The AI-Surveillance Symbiosis in China

Artificial Intelligence and the U.S.-China Relationship

Artificial intelligence (AI) technology is already playing a role in people’s daily lives through applications ranging from driver-assistance systems, to medical diagnostics and financial management. AI is a general-purpose technology that, like electricity, has the potential to deeply change and restructure many economic and social activities. However, AI’s potential to reshape economic activity raises fraught questions for governments around the world. What are AI’s potential economic and national security advantages? What risks does AI pose to democratic institutions and civil rights? What does “leading AI” mean in a world where transnational research and collaboration are the norm?

In Washington, much of the debate is shaped by concerns that Beijing is succeeding in catching up in the AI race, which many worry will support China’s commercial ambitions, assertive national security, and repressive domestic security apparatus. As a result, AI is now one of the leading concerns in the U.S.-China strategic competition. This competition is now a central lens through which to understand AI’s trajectory and significance. Even though AI development involves extensive global collaboration, many believe that AI’s impact is so vast that countries “leading” in its development and deployment could have a geopolitical edge.

By most metrics, the United States and China are two of the leading countries in AI, with multiple research institutions and companies operating at the cutting edge in various domains of applied AI and advanced research and development (R&D). Although there are many international and domestic factors shaping the trajectory of AI in China, this Big Data China feature provides insight into how the relationship between the state and the private sector in China plays an important role in promoting innovation. The analysis highlights recent data-driven scholarship by Professors Noam Yuchtman (London School of Economics) and David Yang (Harvard University), which sheds light on the positive feedback loop between state surveillance and companies operating in one area of AI: facial recognition technology.

China’s AI Rise

Chinese companies have advanced rapidly in AI R&D and commercial applications. One reason for this is the gradual growth of entrepreneurial class that is driving development of China’s large technology platforms over the past two decades and a highly dynamic and competitive business landscape. International collaboration, through corporate investments and partnerships as well as transnational scholarly research, has also been central.

At the same time, companies driving the applications of AI in China have benefited from targeted state support for both investment and R&D. The Chinese government often sets unrealistically ambitious targets, but its approach can be effective in mobilizing massive resources toward targeted industries. As Kaifu Lee’s and other analysis shows, Chinese AI firms have become more innovative in part because China’s policy environment has fostered a nimble commercial ecosystem. Moreover, as this feature explains, the size of China’s population combined with extensive surveillance means that Chinese companies can potentially tap into enormous amounts of data collected by government agencies, including through surveillance cameras and smart-city systems, when they win public contracts. Large datasets are important inputs for spurring growth and innovation in AI, and Chinese tech companies have been able to acquire data with substantial depth, covering a wide variety of behaviors for individuals, thanks to the ubiquity of mobile and online services in the country. While a recent regulatory crackdown affecting some of China’s top internet companies may have dampened private sector growth in some areas, China’s government has remained strongly committed to prioritizing the development of AI technology and its widespread application.

Differences in the political economy of the U.S. and China has resulted in AI developing in different ways. The United States has traditionally been a magnet for international talent and benefited from world-class research institutions. This relatively open system has generally been an advantage in AI and contributed to the United States’ leading position in general AI R&D.

While the growth of China’s AI sector goes back several years, it was largely unnoticed in the United States because many dismissed China’s innovation capacity. But all levels of government in China have engaged in adopting and deploying AI as part of a goal set in 2017 to make China the world’s leading AI hub by 2030.

China now far outpaces the United States in the number of patents filed (although not patents granted) and journal articles published. China is also home to a huge number of competitive AI companies supported by growing private and public investment. According to an estimate from Georgetown’s Center for Science and Emerging Technology (CSET), the Chinese government spent at least $2 billion supporting AI research in 2018. There is no direct comparison between the United States and China for AI R&D spending, but in 2018, by comparison the U.S. government’s non-defense AI R&D was estimated at around $560 million. U.S. private investment in AI, however, far outpaces the rest of the world, and in 2021 it was three times the size of China’s.

