and technology and also shift away from international publications as a way of measuring research output. Many researchers cannot talk to the press—finding sources in China for this story was challenging. One Chinese plant scientist, who asked to remain anonymous, said that she had to seek permission a year in advance to attend overseas conferences. “It’s contradictory—on the one hand, they set restrictions so that scientists don’t have freedoms like being able to go abroad to communicate with their colleagues. But on the other hand, they don’t want China to fall behind.”

Live until old, learn until old

The overwhelming opinion of scientists in China and the West is that collaboration must continue or, better, increase. And there is room to do more. Though China’s science output has grown dramatically, the share that is conducted with international collaborators has remained stable at around 20%—Western scientists tend to have far more international collaborations. Western researchers could pay more attention to the newest science from China, too. Data from a study published last year in Nature Human Behaviour showed that, for work of equivalent quality, Chinese scientists cite Western papers far more than vice versa. Western scientists rarely visit, work or study in China, depriving them of opportunities to learn from Chinese colleagues in the way Chinese scientists have done so well in the West.

Closing the door to Chinese students and researchers wishing to come to Western labs would also be disastrous for Western science. Chinese researchers form the backbone of many departments in top American and European universities. In 2022 more of the top-tier AI researchers working in America hailed from China than from America. The West’s model of science currently depends on a huge number of students, often from overseas, to carry out most day-to-day research.

There is little to suggest that the Chinese scientific behemoth will not continue growing stronger. China’s ailing economy may eventually force the CCP to slow spending on research, and if the country were to become completely cut off from the Western science community its research would suffer. But neither of these looks imminent. In 2019 we also asked if research could flourish in an authoritarian system. Perhaps over time its limits will become clear. But for now, and at least for the hard sciences, the answer is that it can thrive. “I think it’d be very unwise to call limits on the Chinese miracle,” says Prof Marginson. “Because it has had no limits up until now.” ■

Curious about the world? To enjoy our mind-expanding science coverage, sign up to  Simply Science , our weekly subscriber-only newsletter.

Explore more

This article appeared in the Science & technology section of the print edition under the headline “Soaring dragons”

Science & technology June 15th 2024

From the June 15th 2024 edition

Discover stories from this section and more in the list of contents

More from Science & technology

Geothermal energy could outperform nuclear power

Tricks from the oil industry have produced a hot-rocks breakthrough

The world’s first nuclear clock is on the horizon

It would be 1000 times more accurate than today’s atomic timekeepers

Baby formulas now share some ingredients with breast milk

They may one day replicate its benefits

Breast milk’s benefits are not limited to babies

Some of its myriad components are being tested as treatments for cancer and other diseases

Particles that damage satellites can be flushed out of orbit

All it takes is very long radio waves

A common food dye can make skin transparent

The discovery allows scientists to see inside live animals

• Conferences
• Last Updated: September 13, 2024
• In AI Mysteries

Top Machine Learning Research Papers

• by Dr. Nivash Jeevanandam

Advances in machine learning and deep learning research are reshaping our technology. Machine learning and deep learning have accomplished various astounding feats, and key research articles have resulted in technical advances used by billions of people. The research in this sector is advancing at a breakneck pace and assisting you to keep up. Here is a collection of the most important scientific study papers in machine learning.

Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training

The authors of this work examined why ACGAN training becomes unstable as the number of classes in the dataset grows. The researchers revealed that the unstable training occurs due to a gradient explosion problem caused by the unboundedness of the input feature vectors and the classifier’s poor classification capabilities during the early training stage. The researchers presented the Data-to-Data Cross-Entropy loss (D2D-CE) and the Rebooted Auxiliary Classifier Generative Adversarial Network to alleviate the instability and reinforce ACGAN (ReACGAN). Additionally, extensive tests of ReACGAN demonstrate that it is resistant to hyperparameter selection and is compatible with a variety of architectures and differentiable augmentations.

This article is ranked #1 on CIFAR-10 for Conditional Image Generation.

For the research paper, read here .

For code, see here .

