China’s AI Unicorns Can Spot Faces. Now They Need New Tricks

A warehouse in an industrial park about an hour’s drive north of downtown Beijing offers a paradoxical picture of China’s much-hyped, and increasingly controversial, artificial intelligence boom.

Inside the building, a handful of squat cylindrical robots scuttle about, following an intricate and invisible pattern. Occasionally, one zips beneath a stack of shelves, raises it gently off the ground, then brings it to a station where a human worker can grab items for packing. A handful of engineers stare intently at code running on a bank of computers.

The robots and the AI behind them were developed by Megvii, one of China’s vaunted AI unicorns. The impressive demo might seem like further evidence of China’s AI prowess—perhaps even proof that the country is poised to eclipse the US in this critical area. But the warehouse also points to a fundamental weakness with China’s AI. Amazon has been using similar technology in US fulfillment centers for several years.

China’s AI champions have spun AI algorithms into gold in recent years, but that may become more difficult as the technology becomes more widely available. Megvii, a private company that CB Insights says is valued at around $4 billion, hopes to convince customers to buy its warehouse and manufacturing AI technology as it looks to move beyond a business built largely around facial-recognition technology. The trouble is, AI is not yet proven as a general-purpose technology that can easily be applied to different industries. Broader challenges, including newly imposed US trade restrictions, will make things even more difficult.

“These companies are not going to be big companies, like Alibaba or Tencent,” predicts Nina Xiang, a business journalist in Hong Kong and author of Red AI, a recent book about China’s AI boom. “They will remain small operators, and some valuations will have to be corrected.”

In October, Megvii and five other AI-focused Chinese companies were added to a US export blacklist, because Chinese authorities allegedly use their technology to monitor and control Muslim minorities in Xinjiang, a province in western China. The blockade means these companies can no longer buy crucial components such as advanced microchips from US firms.

China’s AI boom has produced more than a dozen unicorns, private companies valued at more than $1 billion. These include SenseTime, valued at $7.5 billion, the company’s CEO told Bloomberg earlier this year, and Yitu and CloudWalk, both valued above $2 billion. Another prominent AI company, iFlytek, has been around longer, having started out making speech-recognition tech, and it carries a market capitalization of $10 billion on the Shenzhen stock exchange.

Megvii, which filed to go public on the Hong Kong stock exchange in September, has genuinely impressive AI expertise, having developed core algorithms and software. It was founded by several graduates of a renowned AI program at Tsinghua University in Beijing. The company’s IPO filing offers a rare insight into the finances of a Chinese AI giant, and highlights just how dependent the company seems to be on face recognition and surveillance for now. Revenue grew four-fold last year to $200 million, compared with 2017; but its “City IoT” segment, which encompasses surveillance and security systems, accounts for nearly three-quarters of that revenue.

State-led development may be both a blessing and a curse for China’s AI enterprises. When the government announced a grand national AI plan in July 2017, it served as a signal for Chinese cities and provinces to pour money into AI projects. Xiang says Megvii and other Chinese AI unicorns seem to be heavily reliant on government contracts, subsidies, and other forms of strategic support. “On average, we can say a significant share of these companies’ revenues is government-reliant,” she says.

The Chinese AI companies have made efforts to move into new areas over the past few years. Besides Megvii’s move into logistics and manufacturing, Yitu touts its work in medical imaging and document analysis, SenseTime is investing in autonomous driving, and iFlytek often demos tools for analyzing legal documents. The catch is that AI is relatively unproven in such areas, and it’s unclear how much revenue the companies have generated from these ventures.

“Applying AI to business requires skills that are more artful,” says Qiang Yang, a professor at the Hong Kong University of Science and Technology and chief AI officer at WeBank, a banking startup founded by Tencent. He says a company needs to understand how to use AI tools to solve real-world problems, how to gather sufficient high-quality data, and how these challenges fit into the business life cycle. “This is hard,” Yang adds.

“The largest problem these companies face may be the dawning realization on investors that, although it seems promising, in most areas AI just isn’t ready for the big time,” says Helen Toner, of the Center for Security and Emerging Technology at Georgetown University, who has studied the development of AI in China.

There’s a technical reason for the predicament. Chinese AI companies built early success by applying deep learning, an AI technique that has dramatically improved machine perception in recent years, to problems like facial and speech recognition. Now, as deep learning becomes more broadly accessible through software packages and APIs, these companies need to expand into areas that require greater domain expertise.

Facial recognition has been particularly lucrative for Chinese companies, and the technology is widely used across the country. A report issued by IHS Market last week concludes that a billion surveillance cameras will be in operation worldwide by 2021, with about half of them in China. SenseTime, for example, recently deployed a system at Beijing’s new Daxing airport for China Eastern airlines. This uses facial recognition to let passengers check in, pass through security, enter the business lounge, and even board a plane without showing a boarding pass.

Megvii’s facial-recognition technology lets people unlock phones made by Oppo, Xiaomi, and Vivo and log into apps with a glance; it’s also bundled with security cameras that automatically check employees into office buildings. Like other Chinese AI companies, Megvii also supplies this technology to police departments that use it to hunt for criminals in surveillance footage. The company’s tech was being used by the authorities in Xinjiang, although Megvii says a developer used its application programming interface without the company’s knowledge.

Megvii declined to comment, citing a quiet period around its planned IPO. (The warehouse demo took place earlier.) Kang Ho, a spokesperson for SenseTime, challenged the idea that facial-recognition technology is now more broadly available. Still, Kang pointed to a range of ongoing projects in other fields, including tools for medical imaging, education, and virtual reality.

There are other signs that China’s AI bonanza, supposedly built on huge quantities of data and government backing, may be less spectacular than often assumed. A report published last week by the analyst firm IDC and Qbitai, a Chinese media company, found that 60 percent of executives surveyed expect significant difficulty deploying AI due to poor-quality data and a scarcity of AI talent.

Andrew Grotto, a professor at Stanford who coauthored a recent report on the financial details of China’s AI industry, agrees that these AI unicorns face significant challenges. Their true value “is a topic of debate in China,” he says.

China’s crop of well-funded AI-centric companies is unusual. The US’ big AI players, like Google, Facebook, Amazon, and Microsoft, all have existing businesses, like advertising, ecommerce, or software licensing, to bankroll their AI efforts. And while China’s own tech powerhouses—Alibaba, Tencent, and Baidu—are also heavily invested heavily in AI, the country’s tech market has been flooded by companies touting AI itself as a business in recent years. “Some Chinese firms, like Tencent, are among the very best in the world,” Grotto says. “But there are a whole lot of pretenders, too.”

The hype around Chinese AI certainly seems to have backfired if it helped prompt the US government’s export ban on Megvii and others. The White House appears concerned that China could soon steal an advantage in this critical area of technology.

It would be helpful for Megvii and others to diversify away from surveillance technologies now under scrutiny, says Rebecca Fannin, author of Tech Titans of China. Fannin adds that becoming less reliant on the West for advanced technology will benefit China long-term, but “could be a challenge” for the targeted AI companies.

Even if Megvii can build a major new business supplying AI-powered robots to manufacturers and ecommerce companies—and even if other Chinese AI companies find their own successes—a technological “decoupling” of China and America may affect companies in both countries in unforeseen ways.

The latest tit-for-tat measure saw the Chinese government order this week that US computers and software be replaced with Chinese technology in official buildings over the next few years. Whether or not a trade agreement is reached between China and the US, the emergence of a more cautious and uneasy relationship seems inevitable. “Beijing’s announcement is a harbinger of things to come,” says Grotto of Stanford.

More Great WIRED Stories

Read More