What Previous Industrial Revolutions Can Reveal about the U.S.-China Race for AI Leadership
Jeffrey Ding, a postdoctoral fellow at Stanford’s Center for International Security and Cooperation and at the Stanford Institute for Human-Centered Artificial Intelligence, argues that China and the United States are both taking the wrong lessons from previous industrial revolutions.
As revolutionary as artificial intelligence may be, he says, both governments are overly preoccupied with being dominant in breakthrough advances. Over the long term, it may be more important to figure out how a wide range of industries can make practical use of them all. In this interview, he explains why.
Attend HAI's Weekly Seminar with Jeff Ding This Month to Learn More About The Rise and Fall of Great Technologies and Powers
Why study prior industrial revolutions to understand today’s competition in AI?
It’s not just the media or analysts who are calling AI the next major revolution. China’s president, Xi Jinping, has talked about how AI is the fourth industrial revolution, after the ones in electrification, mechanization, and computers. He’s also talked about how earlier powers seized the productivity advantage during previous revolutions.
And there’s been a lot of discussion about China already seizing the “commanding heights” of this new industrial revolution and perhaps gaining leverage over the United States and even overtaking it as an economic power.
It’s hard to predict what will happen, so I wanted to see if we could glean insights from previous revolutions.
You argue that the conventional understanding about technological revolutions and a nation’s economic power is wrong. Can you explain?
The traditional view has been to link a nation’s rise in economic power to being pioneers in a few specific leading sectors, such as cotton textiles in the early 19th century and chemicals in the late 19th century.
But I find the rise and fall of great powers are affected by their ability to adapt new general-purpose technologies across many different industrial sectors. General purpose technologies affect growth differentials through a slow process of diffusion across the economy.
That has implications for policy. If a country concentrates on being the first to innovate and to have explosive growth in a new industry, that country might overoptimize for scientific research, advanced R&D, and deepening the expert pool. If instead you focus on spreading general-purpose technologies, then a country might invest more in broadening its engineering skill base, optimizing for widening the pool of general electrical engineers versus producing the top electrical experts. It would also invest more in building strong links between private industry and universities to help rapidly diffuse the knowledge associated with the new technology.
What’s a compelling example of how pioneering a new technology wasn’t as important to economic power as spreading that technology to many fields?
In the late 1700s, Britain developed a major new technology for making wrought iron. It was called the Cort’s-puddling process, and it made iron cheap and widely available. But iron goods didn’t account for a large share of British exports and didn’t provide Britain with a big source of monopoly profits.
The real impact of that technology on economic growth was delayed until about 1815, after it spurred the rise of precision-machining and the broader deployment of iron-based machinery.
In the Second Industrial Revolution, from 1870 to 1914, the United States took advantage of those developments in machine tools to spread the practices of “interchangeable” manufacturing, where products have standardized, interchangeable parts. That gave rise to the “American system” of manufacturing, which in turn drove American productivity growth and allowed the United States to become the world’s preeminent economic power.
What’s important to notice here is that the United States was not the uncontested leader in innovation around machine tools. Britain, France, and Germany were all more sophisticated scientific powers at that time, and they developed very sophisticated machine tools. What propelled the rise of U.S. manufacturing is that it was more successful in adopting machine tools across all these industries.
Why did the U.S. have more success than Britain or Germany in adopting these innovations so broadly?
The U.S. performed better than Britain and Germany in two senses. Britain just did not produce enough mechanical engineers. It was not focused on technical training. Germany actually was producing a lot of mechanical engineers, but they weren’t connected to the industrial economy. They were focused on fundamental theory as opposed to lab applications.
In the U.S., by contrast, you saw a huge flurry of initiatives to increase the talent base in mechanical engineering. First and foremost, you had the Morrill Act, which created land-grant universities in every state. Within a few decades, the number of universities had multiplied from fewer than 20 to more than a hundred, and many of them were setting up technical training programs for engineers. The mission of the Morrill Act was specifically to increase expertise in the mechanical arts.
It wasn’t just the public universities. There was also the Franklin Institute in Philadelphia, a private center for machine tool talent and training mechanical engineers. The center developed public-private partnerships, and you saw a flurry of associations that helped build the linkages between different industries.
In other words, the issue isn’t just how many mechanical engineers you have, but how connected they are to each other and to people in other fields.
It’s also important to standardize best practices in general-purpose technologies via systems that facilitate interoperability between sectors and information flows from one sector to another. Germany’s efforts to standardize such practices in mechanical engineering was much slower than what was happening in the U.S.
What are the implications of all this for the global competition for leadership in AI?
Both China and the U.S. are very optimized toward the traditional leading-sector model — focusing on R&D, attracting high-end talent, claiming innovation leadership, and gaining strong market share in this new strategic industry. The idea is that advances in AI will create explosive growth in the AI industry, which will change the balance of economic power between the U.S. and China. But the history of previous industrial revolutions suggests a fundamental reframing.
First, it suggests there will be a delayed impact from AI on the balance of power. We’re still in the early innings. Major advances in deep learning occurred in the 2010s. If AI is in line with other general-purpose technologies, we might not see the impact on national productivity growth until after four decades. Those who talk about a fast transition in balance of power should take that extended timeline into account.
The history also suggests that the focus shouldn’t be exclusively on being first to introduce new advances in AI. Both sides have an opportunity to invest more in diffusion-centered institutions, such as technology-transfer bodies that transfer innovations from the lab to startups and other companies. This could involve things like “innovation vouchers,” where governments offer subsidies to small- and medium-sized firms to purchase services from universities, research institutes, or companies on the technological frontier. The idea is to stimulate the transfer of leading-edge R&D to a broad pool of adopters. How might advances at the large tech giants, for example, translate to a production line in a fourth-tier Chinese city?
The United States has a lot to do. The U.S. Census Bureau’s most recent Annual Business Survey, in 2018, found that only 3 percent of American firms were adopting AI technologies. Research shows that the gap between when the frontier firm adopts the technology and when it successfully penetrates the entire economy is growing larger and larger.
On the other hand, the U.S. does seem to be doing well by one indicator of having a good skills infrastructure in AI. I look at the number of industry-university collaborations in AI publications. The U.S. has one of the highest percentages of AI publications coming from these joint works. That speaks to the interconnected nature of the infrastructure in this space.
In some cases, the general-purpose technology model may mean not changing things. The United States’ approach to standardization is industry-led, which may be more effective than China’s, which is government-led. When the government attempts to impose standards from the top down early in the process, it’s operating as what Paul David called a “blind giant.” It doesn’t know enough about the technology, so it often locks in dated technologies. China’s top-down approach might stifle the ability of these technologies to spread.
Stanford HAI's mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.