China’s Uncharacteristic Approach to Artificial Intelligence (AI) Development
This policy summary, based on a working paper by Jeffrey Ding, shows how China’s AI strategy diverges from China’s typical approach to industrial policy, which stresses self-sufficiency, support for a limited number of national champions, and the essential role of military investment and demand for progress in dual-use domains. China has proved more flexible in cultivating its AI sector—utilizing more open-source software, an increased number of firms, and the commercial sector.
Artificial Intelligence (AI) Is The New, Most Crucial General Purpose Technology
AI is poised to have transformational effects on societies and national and industrial economies. Like electricity, AI is a technology that can be applied across a range of sectors for many different uses. Stimulating AI is therefore critical for countries to stay on the leading edge.
China is working to develop its AI sector, but the demands of general purpose technologies (GPTs) diverge from China’s usual industrial policy approach. GPTs require open flows of information, coordination between the GPT sector and a broad range of industries, and predominantly civilian-led development. China usually engages in techno-nationalism (employing protectionist policies and indigenous innovation to secure high-tech industries) and targets two or three main national firms to fulfill production. Analysts also generally regard the People’s Liberation Army (PLA) involvement as a boon for progress in certain dual-use technologies.
In short, GPTs impose demands that conflict with the conventional toolkit Chinese policymakers use to target technologies that can more easily be contained within sectoral boundaries. This policy note explores this conflict in three aspects of China’s AI policy: (i) AI open-source software, (ii) “picking winners” and the “national team” in AI; and (iii) the limited relevance of the military in AI development.
China’s Approach to Promoting AI as a GPT
1. China Uses Open-Source Software for AI
Open-source software supports GPTs, because GPT sectors and other organizations that provide complementary capital and skills (e.g., other firms or academia) rely on shared industry-specific data to fine tune algorithms and application scenarios. China traditionally hasn’t gone this route, for fear that this type of coordination is less salient and even detrimental to the nation’s competitive advantage, as the innovating firm could leak its technical secrets. For example, Made in China 2025 has been held up as a prime example of China’s ambitions to achieve certain self-sufficient targets in industries such as aviation and new energy vehicles. However, in AI, China has adapted its policy.
In March 2021, the phrase “open source” appeared in China’s Five-Year Plan (covering the years 2021–2025) for the first time. In a section about promoting the digital economy, the plan stated that China will “support the development of innovative consortia such as digital technology open source communities…encourage enterprises to open up software source code, hardware designs, and application services.” This coincides with a wave of government action in open-source software, including support for Gitee, an alternative to the popular developer collaboration platform GitHub.
Chinese planners see open-source software as a foundation for China’s AI development. In July 2018, the Ministry of Industry and Information Technology (MIIT) published a White Paper on China’s Development of AI Open-Source Software, which mapped the landscape of AI open-source software and put forward recommendations to improve the overall open-source ecology. Specifically, the White Paper pushed for cultivating a better open-source culture and improving licensing agreements for open-source software. “Open-source collaboration is especially central to China’s AI development strategy, which emphasizes the development of platforms where resources are openly shared.”
Open-source collaboration inevitably requires dependence on a global community of software developers, since the United States is home to 66 percent of the world’s AI open-source software developers, and indeed, Chinese technology companies’ participation in global open-source communities has grown exponentially over the past decade.
However, technonationalist impulses toward insulated development still shape China’s open-source strategy. MIIT has tried, with limited success, to push the aforementioned Gitee as a homegrown alternative to GitHub in order to insulate Chinese technologists from potential access restrictions. But Chinese open-source developers largely prefer GitHub to Gitee. Over 70 percent of people employed in the computer vision field reported using PyTorch, a popular machine learning framework originally developed by Facebook; only 6.5 percent had used any frameworks developed by Chinese firms and organizations. The corresponding figure was even lower among students in computer vision, with less than 2 percent reporting use of Chinese machine learning frameworks.
Box 1. China’s divergence from typical industrial policy in AI.
|Typical Chinese Policy
|Open flows of information
|Direct government intervention with few national firms
|Linkages between GPT sector and other organizations and industries
|Military involvement seen as necessary spur for dual-use domains
2. Expanding the AI National Team of Firms
For Chinese planners, the approach to “picking winners” in AI differs markedly from other industrial policies that champion two or three main firms. To begin, the number of so-called winners is much larger. In 2017, the Ministry of Science and Technology (MOST) designated four firms—Baidu, Alibaba, Tencent, and iFlytek—as members of a “national team” (guojiadui) in AI, tasking them to lead the development of AI open innovation platforms (AIOIPs). As AI advances have found applications in one industry after another, MOST’s approach to the AI national team has evolved toward more inclusivity.
In updated guidelines for the construction of national AIOIPs published in August 2019, MOST announced that companies could now apply to join the “team” by selecting a particular subdomain of AI in which they would invest, lowering the barrier of entry for small and medium enterprises, and the membership of this AI national team expanded to fifteen firms. As Benjamin Larsen concludes, “The AIOIP initiative is less about granting preferred access to a few select companies, and more about enabling structural mechanisms that afford greater participation and innovation in emerging ecosystems and sectors that increasingly will be powered by AI technologies.”
3. AI Innovation in the Commercial Sector
Although AI has many potential applications in the defense sector—and is therefore considered a dual-use technology—unlike in other dual-use domains, the development of AI in China was primarily incubated in the commercial sector. “The locus of innovation in AI has shifted to the private sector,” notes Elsa Kania, a leading scholar on Chinese military AI. “Consequently, Chinese leadership seeks to ensure that private sector progress in AI can be rapidly transferred for employment in a military context through a national strategy of military-civil fusion, while building up capability in its defense industrial base.” The effectiveness of this transfer will also be shaped by the overall military civil-fusion effort, which has been a protracted process to date.
This is not to say that military R&D and procurement have played no part in spurring China’s development of AI. The PLA is a source of demand for some Chinese AI companies—for instance, voice recognition and machine translation for counterintelligence and intelligence processing, as well as video surveillance for border defense and training exercises—and it does conduct some research on unmanned vehicles, which has spillovers to the commercial domain. On the whole, nevertheless, the private sector has driven the development of AI in China—a departure from traditional Chinese policy.
AI technology’s importance to the economy and society is demonstrated in China’s decision to adapt its policy in order to compete at the leading edge. China’s approach to promoting AI—often regarded as “the next GPT”—includes policy and institutional adaptations that mediate other characteristics of Chinese industrial policy. Specifically, China’s AI strategy diverges from expectations of typical characterizations of China’s industrial policy, which stress an emphasis on self-sufficiency, support for a limited number of national champions, and the essential role of military investment and demand for progress in dual-use domains.
As China transitions from initiating a new GPT trajectory in AI to spreading AI across broad and varied application sectors, increased attention should be paid to how Chinese policymakers adapt their sectorally-focused approach to a technological domain that permeates so many industries. Crucially, the optimal approach for spurring AI development, which emphasizes diffusion and openness, may conflict with other priorities in Chinese industrial policy, such as national control and security. GPTs like AI are not the same as other technologies, and they demand a different toolkit of strategies.
JEFFREY DING is an assistant professor of political science at George Washington University.