Why China's AI players are struggling to evolve beyond surveillance

BOSTON/PALO ALTO, U.S./HONG KONG -- All around a small city near China's border with Vietnam, posters and billboards carry the slogan "Building Smart City." Surrounded by hills and lakes, the town -- whose name the Nikkei Asian Review is not publishing to protect its sources -- is still a relative backwater, in one of the least developed regions of the country.

Most of the residents did not get access to the internet until the late 2000s. In a corner office on the fifth floor of the city hall, though, they now have a dedicated Department of Smart City Development. The official in charge, who gave his name as Cao, is in his 50s and near retirement. His background is in architecture, and he had no experience in technology until he was promoted to head the department in early 2018.

"I have no idea what 'smart city' or 'artificial intelligence' is. But I know the goal is to build an 'artificial intelligence-powered smart city,'" Cao said. In practice, that means most of his days are spent listening to pitches from private companies. "It is the companies' job to come up with the right projects or products that can help my department to reach that goal," he said.

China's national and local governments have rolled out the red carpet for domestic companies developing artificial intelligence, as Beijing tries to gain a strategic lead over other nations in a new technological arms race. The approach, which has seen the public sector provide money, contracts and, crucially, access to enormous quantities of data, has allowed local players to take huge strides.

Between 2016 and 2018, Chinese companies filed more AI-related patents than their U.S. counterparts. In 2018 alone, the country was responsible for 30,000 new patent filings in artificial intelligence: a tenfold increase in just five years, according to research by Nikkei released in March.

But, impressive as China's technological surge has been, it is far from clear that the country will, as its government expects, become an AI superpower in the near term. For all the hype, China's AI industry is constrained at home by a shortage of talent, hardware, and genuine innovation. And overseas, it is limited by its companies' close links with the state, which has fed suspicions about their motives and ethical positions.

"Private AI companies in China leverage strategic support from the government to form a collaborative national organism of AI," said Vilas Dhar, a U.S.-based venture investor and AI expert. "This integrated ecosystem makes them highly competitive, but also spurs scrutiny by the rest of the world."

State support

China's State Council laid out its ambitions in artificial intelligence in 2017, with a road map envisioning an industry worth 1 trillion yuan (over $140 billion) by 2030. The Ministry of Science and Technology identified a "national team," with a first group made up of internet giants Baidu, Alibaba Group Holding and Tencent Holdings, together with voice recognition intelligence company iFlytek, to lead the country's development of next-generation artificial intelligence technologies.

The announcement was sparse in detail about specific support for the companies, and the amount of money that would be made available. However, on a practical level, it seems to have meant that the government is willing to be the anchor client for major AI projects, creating a market for advanced applications of the technology.

The Ministry of Education drafted its own "AI Innovation Action Plan for Colleges and Universities," calling for 50 world-class AI textbooks, 50 national-level online AI courses, and 50 AI research centers to be established by 2020. By the first half of 2019, China had 35 college-level AI research centers. No data is available on how close the country is to achieving its other goals.

China's AI policy calls for parity with world-leading countries by 2020, major breakthroughs by 2025, and attaining the status of a global AI innovation center by 2030. (Photo by Ken Kobayashi)

Across government, public servants were incentivized to bring AI into their work, and to promote AI-related projects in their jurisdictions.

"The adoption rate of AI technology in the city has become an important measure in evaluating government officials' work," said Wang Lei, a professor at the College of Electronics and Information Engineering at Tongji University and expert in Chinese artificial intelligence development. "Experience in leading AI-related works has become a must-have if you want to get promoted to a higher-level position."

Motivated by the prospect of promotion, local government officials are racing to put out more business-friendly policies and initiatives to attract related investments, according to Wang.

But besides policy support and funding, the government also granted Chinese AI companies something else crucially valuable: data.

Data is the raw material for artificial intelligence. The larger and more comprehensive the datasets, the more effectively companies can train their algorithms. The Chinese government, which has access to enormous amounts of information collected on its citizens, has opened up the resource to its champion companies.

Few have benefited more than Ping An Technology, a subsidiary of the giant Ping An Insurance whose 2018 revenues totaled over 1 trillion yuan. Since its foundation in 1988, the company has built an enormous consumer dataset, and it spun out Ping An Tech in 2008 to leverage that using AI. Ping An Tech's activity now spans financial services, health care, automotive, real estate technology and smart cities.

Ping An Insurance spun off its tech business in 2008, using its troves of consumer data to inform AI development.   © Reuters

In 2017, Ping An Tech entered into a collaboration with the Centers for Disease Control and Prevention in Chongqing and Shenzhen. The authorities opened up a vast dataset -- 20 million health and medical records -- to the private company, which used the data to train an artificial intelligence platform, dubbed PADIA, to predict outbreaks of disease. PADIA is now able to predict with 90% accuracy the epidemic status of influenza, hand, foot, and mouth disease and the risk of chronic obstructive pulmonary disease.