Recognizing the China Challenge

While China’s AI capacity was not a major topic of conversation in Washington at the start of the Trump administration, concerns that American research and deployment of AI was lagging grew significantly by 2018. The National Security Commission on Artificial Intelligence (NSCAI), chaired by former Google CEO Eric Schmidt, was established that year, while the White House launched an official AI strategy in early 2019. By 2021, when the NSCAI released its final report, federal spending on non-defense AI R&D had almost tripled from 2018, to $1.53 billion. Investment from public and private sources has only continued to increase, and AI has become a top concern of policymakers in the United States and around the world.

Tech competition with China, including on AI and semiconductors, used to train AI algorithms in the cloud and run them on the edge, is reshaping some aspects of the U.S. government’s traditional approach to innovation. It is also resulting in more direct government attention to supply-chain resilience, data protection, and strategic industries. Examples include the CHIPS Act and the establishment of a National Science Foundation directorate—something that is under discussion in Congress and may be included in the final version of the Bipartisan Innovation Act—if agreement can be reached.

AI could boost China’s companies’ productivity and competitiveness. It also may carry national security risks for the Unites States and its partners as China’s military expands its AI capabilities. But as the United States develops a response in the competition with China for AI leadership, it will have to be conscious of the implications for its own economy, national security, and governance.

What Does AI Mean for Human Rights?

The implications of the adoption of certain types of applications that leverage AI are still far from settled, especially for civil and human rights. The use of security cameras and drones, for example, which may also be coupled with facial or object recognition systems that leverage AI algorithms, has increased globally in the past 20 years. Globally, governments and societies are still struggling to find the right balance between protecting citizens’ rights to privacy and avoiding stifling innovation in an emerging sector. In the United States alone, the number of surveillance cameras increased 50 percent between 2015 and 2018, with an estimated 85 million such devices in the country by 2021. Many of the cameras installed across the United States are private, and not all may feature the use of facial or object recognition, but rather just use live feeds or recording live video. Over the last several years, the use of facial recognition systems for specific applications, such as law enforcement, has come under growing scrutiny.

In China, there has also been an uptick in concerns about privacy in recent years. The passage last year of two major laws, including the Personal Information Protection Law and the Data Security Law means that China has one of the strongest data governance regimes in the world on paper. However, even before the advent of advanced technology systems using AI, Chinese public security organizations were quite adept at tracking and monitoring criminal suspects and regime dissidents. Facial recognition systems are now used to enhance the public security apparatus’ surveillance capabilities. For example, there is strong evidence that some systems using facial recognition are being used specifically to target citizens of Uyghur background in Xinjiang province. This has resulted in the U.S. sanctioning several Chinese companies that have developed facial recognition technology and provided the technology and some services to public security organs. As a result, several leading Chinese AI firms are now included on the Department of Commerce’s Bureau of Industry and Security’s Entity List.

Chinese AI Firms Lead in Some Areas of Innovation

One consistent challenge in measuring countries’ performances in the AI race is identifying the right metrics. One approach is to analyze outputs such as the number and quality of patents and relevant publications or even more importantly, the performance of algorithms in contests or benchmarks. The challenge is that the results can vary across methodologies and depend on the type of AI that is being considered. For example, patents and papers from Chinese-based companies and researchers outpace those from their American-based counterparts in terms of volume, but they often are of lower quality. There are also differences across sub-fields: for example, contests from 2020 and 2021, where different algorithms are compared by examining how quickly and accurately they can perform tasks, show that Chinese companies perform particularly well in facial recognition technology. But in other areas, including language, Chinese algorithms rank less highly. Similarly, expert interviews suggest that when it comes to basic research, China has generally not surpassed the United States.

A different way to assess performance is to focus on how resources stack up, such as the number of experts and the quality of their training, the degree of access to advanced hardware that enables advances in AI, and the kind of data that AI companies can access. Here, Chinese companies are often thought to have a real advantage thanks to the extensive use of tech platforms by the country’s large population. For example, in 2017, Chinese users paid for goods with smartphones 50 times more often than Americans, and food delivery volume was 10 times higher than in the United States. Plus, ubiquitous public surveillance also creates extensive data.