Dense Unsupervised Learning for Video Segmentation

The authors presented a straightforward and computationally fast unsupervised strategy for learning dense spacetime representations from unlabeled films in this study. The approach demonstrates rapid convergence of training and a high degree of data efficiency. Furthermore, the researchers obtain VOS accuracy superior to previous results despite employing a fraction of the previously necessary training data. The researchers acknowledge that the research findings may be utilised maliciously, such as for unlawful surveillance, and that they are excited to investigate how this skill might be used to better learn a broader spectrum of invariances by exploiting larger temporal windows in movies with complex (ego-)motion, which is more prone to disocclusions.

This study is ranked #1 on DAVIS 2017 for Unsupervised Video Object Segmentation (val).

Temporally-Consistent Surface Reconstruction using Metrically-Consistent Atlases

The authors offer an atlas-based technique for producing unsupervised temporally consistent surface reconstructions by requiring a point on the canonical shape representation to translate to metrically consistent 3D locations on the reconstructed surfaces. Finally, the researchers envisage a plethora of potential applications for the method. For example, by substituting an image-based loss for the Chamfer distance, one may apply the method to RGB video sequences, which the researchers feel will spur development in video-based 3D reconstruction.

This article is ranked #1 on ANIM in the category of Surface Reconstruction. 

EdgeFlow: Achieving Practical Interactive Segmentation with Edge-Guided Flow

The researchers propose a revolutionary interactive architecture called EdgeFlow that uses user interaction data without resorting to post-processing or iterative optimisation. The suggested technique achieves state-of-the-art performance on common benchmarks due to its coarse-to-fine network design. Additionally, the researchers create an effective interactive segmentation tool that enables the user to improve the segmentation result through flexible options incrementally.

This paper is ranked #1 on Interactive Segmentation on PASCAL VOC

Learning Transferable Visual Models From Natural Language Supervision

The authors of this work examined whether it is possible to transfer the success of task-agnostic web-scale pre-training in natural language processing to another domain. The findings indicate that adopting this formula resulted in the emergence of similar behaviours in the field of computer vision, and the authors examine the social ramifications of this line of research. CLIP models learn to accomplish a range of tasks during pre-training to optimise their training objective. Using natural language prompting, CLIP can then use this task learning to enable zero-shot transfer to many existing datasets. When applied at a large scale, this technique can compete with task-specific supervised models, while there is still much space for improvement.

This research is ranked #1 on Zero-Shot Transfer Image Classification on SUN

CoAtNet: Marrying Convolution and Attention for All Data Sizes

The researchers in this article conduct a thorough examination of the features of convolutions and transformers, resulting in a principled approach for combining them into a new family of models dubbed CoAtNet. Extensive experiments demonstrate that CoAtNet combines the advantages of ConvNets and Transformers, achieving state-of-the-art performance across a range of data sizes and compute budgets. Take note that this article is currently concentrating on ImageNet classification for model construction. However, the researchers believe their approach is relevant to a broader range of applications, such as object detection and semantic segmentation.

This paper is ranked #1 on Image Classification on ImageNet (using extra training data).

SwinIR: Image Restoration Using Swin Transformer

The authors of this article suggest the SwinIR image restoration model, which is based on the Swin Transformer . The model comprises three modules: shallow feature extraction, deep feature extraction, and human-recognition reconstruction. For deep feature extraction, the researchers employ a stack of residual Swin Transformer blocks (RSTB), each formed of Swin Transformer layers, a convolution layer, and a residual connection.

This research article is ranked #1 on Image Super-Resolution on Manga109 – 4x upscaling.

Artificial Replay: A Meta-Algorithm for Harnessing Historical Data in Bandits

Ways to incorporate historical data are still unclear: initialising reward estimates with historical samples can suffer from bogus and imbalanced data coverage, leading to computational and storage issues—particularly in continuous action spaces. The paper addresses the obstacles by proposing ‘Artificial Replay’, an algorithm to incorporate historical data into any arbitrary base bandit algorithm. 

Read the full paper here . 

Bootstrapped Meta-Learning

Author(s) – Sean R. Sinclair et al.

The paper proposes an algorithm in which the meta-learner teaches itself to overcome the meta-optimisation challenge. The algorithm focuses on meta-learning with gradients, which guarantees performance improvements. Furthermore, the paper also looks at how bootstrapping opens up possibilities. 

Read the full paper here .

LaMDA: Language Models for Dialog Applications

Author(s) – Sebastian Flennerhag et al.