The sheer scale of the data available to Ping An Tech gave it an advantage that few international companies could match. By contrast, in 2016, the U.K.'s National Health Service signed a contract with Google-owned artificial intelligence company DeepMind, aiming to use AI-informed scans to detect eye disease with an accuracy on par with, or exceeding, human experts.

The project is currently stalled. DeepMind so far has access to only 1.6 million health records provided by the NHS, which experts say is not enough to train an effective AI model. Moreover, the partnership has attracted controversies and lawsuits, with some patients denying having given their consent for their data to be shared.

Ping An says that government agencies provided them with raw data that concealed all personal information, meaning the company only has access to a collective, anonymized dataset where data points cannot be traced back to individuals.

"The machine learning process that trained the AI model is looking for patterns in the data, so individual personal information was really not needed," said Xu Liang, deputy chief engineer at Ping An Tech.

Xu said the AI model is also used in Shenzhen to help local government agencies read and extract key information from documents, and can reduce processing time by 95%. Several provinces in China are using it to detect social welfare fraud, and it has already detected instances of health care fraud, saving nearly 1 billion yuan.

Having built a commanding presence at home, China's AI giants want to take their technology to other markets -- and, in fact, they may need to. Government clients aside, the domestic market for AI is relatively limited in China.

Developing AI platforms is important, but commercialization of AI is driven by actually devising use cases for the technology, said Varun Arora, partner in the digital transformation practice at consultancy A.T. Kearney. While the vast scale of the Chinese market means that there are consumer applications, the real money in AI in the near term will be in industrial and commercial settings.

"One big reason why use cases have not flourished in China, besides, of course, a few big ones, is because of cheap and quality labor in the country. Ultimately, AI business cases will be based on displacement of human capital, and the U.S. is a more ripe market for that compared to China," Arora said.

Facetime

Close relationships between AI companies and the Chinese government have allowed them to flourish, but the same cooperation has proved a massive detriment to their expansion into other markets. Governments, companies and consumers fear that giving Chinese AI players access to data and systems is tantamount to handing them over directly to the Communist Party.

"There is not really a distinct private sector in China," said Dhar, the U.S. venture investor. "What is really happening in China is one very large set of people, data, algorithms that are moving fluidly between government and private companies."

This nervousness is particularly keen in the subsector where Chinese technology companies are probably furthest ahead of their international rivals: facial recognition.

Facial recognition has developed rapidly over the past decade or so, due to advances in so-called deep learning, which allows computers to more accurately read and interpret images. In China, public and private actors saw huge potential in the technology, from face-scan verification, used in place of biometrics or passwords, through to the police's desire to scan public spaces, such as airports and train stations. That market led to the development of at least four "unicorns" -- private companies with valuations of more than $1 billion -- Megvii Technology, SenseTime, CloudWalk Technology and Yitu Technology.

From making payments at stores to going through airport customs, many tasks in China can now be easily done through face scans. On Dec. 1, new rules came into effect in China requiring anyone who registers a new mobile phone number to submit to facial scans by their telecom carriers. Driven by the growing ubiquity of the technology, there have been more than 900 facial recognition patents filed in China -- nearly 10 times more than the number of patents filed in the U.S., according to data analytics company CB Insights. Chinese startups have dominated the U.S. National Institute of Standards and Technology's rankings of facial-recognition accuracy, according to a report published in 2018.

Cameras capture jaywalking pedestrians in Shanghai. (Photo by Ken Kobayashi)

Chinese tech companies are seeking to turn that technological dominance into international clout. The Financial Times recently reported that telecommunications equipment maker ZTE, security camera maker Dahua Technology, and state-owned China Telecom are among those proposing new global standards in the U.N.'s International Telecommunication Union (ITU) for facial recognition, video monitoring, and surveillance of cities and vehicles.

The biggest client for facial recognition technology in China has been the government, which has employed it in its defense and police establishments. That has meant that private companies have become entangled in highly controversial government policies, including the police crackdowns on pro-democracy protests in Hong Kong, and the incarceration of more than a million people from the Uighur minority in "re-education" camps in Xinjiang province.

Founded in 2011, Megvii alone has raised more than $1.3 billion, primarily from Chinese investment funds and companies -- including e-commerce giant Alibaba -- and filed for an initial public offering in Hong Kong in August. While pitching to prospective investors in Hong Kong this past summer, the company touted its public security contracts with the Chinese government as a selling point. 

Then, in October, Megvii was put on a U.S. trade blacklist over what Washington said was Beijing's treatment of Uighurs and other predominantly Muslim ethnic minorities. The U.S. blacklisted a total of 28 Chinese public security bureaus and companies, including facial recognition startups SenseTime and Yitu, video surveillance specialist Hikvision and voice recognition giant iFlytek, all of which are trying to expand overseas to escape an increasingly saturated domestic market.

However, while Chinese cities have been rolling out facial recognition technologies, and other Asian nations have embraced Chinese technology to enable their security apparatuses, Western governments have labored over how to regulate the emerging sector, which has created a multitude of new civil liberty concerns.