There are some clear caveats here. For example, leading American tech companies such as Alphabet and Meta collect enormous amounts of data all around the world through their products. While companies like Bytedance and Alibaba are expanding their global user base, their geographic reach remains less extensive. The quality of data, including its precision, also matters, as does the demographic diversity and the kind of data collected. It is not yet clear, for example, that having access to large data sets on financial transactions is always helpful to applications in biotech, although this may change. Finally, in some areas of AI research, data may have become less important compared to, for example, computing power.

Drawing on a massive data set of government procurement contracts, Professors Yuchtman and Yang confirm that that access to data is indeed a real advantage in AI performance and machine learning, at least for facial recognition. Their analysis of the relationship between the Chinese surveillance state and the AI industry shows that, at least in this field, companies can gain commercial advantage from using data acquired through government public security procurement contracts.

The Chinese surveillance AI industry is expanding, with both security cameras and facial recognition software sales growing rapidly (see Figure 1). This parallels a global expansion in the deployment of surveillance technology as well as China’s increasingly widespread application of AI and advanced technologies for surveillance ends.

Figure 1

Extensive data on public procurement and firms’ product development and innovation analyzed by Yuchtman and Yang shows that this expansion enhances the state’s ability to suppress dissent and is a real boon to companies.

Local Governments in China Are Major Buyers of AI Technology

The Chinese government is a major buyer of AI technology, including devices linked with the enhancement of China’s surveillance state, such as facial recognition (see Figure 2). In addition to the size of the Chinese public investment, large-scale government procurement of systems using some AI technology means that the state has at its disposal significant leverage when it comes to supporting Chinese firms. Whether this is intentional or not, the end result is not unlike the deployment of targeted public procurement in the service of industrial policy, a well-documented practice.

As Figure 2 shows, the purchase of AI technology, including public security AI technology, is increasing over time. However, the government also purchases systems that utilize AI technology for applications unrelated to security. This is not surprising since there are all sorts of areas in which AI can be applied to improve government functions.

Figure 2

China’s political structure plays an important role in how contracts are issued. Local governments tend to play a crucial role in implementing central-level targets, and AI and surveillance are no exceptions. As Figure 2 shows, the central government is not the main buyer of AI technology. Prefecture and sub-prefecture-level governments (large and medium-sized cities) are leading in the purchase of AI technology. This is unsurprising since local governments, in China as elsewhere, are tasked with enforcing public security and other non-public security administrative tasks that AI technology can facilitate. Moreover, local governments tend to focus their acquisition and investment efforts on systems that the central government has signaled its support for to ensure alignment with Beijing.

Data for the entirety of 2019 and following years could not be retrieved, but the trends depicted in Figures 2 have likely grown due to the government’s “Zero-Covid” policy, which requires extensive monitoring capacity to work.

Public Security and AI Technology

One of Yuchtman, Yang, and their collaborators’ key findings is that there is a direct link between instances of social unrest and the purchase of AI technology. After controlling for multiple variables and using different statistical methods, they found that the number of public security AI technology contracts in a locality increased in the period immediately following an instance of unrest. (See Box 1 below for more details on methodology.)

More specifically, the data show that public security AI contracts only increase in the quarter immediately following instances of unrest (see Figure 3). This means that the uptick in public security facial recognition contracts is linked to the occurrence of protests and is not simply part of a general trend toward the acquisition of this technology driven by external factors such as economic growth or central government targets.

Figure 3

By contrast, Yuchtman, Yang, and their co-authors found no connection between unrest and the purchase of facial recognition AI for non-public security purposes. There is no variation over time in the procurement of other kinds of AI, proving that the relationship is only related to unrest and public security facial recognition. However, there is a positive effect on complementary surveillance technology investments: local governments also purchase more high-resolution surveillance cameras and more police are assigned to desk jobs, likely tasked with overseeing the technology deployed for public security purposes.