The research describes the LaMDA system which caused chaos in AI this summer when a former Google engineer claimed that it had shown signs of sentience. LaMDA is a family of large language models for dialogue applications based on Transformer architecture. The interesting feature of the model is its fine-tuning with human-annotated data and the possibility of consulting external sources. This is a very interesting model family, which we might encounter in many applications we use daily. 

Competition-Level Code Generation with AlphaCode

Author(s) – Yujia Li et al.

Systems can help programmers become more productive. The following research addresses the problems with incorporating innovations in AI into these systems. AlphaCode is a system that creates solutions for problems that require deeper reasoning. 

Privacy for Free: How does Dataset Condensation Help Privacy?

Author(s) – Tian Dong et al.

The paper focuses on Privacy Preserving Machine Learning, specifically deducting the leakage of sensitive data in machine learning. It puts forth one of the first propositions of using dataset condensation techniques to preserve the data efficiency during model training and furnish membership privacy.

Why do tree-based models still outperform deep learning on tabular data?

Author(s) – Léo Grinsztajn, Edouard Oyallon and Gaël Varoquaux

The research answers why deep learning models still find it hard to compete on tabular data compared to tree-based models. It is shown that MLP-like architectures are more sensitive to uninformative features in data compared to their tree-based counterparts. 

Multi-Objective Bayesian Optimisation over High-Dimensional Search Spaces 

Author(s) – Samuel Daulton et al.

The paper proposes ‘MORBO’, a scalable method for multiple-objective BO as it performs better than that of high-dimensional search spaces. MORBO significantly improves the sample efficiency and, where existing BO algorithms fail, MORBO provides improved sample efficiencies over the current approach. 

A Path Towards Autonomous Machine Intelligence Version 0.9.2

Author(s) – Yann LeCun

The research offers a vision about how to progress towards general AI. The study combines several concepts: a configurable predictive world model, behaviour driven through intrinsic motivation, and hierarchical joint embedding architectures trained with self-supervised

learning. 

TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data

Author(s) –  Shreshth Tuli, Giuliano Casale and Nicholas R. Jennings

This is a specialised paper applying transformer architecture to the problem of unsupervised anomaly detection in multivariate time series. Many architectures which were successful in other fields are, at some point, also being applied to time series. The research shows improved performance on some known data sets. 

Differentially Private Bias-Term only Fine-tuning of Foundation Models

Author(s) – Zhiqi Bu et al. 

In the paper, researchers study the problem of differentially private (DP) fine-tuning of large pre-trained models—a recent privacy-preserving approach suitable for solving downstream tasks with sensitive data. Existing work has demonstrated that high accuracy is possible under strong privacy constraints yet requires significant computational overhead or modifications to the network architecture.

ALBERT: A Lite BERT

Usually, increasing model size when pretraining natural language representations often result in improved performance on downstream tasks, but the training times become longer. To address these problems, the authors in their work presented two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT. The authors also used a self-supervised loss that focuses on modelling inter-sentence coherence and consistently helped downstream tasks with multi-sentence inputs. According to results, this model established new state-of-the-art results on the GLUE, RACE, and squad benchmarks while having fewer parameters compared to BERT-large. 

Check the paper here .

Beyond Accuracy: Behavioral Testing of NLP Models with CheckList

Microsoft Research, along with the University of Washington and the University of California, in this paper, introduced a model-agnostic and task agnostic methodology for testing NLP models known as CheckList. This is also the winner of the best paper award at the ACL conference this year. It included a matrix of general linguistic capabilities and test types that facilitate comprehensive test ideation, as well as a software tool to generate a large and diverse number of test cases quickly. 

Linformer is a Transformer architecture for tackling the self-attention bottleneck in Transformers. It reduces self-attention to an O(n) operation in both space- and time complexity. It is a new self-attention mechanism which allows the researchers to compute the contextual mapping in linear time and memory complexity with respect to the sequence length. 

Read more about the paper here .

Plug and Play Language Models

Plug and Play Language Models ( PPLM ) are a combination of pre-trained language models with one or more simple attribute classifiers. This, in turn, assists in text generation without any further training. According to the authors, model samples demonstrated control over sentiment styles, and extensive automated and human-annotated evaluations showed attribute alignment and fluency. 

