Beyond the AI Arms Race

The idea of an artificial intelligence (AI) arms race between China and the United States is ubiquitous. Before 2016, there were fewer than 300 Google results for “AI arms race” and only a handful of articles that mentioned the phrase. Today, an article on the subject gets added to LexisNexis virtually every week, and Googling the term yields more than 50,000 hits. Some even warn of an AI Cold War.

One question that looms large in these discussions is if China has, or will soon have, an edge over the United States in AI technology. Dean Garfield, the president of a U.S. trade group called the Information Technology Industry Council, recently told Politico that such fears are “grounded in hysteria.” But many prominent figures disagree. Former Alphabet CEO Eric Schmidt, for instance, warned in 2017 that “By 2020, [the Chinese] will have caught up [to the United States]. By 2025, they will be better than us. And by 2030, they will dominate the industries of AI.” And former Deputy Defense Secretary Bob Work, among others, has argued that China’s advances in AI should spark a “Sputnik moment” for the United States, inspiring a national effort comparable to the one that followed the Soviet Union’s early victories in the space race.

One of the most in-depth discussions of U.S.–Chinese AI competition can be found in AI Superpowers: China, Silicon Valley, and the New World Order, a recent book by the Taiwanese-American venture capitalist and AI expert Kai-Fu Lee. Drawing on his in-depth knowledge of—and personal experience with—the Chinese and U.S. tech sectors, Lee, like Schmidt and Work, concludes that “China will soon match or even overtake the United States in developing and deploying artificial intelligence.”

There is much to like about AI Superpowers. Lee persuasively counters many common misperceptions about China and AI, and he offers thoughtful personal reflections. Unfortunately, the parts of the book that inform his core thesis about China’s relative advantages over the United States are the parts that leave the most to be desired. In the end, Lee’s overly rosy portrayal of China’s AI capabilities both detracts from the book’s other contributions and risks feeding the zero-sum arms race thinking that he himself warns could hurt humanity’s ability to harness AI for good.

AI AND THE GLOBAL ORDER

AI Superpowers unfolds in three parts, each with its own core argument. The first part is about the Chinese technology sector, a topic that Lee—who led Microsoft Research Asia and Google China before founding Sinovation Ventures, one of the biggest venture capital firms in China—is uniquely qualified to talk about. He traces the sector’s evolution from its initial “copycat” approach, in which it sought to replicate Western technologies and business models, to its current state of cutthroat competition and innovation, which makes Silicon Valley look boring by comparison. Along the way, Lee details how China’s tech giants, such as the e-commerce platform Taobao, triumphed against competition from established foreign firms such as eBay.

The core reason for Chinese technology firms’ success, according to Lee, is that they are willing to engage with the nitty-gritty of the real world. They prefer messy vertical integration, in which one firm controls everything from design to production, retail, and marketing, to Silicon Valley’s cleanly siloed digital platforms. Compare the Chinese food delivery giant Dianping—which started out as a reviews site but soon dove into delivery, providing everything from a payment platform to a scooter-riding delivery staff—with the U.S. firm Yelp, which initially offered delivery services that mostly relied on restaurants’ own infrastructure before retreating to focus on its function as a review app. This difference is partly cultural, but it also stems from necessity: intellectual property protections are weaker in China than in the United States, so to deter competitors Chinese firms have to invest huge amounts of capital in building hard-to-replicate physical infrastructure. In the United States, by contrast, Lee believes that the combination of intellectual property laws and a cultural distaste for anything that smacks of copying often allows Silicon Valley companies “to coast on the basis of one original idea or lucky break,” leading to complacency.

