OpenAI and Google have admitted to selling their advanced artificial intelligence (AI) models to Chinese companies on the U.S. Department of Defense blacklist, the Financial Times (FT) reported on the 10th.
The FT noted that this exposes a loophole in the U.S. government's efforts to check China's AI development. According to the report, OpenAI and Google confirmed in interviews with the FT that they have provided AI services to the Singapore subsidiaries of Alibaba, Baidu, and Tencent. The three companies were placed on the U.S. Department of Defense blacklist on grounds of links to the Chinese military.
OpenAI confirmed that while access to its AI models is not permitted in China, it has allowed some companies, even those Chinese-owned or headquartered in China, to use its models "in countries where safety measures can be implemented and 'distillation' can be monitored."
Distillation is a technique that trains a new model using the responses of a superior AI model as data. Originally used to create lightweight models comparable to large ones, U.S. AI companies claim that Chinese AI firms are effectively stealing a significant portion of AI model functions by distilling without permission. OpenAI added, "We want AI based on democratic values to be more widely used than AI controlled by authoritarian governments," and "access rights should not be determined by a company's nationality alone."
Google also said its AI services are available in Hong Kong and Singapore, but that users must instead comply with usage rules including a ban on distillation. Google acknowledged that geographic service restrictions alone cannot fully address the distillation risk, and that experts in this field can easily circumvent such restrictions. The FT reported that while providing such circumvention services does not violate current law, calls are again growing for stronger regulation of AI model use, similar to export controls on chips used to train high-performance AI models.
Chris McGuire, a technology and security expert at the Council on Foreign Relations, said, "The Trump administration says it must always outpace China in AI, but the problem is that in reality it is doing little on export controls, which are the practical means to slow China's growth."
He added, "Advanced AI models should not be provided to companies headquartered in China anywhere in the world." Joe Kambam, an expert on AI policy and national security law at the U.S. nonprofit Law Reform Institute, warned earlier this year that "the U.S. government must use export controls and related measures to block Chinese labs that systematically siphon off the advanced capabilities of American AI without bearing any of the computing, engineering, and safety-related costs that U.S. AI companies shoulder," adding that "this situation weakens the economic foundation of American advanced AI leadership."