In May, San Francisco became the first city in the U.S. to ban the use of facial recognition technology by the city's municipal departments, including the police. In the U.K., a citizen has taken a local police force to court in what could be a test case for the use of the technology. This speaks to a profound ideological gulf in the way that artificial intelligence will be developed and used.

"A constitutional right we have [in the U.S.] is innocent until proven guilty. Facial recognition could flip that idea around," said Mutale Nkonde, a Harvard fellow and AI policy adviser.

Ren Zhengfei, founder and CEO of Huawei Technologies, articulated the difference in views in a panel interview in September. "I believe the protection of personal information should not come at the expense of public security," Ren said. "Western societies have protected privacy at all costs, but the trade-off is its public safety. Just look at how many mass shootings have happened in the U.S. You can protect someone's privacy, but what about the safety of the others?"

There are signs, however, that even in China citizens are becoming concerned about overuse of facial recognition. In November, the Southern Metropolis Daily reported that Guo Bing, a lecturer at Zhejiang Sci-tech University, was suing a local wildlife park for breach of contract, after it changed its biometric ticketing system to one using facial recognition. Guo was reportedly alarmed that the system would increase the risk of identity theft, and asked for a refund, which the park refused to give him. It is the latest of a small but growing number of private actions against companies using the technology.

"People in China are also getting increasingly concerned about the power companies have as they are collecting billions of people's biometrics data every day," said Danny O'Brien, strategy director of the nonprofit Electronic Frontier Foundation.

In September, a new Chinese app, Zao, which lets users "swap faces" with celebrities -- or anyone else -- in video clips, drew millions of downloads immediately after launching. But it also quickly sparked debates over privacy issues, as its user agreement allows the company to store pictures taken and use them for profit. The Chinese Ministry of Industry and Information Technology met with company executives within a week of the app's launch, and asked the company to revise the agreement to require users' consent before using the images.

Cases like this show the progress China has made in protecting consumer data and privacy, according to Yi Zeng, head of AI ethics and safety at the Chinese Academy of Sciences. He added that a stipulation on "the proper use of citizens' data" was also written into China's national AI strategy.

However, while the Chinese private sector is under increasing pressure to be responsible with user data, citizens have few avenues to challenge the government's use of their personal information. According to articles 7 and 14 of the National Intelligence Law of the People's Republic of China, Chinese citizens are required to "support, assist and cooperate with the state intelligence work," which leaves enough leeway for the authorities to collect and use citizens' data without informed consent.

Fundamental flaws

The dramatic growth of China's AI sector, measured by the sheer number of patents that the country produces, masks the fact that, in terms of fundamental research, it is still lagging behind the U.S.

"One critical question left to answer for China is whether its AI ecosystem can produce big breakthroughs in fundamental AI research -- the cornerstone of U.S. structural advantages in this space," said Jeffrey Ding, China lead for the Governance of AI project at the University of Oxford's Future of Humanity Institute.

Earlier this year, the AI specialist Lux Research performed an analysis of highly cited researchers in scientific papers focusing on cutting-edge artificial intelligence. They found that out of 2,100 highly cited researchers, only 158 came from Chinese research entities. Of those, just 37 came from the private sector.

This deficiency is in a large part down to the industry's structure. American AI development is market-driven and dominated by independent, private sector players. Their need to focus on commercially viable applications has meant that U.S. companies have taken leads in more industrial and business-process AI, including in autonomous vehicles.

China's centralized, top-down approach has helped its AI industry focus on specialist areas and created a market for them, but it may also be stifling innovation.   © Reuters

By contrast, China's approach is centralized and top down. While that has helped the sector to focus on specialist areas and created a market for them, it may also be stifling innovation. "The openness of the U.S. technology ecosystem -- to new ideas, new people, and new debates about AI ethics -- provides the bedrock for its AI advantage," Ding said.

The U.S. also has an advantage in hardware: Most of the cutting-edge chips being used in fundamental AI research are designed in America, and some are already finding their way onto lists of restricted technology exports.

"The U.S. is still way ahead of us in artificial intelligence in terms of research," said Xu Liang at Ping An Tech. "China needs time and talents to catch up, and it's critical to catch up before our development of AI [is] limited by fundamental infrastructures such as computing powers."

That said, the lack of a collective goal and a unifying strategy in the U.S. could ultimately hamper the sector, as the technology becomes more mature -- and, while the private sector-led approach may, on a superficial level, appear less open to abuse than one backed by the security apparatus, it is not without risks. Big tech companies in the U.S. are already under regulatory pressure over their market dominance, and huge ethical questions about their collection and use of data remain unresolved.

"If in the U.S., it continues to be only private companies extracting and using consumer data for profit and leveraging artificial intelligence for private gain, without accompanying a unified governmental position that guides them for social good, that raises a lot of alarm bells," said Dhar.

Additional reporting by Nikkei staff writer Coco Liu in Hong Kong.

Correction: An earlier version of this article misstated Megvii's description of its business activities in China. This article has been amended to reflect information contained in its IPO filing documents.