The second part of the book develops Lee’s argument about the U.S.–Chinese technological rivalry, which we will return to below. In the third part, he deals with the socioeconomic effects of AI. Here again, Lee is full of insights. He first takes issue with the argument that the economic impact of AI will resemble that of other general-purpose technologies, such as the steam engine and electricity, which converted many previously high-skill tasks (such as handcrafting textiles) into low-skill tasks (such as operating a steam-driven power loom), thereby creating new employment opportunities for large numbers of lower-skill workers. In contrast to this “de-skilling” effect, the trend with information technologies has generally been to increase the productivity of high-skill workers, often by reducing their reliance on other auxiliary roles—think, for example, of the decreasing need for secretaries as e-mail and word processors have replaced telephones and typewriters. Lee argues that this “skill bias” in the economic effects of AI will leave large swaths of low- and middle-skill workers unable to contribute productively to the economy. The AI revolution will also come much faster than previous technological revolutions, mainly because software, in contrast to steam engines, can be instantly replicated and distributed across the world. Lee predicts that 40–50 percent of U.S. jobs will be capable of being automated within the next couple of decades, which will probably increase unemployment rates by around 20–25 percent.

AI also differs from many past technologies in its natural tendency toward monopoly, thanks to the self-reinforcing cycles that strengthen the best companies: the more users a company has, the more data it can access, allowing it to develop a better product that will attract even more users. Lee argues that this tendency toward monopoly will exacerbate both domestic and international inequality: domestic, because of the rise of “superstar firms” that tend to decrease labor’s share of national income; international, because most AI talent and resources are concentrated in China and the United States. (The auditing firm PricewaterhouseCoopers predicts that nearly $16 trillion in GDP growth could accrue from AI by 2030, of which 70 percent will go to the United States and China alone.) Lee sees the main threat from AI as “tremendous social disorder and political collapse stemming from widespread unemployment and gaping inequality.”

THE RACE FOR AI SUPREMACY

Lee’s treatments of the Chinese technology sector and the socioeconomic impacts of AI are both excellent. But it is the remainder of the book that has received the most attention: the part dealing with the U.S.–Chinese struggle for AI supremacy, in which Lee argues that China is poised to catch up with or overtake the United States in most areas of AI. Unfortunately, this is also the section that leaves most to be desired, both in what it argues and what it leaves out.

China’s particular potencies, according to Lee, include “abundant data, hungry entrepreneurs, AI scientists, and an AI-friendly policy environment.” Lee points to two underlying trends that play to these strengths. The first is the shift from “the age of discovery” to “the age of implementation.” That is to say that, according to Lee, “much of the difficult but abstract work of AI research has been done.” He claims that recent progress in AI rests almost entirely on a single breakthrough—deep learning, a class of algorithms loosely inspired by the interconnection of neurons in the brain—and that all that is left is for “entrepreneurs to roll up their sleeves and get down to the dirty work of turning algorithms into sustainable businesses.” This will benefit China by making its disadvantage in fundamental research less important while playing to its strengths—entrepreneurialism and the country’s vast quantity of good-enough engineers.

Lee’s view that today’s progress in AI is the product of a single breakthrough, however, is by no means a consensus one, and he does little to defend it. In fact, the conceptual foundations of deep learning were laid decades ago, with later researchers finding clever new ways of implementing existing concepts and exploiting the capabilities afforded by faster hardware and larger data sets. In other words, far from being a one-time breakthrough, deep learning has been advanced by a series of small to medium-sized innovations, which are continuing to this day. This dynamism undermines Lee’s argument that the “age of exploration” has ended. In all likelihood, labs that focus not just on applications and products but also on more fundamental research—an area in which Lee concedes that the United States has a strong advantage over China—will continue to push the envelope in ways that profit-hungry entrepreneurs cannot.

The second trend that Lee sees tilting the playing field toward China is a shift from “the age of expertise” to “the age of data.” Although he acknowledges that the United States has more elite researchers than China does, he claims that in today’s world data is more important than top talent, “because once computing power and engineering talent reach a certain threshold, the quantity of data becomes decisive in determining the overall power and accuracy of an algorithm.” And China, as the “Saudi Arabia of data,” has much more of it than the United States. The latter claim is indisputable: China has more Internet users than the United States and Europe combined. These users rely on “super-apps” such as WeChat for everything from booking doctors’ appointments to filing taxes, and provide the apps with additional data by using them to process the huge number of mobile payments in China’s virtually cashless economy.

