By Carl Benedikt Frey and Michael Osborne
Once upon a time, Japan was widely expected to eclipse the United States as the technological leader of the world. In 1988, the New York Times reporter David Sanger described a group of U.S. computer science experts, meeting to discuss Japan’s technological progress. When the group assessed the new generation of computers coming out of Japan, Sanger wrote, “any illusions that America had maintained its wide lead evaporated.”
Replace “computers” with “artificial intelligence,” and “Japan” with “China,” and the article could have been written today. In AI Superpowers: China, Silicon Valley, and the New World Order, which unsurprisingly became an instant bestseller, former Google China President Kai-Fu Lee argues that China’s unparalleled trove of data, culture of copying, and strong government commitment to artificial intelligence give it a major leg up against the United States. The Harvard University political scientist Graham Allison has recently argued that China’s embrace of what most Americans view as a nightmare surveillance state gives it a significant data advantage over the United States.
As scholars who study the applications and implications of artificial intelligence, we respectfully disagree. China, if anything, looks less likely to overtake the United States in artificial intelligence than Japan looked to dominate in computers in the 1980s. For while China is rich in data and has excelled in refining technology invented elsewhere, much impedes it from becoming the site of the next big breakthrough that artificial intelligence sorely needs.
DATA ALONE ARE NOT ENOUGH
China made international headlines by effectively leveraging its surveillance technology for contact tracing in response to COVID-19, the disease caused by the novel coronavirus. And yet the country’s alleged data advantage is hugely overblown. One reason is that data are highly domain specific and don’t often solve more than the problem for which they were gathered. China’s disregard for privacy enables it to snoop on its citizens, but not much else. And an abundance of surveillance data doesn’t give China an advantage in applying artificial intelligence to such ends as drug discovery or self-driving cars, for example.
The puzzle of artificial intelligence lies not in the quantity of data to which its algorithms have access but in the efficiency with which it learns from that data. Even with huge amounts of data, artificial intelligence systems are easily tricked into making errors. The Google researcher Christian Szegedy and his collaborators proved this point by fooling an algorithm that had once confidently and correctly classified images of dogs and school buses. The researchers manipulated the pixels of images in a manner that would have been completely undetectable to the human eye—but that led the algorithm to classify both dogs and school buses as ostriches. Artificial intelligence algorithms can often identify objects, but they lack any conceptual understanding of the relationships between those objects or of their respective properties. As the deep learning researcher Yoshua Bengio has warned, “We can’t realistically label everything in the world and meticulously explain every last detail to the computer.”
Artificial intelligence systems are easily tricked into making errors.
Many think of China as “the Saudi Arabia of data.” But if data are the new oil, they might just be China’s natural resource curse. For example, in the early twentieth century, electric cars looked more promising than gasoline-powered cars. Huge oil discoveries, among other things, tipped the balance in favor of the internal combustion engine. A century later, we are trying to get back into electric cars. The current focus on data-thirsty AI applications could lead to a similar lock-in into the wrong sort of AI.
We have seen this movie before. In the 1980s, the grand promises and overwhelming focus on symbolic AI prompted immense funding and media hype. This meant that funding for “deep learning” dried up. But deep learning has its own problems and has recently caused companies to focus on easy AI problems, such as classifying cats and dogs, where data are abundant. This approach alone is likely to run into diminishing returns that could even prompt another AI winter.
Data efficiency is the holy grail of further progress in artificial intelligence. The reason most people associate the steam engine with James Watt and not Thomas Newcomen (who developed a coal-powered steam engine decades earlier) is that Watt’s separate condenser first made the technology energy efficient. Artificial intelligence is still waiting for its separate condenser moment. Indeed, to learn enough to win a game of Go against Lee Sedol, a champion of the strategic board game, DeepMind’s AlphaGo software first had to play many millions of games against itself. It learned to play far slower than any human. Humans are incredibly data efficient; recent breakthroughs in artificial intelligence are much less so. Whether the United States or China will lead the world in artificial intelligence depends far less on who controls the most data than on who will be the first to innovate past this impasse.
EXPERIMENTATION DRIVES INNOVATION
Those who warn of China’s inexorable advance in the field of artificial intelligence worry that because the technology is by nature centralizing, authoritarian governments are better able to encourage AI innovation than democratic ones—and that AI technology, in turn, will advantage authoritarian governments. The concern recalls a belief about electricity that held sway a century ago—and like that belief, today’s is also misplaced.
In 1923, the pioneering electrical engineer Charles Steinmetz—whose work for the General Electric Company around the turn of the twentieth century made him a celebrity of the time—predicted that electricity would give rise to a more collectivist society. Steinmetz argued, somewhat circularly, that the development of a national electrical grid would lead to socialism, because only a socialist system could effectively manage the new interdependencies that progress toward a national grid would require. The Rural Electrification Act of 1936 did indeed provide funds to rural cooperatives that had been neglected by major private power companies. But the real transition to electrical power came out of capitalist competition, in the form of experimentation on the factory floor. When engineers figured out how to equip every machine with its own electrical motor, rather than relying on one central power source, they could sequence the machines according to the natural flow of production—a breakthrough that gave rise to mass production.