A corollary of these findings is that the deployment of surveillance technology is a decentralized process in China. The purchase of AI technology is not as centrally coordinated as might be otherwise assumed looking at China’s AI ambitions from the outside. While Beijing does actively encourage this expansion in surveillance through multiple programs, local public security agencies are key decisionmakers. The role of local governments may complicate further the job of researchers collecting data and analyzing in this area.

AI Tech Prevents Future Unrest

While trends in the purchasing of AI technology are clearly driven by local governments’ response to unrest, Professors Yuchtman, Yang, and their co-authors show that it can have a real political impact. They conducted advanced statistical analyses to prove that localities which had previously invested in public security facial recognition platforms were subsequently less likely to experience unrest. (See Box 1 below for more details on methodology.) They also found no similar effect on unrest for localities that purchased non-public AI technology.

Local governments have multiple incentives to purchase AI surveillance technology. For example, if agencies do not spend their full budget, they may see their future budget revised downward, and public surveillance technology is generally accepted as a responsible investment. But most importantly, career progression for government officials in China is tied to their success in preventing or quelling unrest. As a result, there is a particularly strong incentive to deploy this type of surveillance technology if it reduces unrest.

AI Firms Gain Commercial Advantage from Public Security Contracts

The data presented so far explain some of the drivers of the Chinese government’s push to purchase public security AI software. But growth in the acquisition of public security AI also has innovation benefits for firms as well. This may be the part of Yuchtman and Yang’s research that has some of the greatest implications for the U.S.-China tech competition. As Figure 4 shows, winning a contract to supply the government with facial recognition AI for surveillance purposes brings clear commercial advantages to firms. After receiving a public security contract from a city with above median surveillance capacity, which can be characterized as ‘data-rich,’ firms released more software products (including new software or major releases), all else equal. The contracts are beneficial because they unlock access to a wealth of data that firms may otherwise struggle to acquire. And the fact that bids for data-rich government contracts are more abundant and lower than for other types of contracts suggests that firms are aware of these advantages.

Figure 4

To prove the link between public procurement and firm performance, and the mechanisms that provide the commercial advantage, Yuchtman and Yang and their collaborators analyzed information on all Chinese facial recognition firms and almost 3 million public procurement contracts from all levels of Chinese government between 2013 and 2019 from the Chinese Government Procurement Database, maintained by China’s Ministry of Finance. Among those, they found over 28,000 contracts that involved at least one AI company. They matched those contracts with their list of AI companies and found that 1,095 AI companies in their list received at least one pubic contract (See Box 1 below for more details on methodology.)

By matching the performance of AI firms relative to their first public security contract, they found that a data-rich government contract was linked with an immediate increase in software products released, an effect that held for at least three years. The scholars also tracked whether the software released had government or commercial applications and found that the impact was significant for both types of software. Firms released on average two more software products with commercial applications and three more with government applications compared to companies that hadn’t won a government contract. Firms with data-rich public security contracts also saw an increase in sales of non-AI, data-complementary software (for example, software supporting data storage and transmission). This suggests that government procurement, at least in this area, creates opportunities for economies of scope, the economic efficiencies that emerge when a company expands the variety of products and services it offers (as opposed to expanding the volume, as in economies of scale).

Figure 5

In short, there is a clear positive feedback loop between the surveillance state and companies producing facial recognition software (see Figure 5). Local governments purchase AI surveillance technology in direct response to unrest, and these investments help deter protests. At the same time, the AI firms that supply that technology benefit commercially from the access to data through government contracts. Other research has shown that this relationship has led private firms to develop and offer services that go even beyond what the government had initially requested.

Policy Implications

While Yuchtman and Yang’s papers do not include specific policy analysis and recommendations, their research has important policy implications for China, the United States, and the international community. With regard to China, it appears that AI is enhancing the regime’s surveillance state capabilities, which could have serious human rights implications. In addition, Chinese AI companies are benefiting from access to data acquired through government contracts. Given the extensive state surveillance in place in China, this could be a real advantage to firms and may set them ahead of their international competitors. This may enhance China’s “data advantage,” which analysts think provides a comparative advantage in some areas of AI. This government-induced tilt toward sensory AI and surveillance also means that fewer public resources may be going towards other areas of AI, like language.