Reformer 

The researchers at Google, in this paper , introduced Reformer. This work showcased that the architecture of a Transformer can be executed efficiently on long sequences and with small memory. The authors believe that the ability to handle long sequences opens the way for the use of the Reformer on many generative tasks. In addition to generating very long coherent text, the Reformer can bring the power of Transformer models to other domains like time-series forecasting, music, image and video generation. 

An Image is Worth 16X16 Words

The irony here is that one of the popular language models, Transformers have been made to do computer vision tasks. In this paper , the authors claimed that the vision transformer could go toe-to-toe with the state-of-the-art models on image recognition benchmarks, reaching accuracies as high as 88.36% on ImageNet and 94.55% on CIFAR-100. For this, the vision transformer receives input as a one-dimensional sequence of token embeddings. The image is then reshaped into a sequence of flattened 2D patches. The transformers in this work use constant widths through all of its layers.

Unsupervised Learning of Probably Symmetric Deformable 3D Objects

Winner of the CVPR best paper award, in this work, the authors proposed a method to learn 3D deformable object categories from raw single-view images, without external supervision. This method uses an autoencoder that factored each input image into depth, albedo, viewpoint and illumination. The authors showcased that reasoning about illumination can be used to exploit the underlying object symmetry even if the appearance is not symmetric due to shading.

Generative Pretraining from Pixels

In this paper, OpenAI researchers examined whether similar models can learn useful representations for images. For this, the researchers trained a sequence Transformer to auto-regressively predict pixels, without incorporating knowledge of the 2D input structure. Despite training on low-resolution ImageNet without labels, the researchers found that a GPT-2 scale model learns strong image representations as measured by linear probing, fine-tuning, and low-data classification. On CIFAR-10, it achieved 96.3% accuracy with a linear probe, outperforming a supervised Wide ResNet, and 99.0% accuracy with full fine-tuning and matching the top supervised pre-trained models. An even larger model, trained on a mixture of ImageNet and web images, is competitive with self-supervised benchmarks on ImageNet, achieving 72.0% top-1 accuracy on a linear probe of their features.

Deep Reinforcement Learning and its Neuroscientific Implications

In this paper, the authors provided a high-level introduction to deep RL , discussed some of its initial applications to neuroscience, and surveyed its wider implications for research on brain and behaviour and concluded with a list of opportunities for next-stage research. Although DeepRL seems to be promising, the authors wrote that it is still a work in progress and its implications in neuroscience should be looked at as a great opportunity. For instance, deep RL provides an agent-based framework for studying the way that reward shapes representation, and how representation, in turn, shapes learning and decision making — two issues which together span a large swath of what is most central to neuroscience. 

Dopamine-based Reinforcement Learning

Why humans doing certain things are often linked to dopamine , a hormone that acts as the reward system (think: the likes on your Instagram page). So, keeping this fact in hindsight, DeepMind with the help of Harvard labs, analysed dopamine cells in mice and recorded how the mice received rewards while they learned a task. They then checked these recordings for consistency in the activity of the dopamine neurons with standard temporal difference algorithms. This paper proposed an account of dopamine-based reinforcement learning inspired by recent artificial intelligence research on distributional reinforcement learning. The authors hypothesised that the brain represents possible future rewards not as a single mean but as a probability distribution, effectively representing multiple future outcomes simultaneously and in parallel. 

Lottery Tickets In Reinforcement Learning & NLP

In this paper, the authors bridged natural language processing (NLP) and reinforcement learning (RL). They examined both recurrent LSTM models and large-scale Transformer models for NLP and discrete-action space tasks for RL. The results suggested that the lottery ticket hypothesis is not restricted to supervised learning of natural images, but rather represents a broader phenomenon in deep neural networks.

What Can Learned Intrinsic Rewards Capture

In this paper, the authors explored if the reward function itself can be a good locus of learned knowledge. They proposed a scalable framework for learning useful intrinsic reward functions across multiple lifetimes of experience and showed that it is feasible to learn and capture knowledge about long-term exploration and exploitation into a reward function. 

