A story in Hong Kong’s South Morning China Post, citing an obscure research paper linked to the Chinese military, wrongly suggested that Chinese AI giant, Baidu, is working with the Chinese military, setting off a rout in the company’s stock.
In fact, the researchers who wrote the paper only used Baidu’s large language model, Ernie Bot, in the same way that any public user would. The newspaper issued a correction, but much of the damage was already done: Baidu lost an estimated $4.68 billion in market capitalization on the Hong Kong Stock Exchange on Monday, presumably over fears that Baidu could face US sanctions for working with the military, as has Chinese telecommunications giant Huawei.
“Baidu has not engaged in any business collaboration or provided any tailored service to authors of the academic paper or any institutions with which they are affiliated,” Baidu said in a statement. The company’s primary listing is on the NASDAQ stock exchange.
Nonetheless, the incident highlights the difficulty that tech companies face in keeping a distance from their national militaries while developing and deploying dual-use technologies such as generative AI.
Google faced a backlash several years ago when it participated in Project Maven, a program to develop computer vision for military drones, eventually ending its involvement under public pressure – though it has quietly continued to cooperate with the military and intelligence establishment.
Most recently OpenAI has quietly amended its usage policies, deleting “military and warfare” from a list of prohibited uses of its technology.
U.S. companies have the option to withhold proprietary technology from the government. But in China, it isn’t as clear, putting Chinese companies – particularly those listed on U.S. stock exchanges – in a perilous position.
Tech companies hold the intellectual property and expertise to develop and deploy these powerful systems that are inherently dual use and of tremendous interest to their national militaries.
U.S. laws governing dual-use technology, which has both civilian and military applications, has long focused on limiting proliferation rather than U.S. government or military access to proprietary technology. However, most recently, the Biden Administration’s October 2023 Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence mandated that companies provide the federal government, on an ongoing basis, with “information, reports, or records regarding any ongoing or planned activities related to training, developing, or producing dual-use foundation models.”
For years, the US was seen at a disadvantage to China because US corporations cannot be compelled to cooperate with the military. However, China’s ‘military-civilian fusion’ policy (MCF), while undeveloped in law, leaves Chinese companies vulnerable to US sanctions because they have difficulty convincing the US that they can refuse to cooperate with the Chinese military.
But the Center for a New American Security, a Washington, D.C.-based bipartisan think tank that conducts in-depth research and analysis on national security and defense policies, found in 2021 that the Chinese policy has been often misunderstood and oversimplified by US commentators. “There is no explicit legal obligation for Chinese companies to participate in MCF,” the report stated.
Baidu, whose primary business is its search engine, is not comparable to Huawei. Nonetheless, this is the context of the South China Morning Post article and its subsequent impact on the Chinese tech company. The newspaper report, published on January 12, alleged that Chinese military researchers connected their “military AI” to commercial large language models including Baidu’s Ernie Bot – China’s leading challenger to OpenAI’s ChatGPT.
The newspaper later corrected the report, saying that its original claim that a Chinese military lab “had forged a physical link between its Al system and Baidu’s Ernie” was wrong and that “there is no dedicated line between the two.”
But the newspaper report had deeper problems, according to Pekinology, an ostensibly independent newsletter that follows US-China relations, edited by a former correspondent for Xinhua, the Chinese state news agency, who is now a research fellow at the Center for China and Globalization, a leading non-government think tank in China.
Pekinology systematically debunks the newspaper report, explaining that it misrepresents the content and implications of the Chinese paper. For example, while the newspaper suggests that the military-linked research lab had built a “military AI” that used Ernie Bot for battlefield analysis, it’s clear from the lab’s paper that it was only proposing such an AI system and that no such system exists.
Moreover, the researchers’ use of Ernie Bot consisted of prompting the public model and recording the model’s responses – something that anyone in China can do.