Given these trends in China, what should the United States do? There may be some steps the United States can take to stymie the growth of China’s surveillance state or homegrown AI industry, but there are reasons to doubt the effectiveness of such measures. A more effective approach will involve additional efforts to promote American innovation and safeguard national security, which will require Washington to overcome partisan gridlock. The proposals here are divided into three categories: expanded regulation, advance innovation, and address China’s surveillance-AI symbiosis.

1. Expand regulation of data and of surveillance technology at home and internationally.

The use of facial recognition technology and the surveillance application of AI is highly controversial in democratic countries—for good reason. U.S. policymakers should be careful to avoid a “race to the bottom” over data protection rules in the name of competition with China. Regulation could clarify how companies should operate in this rapidly changing space and allow the United States to engage in the global standard-setting debate. To this end, U.S. policymakers should seek to promote global standards and norms to ensure that AI is used ethically.

By not weighing in on data regulation, Washington is letting Beijing and Brussels establish principles in the private sector and could be ceding its power to shape global tech standards. This may ultimately hurt competitiveness rather than aid it. Engaging more seriously with the model set by the European Global Data Protection Regulation (GDPR) and developing a data protection regime that can work in the United States could be beneficial to this end. The European Union has been particularly active compared to the United States when it comes to setting standards on data, AI, and its applications.

A positive sign is that there are ongoing efforts to coordinate with other like-minded countries within platforms focused on developing common standards and regulatory frameworks, particularly the U.S.-EU Trade and Technology Council, the Quadrilateral Security Dialogue (Quad), and potentially the Indo-Pacific Economic Forum. But these should be further enhanced to cover the digital space. Accelerating efforts to develop common standards on the use and sale of AI technology, including for facial recognition software, could be particularly helpful. To this end the United States might consider joining existing efforts to promote collaboration in the digital space, including ethical AI like the Digital Economy Partnership Agreement (DEPA).

Finally, U.S. policymakers should continue to help deploy AI to advance global common goods, such as public health and climate change. Releasing some public data sets could help achieve some of these goals. This may also mean finding ways of cooperating with China, although this would have to be done in a way that addresses concerns over unfair competition and technology transfers.

2. Advance U.S. innovation in AI.

The mainstays of competitiveness in the field of AI are data, talent and computing power. To advance U.S. interests, policymakers should focus on ways to enhance comparative advantages in these areas but avoid creating cleavages with its partners.

To advance talent, several complementary strategies could be implemented or enhanced. First, many of the top graduate students and researchers focusing on AI in the United States are not U.S. citizens. Those individuals and their families should be able to come to the United States and should also be able stay and be encouraged to work for American companies and universities. Multiple industry and national security experts have called for this kind of immigration reform.

The COMPETES Act, passed by the House, includes provisions that would facilitate the acquisition of permanent residency status for individuals with advanced science, technology, engineering, and mathematics (STEM) degrees in certain fields. This is a small step that could be helpful in countering restrictive immigration rules that are undercutting U.S. competitiveness. However, it is far from certain that these provisions will be retained in the final version of the Bipartisan Innovation Act after reconciliation—or whether the ongoing reconciliation process with the Senate-passed U.S. Innovation and Competition Act (USICA) will be successful at all. Immigration is an area of unique comparative advantage for the United States since it attracts far more talent than any other country, including many researchers from China. However, immigration reform has become particularly toxic in the U.S. political system, and a comprehensive overhaul is unlikely.

A longer term but important step to enhance U.S. competitiveness overall is to improve the quality of and expand access to the education system. In addition to university and post-graduate-level scholarships and funding, the United States should also ensure that a larger number of students arrive in universities with the right skills to excel. To that end, more funding should be allocated to improve the K-12 system, including in underserved areas, and expand education in STEM, including data science, analysis, and even ethics. This is in line with the recommendations of the NSCAI from 2021. Ultimately, a robust general education is crucial to making informed decisions, and STEM education should not come to the expense of social science and humanities, which inform the field of ethics. This will help advance innovation and equality more broadly in the United States and yield significant, positive economic effects.