AutoML- Zero

The progress of AutoML has largely focused on the architecture of neural networks, where it has relied on sophisticated expert-designed layers as building blocks, or similarly restrictive search spaces. In this paper , the authors showed that AutoML could go further with AutoML Zero, that automatically discovers complete machine learning algorithms just using basic mathematical operations as building blocks. The researchers demonstrated this by introducing a novel framework that significantly reduced human bias through a generic search space.

Rethinking Batch Normalization for Meta-Learning

Batch normalization is an essential component of meta-learning pipelines. However, there are several challenges. So, in this paper, the authors evaluated a range of approaches to batch normalization for meta-learning scenarios and developed a novel approach — TaskNorm. Experiments demonstrated that the choice of batch normalization has a dramatic effect on both classification accuracy and training time for both gradient-based and gradient-free meta-learning approaches. The TaskNorm has been found to be consistently improving the performance.

Meta-Learning without Memorisation

Meta-learning algorithms need meta-training tasks to be mutually exclusive, such that no single model can solve all of the tasks at once. In this paper, the authors designed a meta-regularisation objective using information theory that successfully uses data from non-mutually-exclusive tasks to efficiently adapt to novel tasks.

Understanding the Effectiveness of MAML

Model Agnostic Meta-Learning (MAML) consists of optimisation loops, from which the inner loop can efficiently learn new tasks. In this paper, the authors demonstrated that feature reuse is the dominant factor and led to ANIL (Almost No Inner Loop) algorithm — a simplification of MAML where the inner loop is removed for all but the (task-specific) head of the underlying neural network. 

Your Classifier is Secretly an Energy-Based Model

This paper proposed attempts to reinterpret a standard discriminative classifier as an energy-based model. In this setting, wrote the authors, the standard class probabilities can be easily computed. They demonstrated that energy-based training of the joint distribution improves calibration, robustness, handout-of-distribution detection while also enabling the proposed model to generate samples rivalling the quality of recent GAN approaches. This work improves upon the recently proposed techniques for scaling up the training of energy-based models. It has also been the first to achieve performance rivalling the state-of-the-art in both generative and discriminative learning within one hybrid model.

Reverse-Engineering Deep ReLU Networks

This paper investigated the commonly assumed notion that neural networks cannot be recovered from its outputs, as they depend on its parameters in a highly nonlinear way. The authors claimed that by observing only its output, one could identify the architecture, weights, and biases of an unknown deep ReLU network. By dissecting the set of region boundaries into components associated with particular neurons, the researchers showed that it is possible to recover the weights of neurons and their arrangement within the network.

Cricket Analytics and Predictor

Authors: Suyash Mahajan,  Salma Shaikh, Jash Vora, Gunjan Kandhari,  Rutuja Pawar,

Abstract:   The paper embark on predicting the outcomes of Indian Premier League (IPL) cricket match using a supervised learning approach from a team composition perspective. The study suggests that the relative team strength between the competing teams forms a distinctive feature for predicting the winner. Modeling the team strength boils down to modeling individual player‘s batting and bowling performances, forming the basis of our approach.

Research Methodology: In this paper, two methodologies have been used. MySQL database is used for storing data whereas Java for the GUI. The algorithm used is Clustering Algorithm for prediction. The steps followed are as

• Begin with a decision on the value of k being the number of clusters.
• Put any initial partition that classifies the data into k clusters.
• Take every sample in the sequence; compute its distance from centroid of each of the clusters. If sample is not in the cluster with the closest centroid currently, switch this sample to that cluster and update the centroid of the cluster accepting the new sample and the cluster losing the sample.

For the research paper, read here

2.Real Time Sleep / Drowsiness Detection – Project Report

Author : Roshan Tavhare

Institute : University of Mumbai

Abstract : The main idea behind this project is to develop a nonintrusive system which can detect fatigue of any human and can issue a timely warning. Drivers who do not take regular breaks when driving long distances run a high risk of becoming drowsy a state which they often fail to recognize early enough.

Research Methodology : A training set of labeled facial landmarks on an image. These images are manually labeled, specifying specific (x, y) -coordinates of regions surrounding each facial structure.

• Priors, more specifically, the probability on distance between pairs of input pixels. The pre-trained facial landmark detector inside the dlib library is used to estimate the location of 68 (x, y)-coordinates that map to facial structures on the face.

