What is and isn’t concerning about China’s AI surprise

Reuters recently reported that researchers affiliated with China’s People’s Liberation Army released studies demonstrating that they had applied versions of Meta’s Llama AI system to experimental military applications.

Within hours, there were renewed congressional calls to slam the gate on unrestricted commercial exports of AI products. Although concern about China’s global AI ambition is warranted, officials are misinterpreting this event and drawing the wrong lessons.

In one case, Llama’s 13b version from 2023 was found to be useful for intelligence analysis and military-relevant information queries. In another, Llama 2 was applied to “support electronic warfare and self-defense jamming strategies.” This raised immediate concern that U.S. companies were brazenly sharing our most advanced technology without restrictions. To address that, we need to ground the conversation and determine whether this incident truly represents a threat.

The models used by the PLA can hardly be considered cutting-edge. According to Lymsys’ Chatbot Arena, an AI ranking service, Llama 2 is considered the 98th-most capable model and Llama 13b the 155th. In other words, this is outmoded tech.

In more familiar terms, worrying about the misuse of these models is akin to worrying that the Chinese military might find value in Windows XP. Dinosaur technology might provide value to an outmatched military, but not a strategic edge against the U.S. and its ability to counter.

Also notable is the suspect information source: People’s Liberation Army research, which could well be either fabricated or released with a political agenda.

Chinese researchers have a recent history of ruffling policy feathers by trumping up tech studies. In 2022, some made a similar splash by claiming they discovered a quantum computer-powered method of cracking the encryption that secures most systems. Years later, there is little validating evidence of a real threat.

These new AI studies must likewise be taken with a massive grain of salt. By brandishing such provocative studies, China’s intent could be to prompt a U.S. policy fumble. Basing any legislation on the PLA’s message demands unique caution.

More specifically, hasty new AI export controls would risk impeding America’s ability to compete globally. We cannot ensure that AI is used safely and for shared prosperity if American firms in Meta’s class cannot export AI.

That speaks to a real takeaway, which is not advanced risk, but something more subtle. China’s use of Meta's product signals that the U.S. presumption of openness is helping to produce technology that is the global go-to. This is a supreme competitive advantage and ultimately loads the economic dice in our favor.

In the past, openness has enabled American technology to become the global default. The result was to draw other nations to invest in, build atop of and secure our technology. It’s precisely because the world has defaulted to openly accessible systems like Windows, Android and iOS that Silicon Valley has the revenue and data needed to power current AI innovation. America must maintain this position to drive security-essential investments.

What’s genuinely concerning is China’s clear commitment to applying AI deeply. Complex tasks such as electronic warfare mean working under constraints and in truly innovative ways.

Strategically, this is key, because the country that wins a potential conflict won’t necessarily be the one with the best tech, but the one that that best uses the tech it has. Making this principle more important is that China will find a way to develop or acquire functional AI tech regardless of U.S. policies.

On the usage front, we may be falling behind. According to a July study of industry leaders, Chinese businesses hold an 18 percent lead over the U.S. in applying generative AI. If we continue to lag in application, the superiority of our AI models won’t be enough.

This is the lesson that should command policymakers’ attention. Instead of slamming China with export restrictions, national security demands that we meaningfully, rapidly and deeply apply this technology ourselves. The application of AI is the real AI race.

It will take work. The most important near-term step is avoiding reactions that might restrain both U.S. industry and government. Instead, Congress and the next Trump administration should shift towards enabling better, safer, more transparent AI. A starting point is to commission AI regulatory mapping exercises to better understand and scale back the web of rules that bind the tech’s responsible use.

Naturally, there is no silver bullet. While the latest reports don’t represent a unique threat, China’s continued innovation and deep application might. If we want to succeed, it’s imperative the United States sustain its commitment to openness, complemented by an emphasis on AI application.

If we can avoid policy missteps at this critical moment, the AI race is ours to win.

Matthew Mittelsteadt is a technologist and research fellow with the Mercatus Center at George Mason University.