Decentralized experimentation and decision-making are critical to harnessing the benefits of AI.
Decentralized experimentation and decision-making will likewise be critical if the world is to harness the benefits of artificial intelligence. China is at a disadvantage in this regard. The country’s recent surge in patent filings is often cited as evidence of its innovativeness, but simply counting patents isn’t a good way to measure innovation: studies show that ten percent of patents account for roughly 90 percent of total patent value, meaning that the vast majority are of little value. Patent citations offer a more useful indicator, and if we look at the 100 most cited patents since 2003, not a single one comes from China. Moreover, China’s leading artificial intelligence companies, including Tencent, Alibaba, and Baidu, are merely copies of Facebook, Amazon, and Google, tailored to the Chinese market.
As the late economic historian Alexander Gerschenkron observed, when a country lags behind the technological frontier, imitation and the adoption of foreign technology can take it a long way—and, in general, the further a country has fallen behind, the greater the role the state must play in driving industrial catch-up. Thanks to state investment in mass production technology, the Soviet Union grew rapidly during much of the Cold War, as did Japan, South Korea, and Taiwan. Indeed, numerous scholars have attributed the “Asian Miracle” to state-driven industrial catch-up. But while they were successful in closing some of the gap, these countries never managed to overtake the United States. Unlike imitation, which can be planned and coordinated, innovation is a voyage of exploration into the unknown, to paraphrase the economist and philosopher Friedrich von Hayek. And switching from imitation to innovation is hard: if it were easy, most countries would be innovating at the technological frontier.
By observing that China is unlikely to overtake the United States in technological innovation, we mean in no way to downplay China’s tremendous economic achievements since Deng Xiaoping came to power in 1978. China has plenty of talent, but the fact remains that, so far, Chinese innovation has mainly focused on incrementally improving technologies that were conceived elsewhere. Chinese companies currently lead the world in the development of 5G, for example, but their work builds on several previous generations of telecommunications technology. What Huawei demonstrates is that China has significant engineering capabilities, just like Japan and indeed the Soviet Union.
DYNAMISM VERSUS STABILITY
Artificial intelligence is not yet a mature technology, and continued progress will require radical innovation on multiple fronts. Breakthroughs will happen the way they usually do: through serendipity and recombination, as inventors and entrepreneurs interact and exchange ideas. China’s strong state and collectivist structure have significant advantages in swiftly building infrastructure or mounting a coherent response to a pandemic. But radical innovation is a different matter, and historically, the most innovative societies have always been those that allowed people to pursue controversial ideas. As the eminent economic historian Joel Mokyr has argued, that is why the Industrial Revolution happened in the West rather than in China in the first place.
China’s efforts to restrict the flow of ideas on the Internet and elsewhere are likely to hold back innovation. Since September 2019, China and Huawei have been proposing radical changes to the Internet infrastructure that underpins networks worldwide. If implemented, the changes would likely splinter the Internet and further reduce Chinese citizens’ exposure to new ideas from outside the country. The initiative underlines Beijing’s preference for maintaining the political status quo, even if that means slower innovation and less dynamism.
That said, the United States is not destined to win the race for supremacy in artificial intelligence. China could still change its trajectory, and new immigration restrictions imposed by the administration of U.S. President Donald Trump could stifle innovation in the United States. Research shows that immigration has been a key driver of American innovation over the past 130 years. The Trump administration’s alleged plans to restrict H-1B visas is particularly worrying in this regard. But while Trump might hold on to power for another term, Xi Jinping could rule indefinitely.
Under Xi, the Chinese Communist Party has stepped up efforts to penetrate private-sector businesses and consolidate political power. A surveillance state with a censored Internet, together with a social credit system that promotes conformity and obedience, seems unlikely to foster creativity: innovation is about breaking the rules, not abiding by them. Indeed, a recent study published in the Proceedings of the National Academy of Sciences found that positive attitudes toward conformism and obedience predict less disruptive innovation.
Japan failed to overtake the United States, even without heavy restrictions on the flow of ideas and an authoritarian regime that promotes obedience. Hence, the United States has critical advantages that should enable it to remain the world’s leader in artificial intelligence. If it cedes that position to China, the reason will likely be that Washington has tried to emulate the Chinese model by propping up national champions rather than embracing the competition and dynamism that have made the United States the world’s technological front-runner for more than a century.
CARL BENEDIKT FREY is Oxford Martin Citi Fellow and Future of Work Director at the Oxford Martin School at Oxford University and the author of The Technology Trap: Capital, Labor, and Power in the Age of Automation.
MICHAEL OSBORNE is Professor of Machine Learning at the University of Oxford, a Fellow at the Oxford Martin School, and Co-Founder of Mind Foundry.
No comments:
Post a Comment