A Study of Various Text Augmentation Techniques for Relation Classification in Free Text

Authors: Chinmaya Mishra Praveen Kumar and Reddy Kumar Moda,  Syed Saqib Bukhari and Andreas Dengel

Institute: German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany

Abstract: In this paper, the researchers explore various text data augmentation techniques in text space and word embedding space. They studied the effect of various augmented datasets on the efficiency of different deep learning models for relation classification in text.

Research Methodology: The researchers implemented five text data augmentation techniques (Similar word, synonyms, interpolation, extrapolation and random noise method)  and explored the ways in which we could preserve the grammatical and the contextual structures of the sentences while generating new sentences automatically using data augmentation techniques.

Smart Health Monitoring and Management Using Internet of Things, Artificial Intelligence with Cloud Based Processing

Author : Prateek Kaushik

Institute : G D Goenka University, Gurugram

Abstract : This research paper described a personalised smart health monitoring device using wireless sensors and the latest technology.

Research Methodology: Machine learning and Deep Learning techniques are discussed which works as a catalyst to improve the  performance of any health monitor system such supervised machine learning algorithms, unsupervised machine learning algorithms, auto-encoder, convolutional neural network and restricted boltzmann machine .

Internet of Things with BIG DATA Analytics -A Survey

Author : A.Pavithra,  C.Anandhakumar and V.Nithin Meenashisundharam

Institute : Sree Saraswathi Thyagaraja College,

Abstract : This article we discuss about Big data on IoT and how it is interrelated to each other along with the necessity of implementing Big data with IoT and its benefits, job market

Research Methodology : Machine learning, Deep Learning, and Artificial Intelligence are key technologies that are used to provide value-added applications along with IoT and big data in addition to being used in a stand-alone mod.

Single Headed Attention RNN: Stop Thinking With Your Head 

Author: Stephen Merity

In this work of art, the Harvard grad author, Stephen “Smerity” Merity, investigated the current state of NLP, the models being used and other alternate approaches. In this process, he tears down the conventional methods from top to bottom, including etymology.

The author also voices the need for a Moore’s Law for machine learning that encourages a minicomputer future while also announcing his plans on rebuilding the codebase from the ground up both as an educational tool for others and as a strong platform for future work in academia and industry.

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Authors: Mingxing Tan and Quoc V. Le 

In this work, the authors propose a compound scaling method that tells when to increase or decrease depth, height and resolution of a certain network.

Convolutional Neural Networks(CNNs) are at the heart of many machine vision applications. 

EfficientNets are believed to superpass state-of-the-art accuracy with up to 10x better efficiency (smaller and faster).

Deep Double Descent By OpenAI

Authors: Mikhail Belkin, Daniel Hsu, Siyuan Ma, Soumik Mandal

In this paper , an attempt has been made to reconcile classical understanding and modern practice within a unified performance curve. 

The “double descent” curve overtakes the classic U-shaped bias-variance trade-off curve by showing how increasing model capacity beyond the point of interpolation results in improved performance. 

The Lottery Ticket Hypothesis

Authors: Jonathan Frankle, Michael Carbin

Neural network pruning techniques can reduce the parameter counts of trained networks by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy. 

The authors find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively. Based on these results, they introduce the “lottery ticket hypothesis:”

On The Measure Of Intelligence 

Authors: Francois Chollet

This work summarizes and critically assesses the definitions of intelligence and evaluation approaches, while making apparent the historical conceptions of intelligence that have implicitly guided them.

The author, also the creator of keras, introduces a formal definition of intelligence based on Algorithmic Information Theory and using this definition, he also proposes a set of guidelines for what a general AI benchmark should look like. 

Zero-Shot Word Sense Disambiguation Using Sense Definition Embeddings via IISc Bangalore & CMU

Authors: Sawan Kumar, Sharmistha Jat, Karan Saxena and Partha Talukdar

Word Sense Disambiguation (WSD) is a longstanding  but open problem in Natural Language Processing (NLP).  Current supervised WSD methods treat senses as discrete labels  and also resort to predicting the Most-Frequent-Sense (MFS) for words unseen  during training.

The researchers from IISc Bangalore in collaboration with Carnegie Mellon University propose  Extended WSD Incorporating Sense Embeddings (EWISE), a supervised model to perform WSD  by predicting over a continuous sense embedding space as opposed to a discrete label space.

