Resetting America's AI strategy
DeepSeek's success should prompt serious reevaluation of underlying assumptions and approach
The rise of DeepSeek, the Chinese hedge fund turned leading AI lab, has raised profound concerns about the United States’ strategy to win technology competition with China. Despite a fraction of the resources and increasing US export and investment controls, DeepSeek has delivered models that perform favorably with those of the leading American firms. DeepSeek’s success demonstrates the need for a US strategy that balances frontier innovation with practical applications and greater openness to ensure AI leadership and its benefits are widely shared.
At the heart of this challenge are several debates shaping US AI policy: should the focus be on achieving artificial general intelligence or ‘good enough’ applications in areas like industrial automation or life sciences? Should innovation be driven by state-led investment and governance or by market forces? Should systems be closed and controlled or open and widely diffused?
The current US strategy rests heavily on the assumption that reaching artificial general intelligence first will confer decisive and enduring security and economic advantages. This in turn demands significant investments to advance the technological frontier and meet the technology’s energy demands; the privileging of closed systems that give greater assurances of control; strict limits on China’s access to fundamental enablers, such as advanced semiconductors, while reducing dependencies on Taiwan; and, for now, prioritizing innovation while neither constraining nor encouraging adoption of commercial and consumer use cases.
The US emphasis on prioritizing artificial general intelligence over more targeted applications is consistent with a broader obsession with building the best instead of the good enough. This has become most stark in the military, which now recognizes it must rapidly pivot from a decades-long emphasis on building expensive, exquisite systems to producing large numbers of attritable ones. China, for its part, has consistently proven in sector after sector that being “good enough” is all that is necessary to triumph.
But America’s emphasis on artificial general intelligence is also consistent with a persistent and flawed pathology that has accepted deindustrialization as permissible in favor of higher value added activities, such as software and services. An even more disconcerting interpretation is possible too, one that implicitly concedes that the US has long passed the point of industrial no return and that AI is the country’s only hope of a fair fight, or the prospect of deterring one, against China.
Avoiding false choices
A more balanced path forward would acknowledge that the possibility, promise, and risk of artificial general intelligence is sufficient to merit the significant investment in its pursuit. But the US can also accelerate investments in practical applications. And it should do so without losing sight of the extent to which the Chinese political system’s obsession with control and inability to attract global talent creates its own strategic disadvantages.
Despite some calls for a new Manhattan Project or outright nationalization of the AI industry, the US has been well-served by its tech ecosystem. Market-led innovation creates more responsive demand signals that reward the most productive uses of technology and minimize the risk of stumbling into mandated technological dead-ends. At the same time, there is a near term risk that, if concerns of a bubble are proven true, it could lead to a hangover that stifles continued momentum. This is where the government can make a difference by being prepared to provide various forms of through-cycle support.
Embracing openness
America’s AI rethink should extend to a bias against openness that extends from the systems themselves to China’s access to technology and broader diplomacy. The government has scrutinized whether open systems introduce risks that closed ones do not. This is based, in part, on the belief that AI’s tendency to exhibit unexpected emergent capabilities is best managed within a controlled environment. But this ignores that closed systems remain vulnerable to misappropriation and misuse and forgo the security and economic benefits of decentralizing technology to reduce single points of failure.
The Biden administration’s “small yard, high fence” approach was an attempt to limit China’s access to critical technologies. But, in practice, technological evolution, China’s ingenuity and resources, the need for coordination with allies, and corporate lobbying resulted in a perpetual state of too little, too late. The Trump administration’s day one executive order on trade appears to signal a further intensification and broadening of technology controls. An alternative approach would be to blunt the indigenization of China’s semiconductor industry at the high-end by offering controlled, cloud-based access to Chinese end users. A combination of approaches will be needed to address the threat of overcapacity in legacy chips, including the possibility of novel component tariffs targeting semiconductors embedded in final exports.
America’s everpresent tension between confrontation and unilateralism and diplomacy extends to technology competition. Diplomacy is not discretionary, but essential to mitigating the potential harms of artificial intelligence and fully maximizing its benefits. With allies, the Biden administration focused on an aligned approach to export controls and too little on aligning its industrial policy. Apart from a late-term summit with China, more substantive engagement has been slowed by bureaucratic uncertainty on China’s part about which part of its government will lead on AI policy. President Trump, known for his interest in the existential risk of nuclear weapons, should not neglect the opportunity to engage Beijing on the no less profound risks AI could present. An affirmative diplomatic vision for third countries is needed too; otherwise China will further entrench its position in those countries’ tech stacks, offering the possibility of AI enabled governance that advances Beijing’s norms and interests globally.
Supporting innovation and adoption
As scholars such as Jeffrey Ding have pointed out, innovation only counts for so much if it does not radiate into the real economy. Here, the American public is sending troubling signals about the adoption of AI. In a recent Google survey, Americans report far lower levels of use of the technology in the past year (29%) than counterparts in Europe (42%) and Asia (52%). Americans are also notably least optimistic about the technology’s possibilities and the probability of positive impact on the economy. If this pessimism persists, it will deny the US not only the upside of AI, but the basis for continued investment in research to drive the frontier forward.
To allay some of their concerns, American workers deserve modern livelihood insurance and workforce development offerings that would enable them to embrace the future with confidence. Instead, policymakers are betting implicitly that America will outperform China because its democracy can better absorb the frustrations of displaced workers than China’s authoritarian system. The latter economy’s high degree of concentration in state-owned enterprises will serve as a further constraint: squeezing out the potential for private sector gains and unable to capture much of automation’s potential given their often far lower levels of digitization.
Continued US leadership in artificial intelligence is more likely to result in a safer and more prosperous world than the alternative. But neither continued leadership nor positive outcomes are assured. It will require foresight, agility, and commitment of resources with little precedent. It is exactly the kind of challenge that an open society is best positioned to solve. The question is whether the US will remain enough of one to succeed.