Yet just how important data is—and how well Chinese companies can utilize the data they have access to—remains an open question. Andrew Ng, one of the world’s leading AI researchers, has argued that “big data is overhyped,” since many problems don’t produce big enough data sets to train AIs on. And insiders at Tencent, the owner of WeChat, have revealed that the company struggles to integrate data streams gathered by its different internal departments due to a combination of bureaucratic and technical hurdles.

Exaggerating Chinese advantage risks worsening the increasingly zero-sum conversation around AI competition.

More generally, a big problem with the common comparison between data and oil (implicit in Lee’s “Saudi Arabia of data” analogy) is that the utility of any data set is mostly limited to questions relating to that specific data set. It’s true that non-Chinese companies would be hard pressed to match their Chinese counterparts at, say, predicting which news stories Chinese consumers will be interested in. And facial recognition based on street-level surveillance footage is likely to remain a stronghold of Chinese companies, given the (highly concerning) amount of effort currently directed there. But it is far from clear that these specific advantages will help China make progress in the domains that policymakers often talk about, such as AI’s effect on the military balance of power. Perhaps because Lee’s focus is on novel commercial products and services, he is mostly silent on the question of why we should expect these narrow advantages to have the large geopolitical implications the book alludes to.

Lee also provides an overly rosy description of China’s top-down, government-sponsored innovation projects. The state has poured billions into developing technologies such as semiconductors and robots—efforts that Lee sees as part of a “techno-utilitarian” approach that is extremely effective despite some inefficiencies. Yet China’s overhyped government plans usually under-deliver. For instance, the analysts Gregory Allen and Elsa B. Kania have cited China’s $150 billion semiconductor fund as evidence of the country’s edge in spurring the development of strategic industries. But the fund has invested only $12 billion since its establishment in 2014. (For context, the South Korean firm Samsung spent nearly $27 billion in capital expenditures for its semiconductor group in 2017 alone.) Finally, even if money does get spent, the effect can even be counterproductive: historically, Chinese science and technology megaprojects have often furnished pretexts to divert funds from high-quality labs toward more politically connected entities.

To be clear, the picture that Lee paints is not entirely one-sided. He argues, for example, that the United States has a decisive advantage in many business applications of AI due to its corporate culture and more standardized data practices, and that American firms have a sizeable lead in autonomous vehicles. Yet Lee seems so eager to counter those who underestimate China’s AI capabilities that he tends to overestimate them instead. This is understandable. But exaggerating Chinese advantage risks worsening the increasingly zero-sum conversation around AI competition that Lee’s book is a part of—whether he likes it or not.

RESETTING THE ARMS RACE NARRATIVE

Lee is clearly aware of the risk of accelerating a downward spiral in the conversation around AI, and he explicitly tries to address this risk in his final chapter. He argues that the rhetoric around AI races could lead to zero-sum thinking that undermines progress and the ability of China and the United States to reap mutual gains. Instead of militarizing AI, Lee wants us to see that the technology’s true value “lies not in destruction but in creation”; he suggests that readers should think of international AI competition as “more akin to the early export of steam engines and lightbulbs than as an opening volley in the global arms race.”   

These arguments, unfortunately,come as too little, too late within the context of the book. Given the book’s title and overall tone, it is hardly surprising that coverage of it has focused on its predictions of geopolitical competition. A headline in The Economist summarized Lee’s thesis as: “In the struggle for AI supremacy, China will prevail.” If Lee had wanted to avoid fanning these flames, he would have done much better to linger on U.S. as well as Chinese strengths, and to acknowledge just how much uncertainty still exists about AI and its impacts rather than speculatively filling in gaps in ways that make China look good.

Ultimately, however, Lee is correct in cautioning against intense arms race dynamics. There are, to be sure, many things the United States can and should be doing to reap the gains from AI, such as funding basic research and development, taking measures against illicit technology transfer, and crafting a supportive regulatory approach. But a full return to a Cold War mindset comes with many significant risks, such as prematurely deploying accident-prone weapons systems or inadvertently proliferating dangerous technological capabilities. This would, in all likelihood, harm rather than help national security, to say nothing of society at large. If this is the path that policymakers choose to follow, they should do so with a full appreciation of the dangers and on the basis of solid intelligence. The stakes are much too high to act on the basis of fear and speculation alone.