Deep Equilibrium Models 

Authors: Shaojie Bai, J. Zico Kolter and Vladlen Koltun 

Motivated by the observation that the hidden layers of many existing deep sequence models converge towards some fixed point, the researchers at Carnegie Mellon University present a new approach to modeling sequential data through deep equilibrium model (DEQ) models. 

Using this approach, training and prediction in these networks require only constant memory, regardless of the effective “depth” of the network.

IMAGENET-Trained CNNs are Biased Towards Texture

Authors: Robert G, Patricia R, Claudio M, Matthias Bethge, Felix A. W and Wieland B

Convolutional Neural Networks (CNNs) are commonly thought to recognise objects by learning increasingly complex representations of object shapes. The authors in this paper , evaluate CNNs and human observers on images with a texture-shape cue conflict. They show that ImageNet-trained CNNs are strongly biased towards recognising textures rather than shapes, which is in stark contrast to human behavioural evidence.

A Geometric Perspective on Optimal Representations for Reinforcement Learning 

Authors: Marc G. B , Will D , Robert D , Adrien A T , Pablo S C , Nicolas Le R , Dale S, Tor L, Clare L

The authors propose a new perspective on representation learning in reinforcement learning

based on geometric properties of the space of value functions. This work shows that adversarial value functions exhibit interesting structure, and are good auxiliary tasks when learning a representation of an environment. The authors believe this work to open up the possibility of automatically generating auxiliary tasks in deep reinforcement learning.

Weight Agnostic Neural Networks 

Authors: Adam Gaier & David Ha

In this work , the authors explore whether neural network architectures alone, without learning any weight parameters, can encode solutions for a given task. In this paper, they propose a search method for neural network architectures that can already perform a task without any explicit weight training. 

Stand-Alone Self-Attention in Vision Models 

Authors: Prajit Ramachandran, Niki P, Ashish Vaswani,Irwan Bello Anselm Levskaya, Jonathon S

In this work, the Google researchers verified that content-based interactions can serve the vision models . The proposed stand-alone local self-attention layer achieves competitive predictive performance on ImageNet classification and COCO object detection tasks while requiring fewer parameters and floating-point operations than the corresponding convolution baselines. Results show that attention is especially effective in the later parts of the network. 

High-Fidelity Image Generation With Fewer Labels 

Authors: Mario Lucic, Michael Tschannen, Marvin Ritter, Xiaohua Z, Olivier B and Sylvain Gelly 

Modern-day models can produce high quality, close to reality when fed with a vast quantity of labelled data. To solve this large data dependency, researchers from Google released this work , to demonstrate how one can benefit from recent work on self- and semi-supervised learning to outperform the state of the art on both unsupervised ImageNet synthesis, as well as in the conditional setting.

The proposed approach is able to match the sample quality of the current state-of-the-art conditional model BigGAN on ImageNet using only 10% of the labels and outperform it using 20% of the labels.

ALBERT: A Lite BERT for Self-Supervised Learning of Language Representations

Authors: Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin G, Piyush Sharma and Radu S

The authors present two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT and to address the challenges posed by increasing model size and GPU/TPU memory limitations, longer training times, and unexpected model degradation

As a result, this proposed model establishes new state-of-the-art results on the GLUE, RACE, and SQuAD benchmarks while having fewer parameters compared to BERT-large.

GauGANs-Semantic Image Synthesis with Spatially-Adaptive Normalization 

Author: Taesung Park, Ming-Yu Liu, Ting-Chun Wang and Jun-Yan Zhu

Nvidia in collaboration with UC Berkeley and MIT proposed a model which has a spatially-adaptive normalization layer for synthesizing photorealistic images given an input semantic layout.

This model retained visual fidelity and alignment with challenging input layouts while allowing the user to control both semantic and style.

📣 Want to advertise in AIM? Book here

Subscribe to The Belamy: Our Weekly Newsletter

Biggest ai stories, delivered to your inbox every week..

Discover how Cypher 2024 expands to the USA, bridging AI innovation gaps and tackling the challenges of enterprise AI adoption

© Analytics India Magazine Pvt Ltd & AIM Media House LLC 2024

• Terms of use
• Privacy Policy