Military metaphors are a symptom of intellectual laziness. “AI arms race” is an example. AI is not an arm, nor does the possession of advanced AI guarantee strategic advantage. The image of an arms race is stale and obscures the complexity of AI’s effects and the processes that create it. A new vocabulary would better guide policy and help Washington and other capitals better understand the processes, effects, and challenges of technology, disruption and innovation.
Declarations that there is an “arms race” are usually tools of political rhetoric, used to press for more spending, to criticize incumbents, or to appeal for measures to reduce perceived risks. Using the arms race metaphor for AI raises several questions. If this is a race, how do we measure who is ahead? What are the right metrics to gauge national AI success? How does better AI translate into tangible national power? Precedents from the last century are unpromising as a source of answers to these questions.
The idea of an “arms race” became prominent in the late 19th century, when Britain and its European competitors began an intense competition for naval supremacy. Each competitor built powerful ships; the British to maintain their naval advantage and the others to match or surpass Britain. “Arm Race” reappeared in the 1930s as German rearmament triggered unease and last-minute arms buildups in the US and UK. In the 1950s and 1960s, the arms race was driven by the acquisition of nuclear weapons and their delivery systems (and accompanied by periodic claims that the US was falling behind the Soviets).
The term became pejorative for some, part of a critique of the “military industrial complex,” which implied mindless competition, a waste of money, and growing existential risk for the entire human race. Today’s discussion of risks from AI and autonomous weapons often parallels nuclear-era fears shown by the “Doomsday Clock,” which since 1947 has unhelpfully and inaccurately predicted that the destruction of the human race is imminent.
Previous arms races had straightforward quantitative metrics, since merely counting the number of battleships, missiles, or bombs would indicate who was in the lead – although quantitative measures beg the question of whether more weapons can be used effectively. Metrics for any new tech race are not as simple. There is clearly competition, as China believes it must surpass the US to restore its former glories and the US now recognizes that it needs to respond to this. Both countries’ governments have plans (long-standing in China, erratic in the US) to strengthen their technological capabilities. Neither side is willing to compromise or enter into security negotiations. In the US, the AI race can be a symptom of occasional panic over China’s efforts to overtake and replace it as the leading world power, although this panic usually confuses China’s technological rise with the end of the unipolar moment. But tech competition between the US and China is not an arms race since it entails far more than building arms.
Arms races are an effort to preserve stability by ensuring balance in military capabilities among opponents. Past arms races involved building new weapons that were believed to confer military superiority. One side begins to build, and the others respond with matching efforts. But AI is not a weapon; it is a software tool and not easily made lethal. Creating AI tools is largely a business competition, not two states stockpiling weapons. AI competition involves dozens of private actors guided by market forces. Innovation and production are driven by profit motives. The leading companies involved are not building weapons and some, at least in the US, even assert altruistic motives. The most reliable metrics of success are company revenue and market share, which reflect both performance and the ability to innovate. The same metrics – national income and market share – apply to nations in determining their technological strengths.
Gauging success in an AI “race” is complicated. There are so many statistics and so little agreement on which are meaningful for strategic advantage that an artful selection can present whatever story one wants. Possible metrics could include deployed military systems using some form of AI, R&D spending, manufacturing capabilities, or business adoption. A review of this data produces mixed results but point to a US lead. The US private sector has twice as many AI researchers, invests almost three times as much as China, and dominates the production of key technologies like semiconductors. Some studies suggest that China may lead in AI startup investment and hopes to close any “gap” by 2030.
The nation that “dominates” AI will not be “the ruler of mankind.” It is more likely that the nation that is best at creating and accommodating technological change will do better economically than others. Both China and the US have strengths and weaknesses in this. Since AI requires hardware (like GPUs), math, coding skills, and access to data, there is no “chokepoint” where one side can deny AI to the other. A competitor can compensate for lower-performance GPUs by aggregating them or by using software workarounds.
To gain advantage, policies that accelerate research and ease connection to the global innovation system would best serve the United States. Attempting to win the “race” by tripping opponents can actually undercut American leadership by harming the ability to invest and innovate. Changing course in Washington may require a new generation of policymakers who better understand technology and are comfortable with it. This change has already happened in business, where most large companies embrace technology to cut costs and increase revenue, with many exploring how to use AI to “disrupt” their own internal processes to become more competitive.
The concept of an AI arms race is appealing as it simplifies complex issues and provides a handy rhetorical device, but it is unhelpful both as a metric and as a guide for policy. Adoption of AI promises economic benefits, but not necessarily military strength or global influence. Discussions on the effect of AI often start with the technology and leap to the conclusion of increased power without considering the many intermediate steps and difficulties in getting from software to outcomes. Discussions of an AI arms race are often an imperfect retread of nuclear debates. Abandoning the use of terms borrowed from 20th century defense and strategy would lead to better policy. We did not talk about an “internet arms race” 25 years ago even though that technology led to immense change. We should not talk about an AI arms race now.
James Lewis is a Distinguished Fellow at CEPA’s Tech Policy program.