The United States has already ramped up spending significantly in support of the AI industry. Government contracts for AI have more than tripled between 2017 and 2020, providing a significant form of support for the industry and no doubt enriching algorithms with data. This is positive, but a review of the kind of spending and types of contracts would be helpful in evaluating what kind of further support the government can provide to its industry without duplicating or crowding out private investment. Moreover, the U.S. government should explore supporting a variety of different applications to avoid overconcentration of investment in areas like military applications and surveillance. Overall, the United States’ traditional lead in basic research can be a true advantage and can lead to game-changing breakthroughs if supported appropriately.

The other area of competition is data. Here, China’s large population and large numbers of participants in the digital economy is thought to provide a real advantage. This has long been recognized, and Yuchtman and Yang’s research provides further evidence. This is especially true in areas such as those explored in this feature, where companies can gain real commercial benefits from fulfilling public contracts. But, as noted before, there are limits to this advantage and having diversified data sources (through access to global markets) can be a real asset.

It may also be beneficial to find ways to support smaller companies to enter the field to prevent all AI research from concentrating in a handful of large firms. One way to do this is to provide more funding to support firms across the phase between R&D and commercialization, referred to as the “valley of death,” which is often problematic for U.S. companies. Once again, this is a solution that could be devised to help U.S. companies operating in emerging industries more broadly, including climate tech.

Finally, when it comes to computing power, the issues at play are linked more broadly to trends in the semiconductor industry, which is perhaps the most contentious industry in the U.S.-China relationship. When it comes to advanced chip design, American companies are still far ahead of Chinese ones—despite the latter’s quick advances. In terms of supply chains, the issue is far more complex, but the United States is generally in an advantageous position here as well. However, the Chinese government is taking muscular action to expand its supply chains as well as data center infrastructure to support computational power.

Any reconciled version of the USCIA and the COMPETES Act, if passed, would provide support for expanding semiconductor manufacturing domestically and strengthening supply chain resilience. Beyond that, however, the United States should be careful to support partners that are key players in the industry, including Taiwan and South Korea. While advancing domestic manufacturing could help diversify the industry in the long term, securing and enhancing relationships overseas will be fundamental in the short term.

3. Address China’s AI-surveillance symbiosis.

This may be the thorniest issue to address. While enhancing U.S. competitiveness may be politically challenging, many agree on the general actions that would be effective. When it comes to how to act to constrain China’s repressive political actions, its surveillance state, or its state capitalism model, there is less consensus on what is feasible and what would be effective.

Some export controls and bans on certain companies are in place. As of mid-2022, 19 Chinese AI facial recognition companies linked with surveillance efforts in Xinjiang have been placed on the Bureau of Industry and Security’s Entity List. But enforcement can be challenging. Moreover, while the federal government has been prohibited from purchasing products from Dahua and Hikvision, two major Chinese surveillance camera makers, on national grounds since 2019, some research indicates that U.S. companies have been selling relabeled Chinese cameras unwittingly to federal agencies as well as cameras that use Chinese components, including Huawei processors.

Better enforcement will be crucial to ensuring that existing rules work. Ultimately, expertise and capacity for detailed supply chain analysis rely largely on the private sector. Consequently, the onus of proof will have to fall on private companies, which will require an expansion of auditing and due diligence capacity for companies and monitoring and enforcement for the government. Similar challenges will lie ahead for any potential international export control regime introduced in this area.

Finally, it is worth noting that AI may expand the state’s surveillance capacity, but it does not change the underlying aims that predate the technology. China’s public security apparatus has long been very effective in tracking dissidents. The United States is already targeting firms supporting human rights violations in China, particularly in Xinjiang, and has included several companies on its Entity List. It should continue to do so and provide more resources for enforcement. But there are limits to the effectiveness of economic tools, as Russia’s continued invasion in Ukraine shows. The issue transcends AI policy, and policymakers will have to explore alternative diplomatic tools to address it, including, for example, providing further support for asylum seekers.

Cite this Page