Export Controls Not Best Strategy to Counter China in AI, Researchers Say

U.S. export controls on artificial intelligence may not be the right strategy to hinder Chinese progress in certain AI subfields, including machine learning, Georgetown University’s Center for Security and Emerging Technology said in a report this week. While the controls may seem “attractive in the abstract,” the report said most decoupling regimes are “imperfect and frequently act as a hindrance, rather than an absolute bar, to a rival’s technological progress.”

The report comes as the Bureau of Industry and Security considers new controls on a range of emerging technologies, including certain slices of AI technologies (see 2201050027). While the report said certain controls are useful, it said an AI decoupling strategy against China won’t be effective in the long term.

The report specifically said an AI decoupling regime would have to target data, algorithms and computational power, which would be a difficult task. “Each of these key inputs feature characteristics that raise doubts about the ability of decoupling regimes to be effective,” CSET said. “Data and algorithms are too geographically distributed to be easily amenable to an export control approach. Computing power, while geographically concentrated, is for a number of reasons only a narrow lever for influencing rival progress in the technology.”

The researchers also pointed to the AI sector’s reliance on international research collaboration, which would be crippled by export restrictions. “Imposing meaningful export controls would be nothing short of a mortal blow to AI research and development,” the report said. University researchers have told BIS that export controls could hurt academic cooperation in a range of fields, including brain computer interface technology (see 2201100010).

While the report also said export controls on computational power in AI -- including various advanced chips -- may “initially appear to be a promising candidate for a decoupling regime,” a “deeper examination suggests that such a compute-focused decoupling regime would be fragile at best for achieving this end.”

The researchers said better access to computational power over China “may ultimately offer only limited advantages in advancing the capabilities of AI systems.” They also said cutting-edge chips are important for machine learning training, but it’s “unclear that limiting rival access to training new models will be the strongest determinant of national competitiveness in this emerging technology.”

More generally, CSET said decoupling is a “narrow tool” to help the U.S. compete in emerging technologies, adding it would be difficult to use the strategy to “permanently bar a rival from catching up in a given technology.” Still, the strategy shouldn't be completely tossed aside, the report said. “To the extent that imposing even a temporary delay in the progress of a rival is valuable, and where the technology in question is amenable to effective decoupling, the tactic offers an attractive option.”

“In the context of AI, imposing even temporary delays on rival progress in the technology may be highly valuable,” the report said. “However, the diffuse nature of its inputs in data and algorithms makes it unlikely that decoupling is an effective means of achieving this end.” CSET suggested other methods to slow China’s progress in AI, including “leveraging” immigration policies to attract more top talent and better shaping technical standards.