China Power

Is China Waking up to the Dangers of AI?

Recent Features

China Power | Security | East Asia

Is China Waking up to the Dangers of AI?

Beijing has thus far taken a more hands-off approach to AI development than might be expected. But as AI’s power becomes clearer, that could soon change.

Is China Waking up to the Dangers of AI?
Credit: Depositphotos

Early last month, reports emerged that the People’s Liberation Army (PLA) had begun deploying a version of Meta’s Llama model for military purposes. Meta had made the artificial intelligence (AI) model open source, meaning anyone could build a version of their own. Its use for supporting an authoritarian regime – and Meta’s impotent response that such usage violated terms and conditions the U.S. firm could never enforce on the PLA – seemed like yet another parable about how democratic openness could easily turn to authoritarian advantage. 

But by the end of the month, the script had flipped. An independent Chinese lab called DeepSeek announced their R1-Lite model achieved a 52.5 percent success rate on advanced mathematics problems. OpenAI’s o1-preview, the hitherto leader, had scored only 44.6 percent. Moreover, DeepSeek was open sourcing its model. 

Despite China’s closed internet and media environment, this was hardly unusual. Beijing has thus far taken a more hands-off approach to AI development than would be expected given the stakes. But as AI’s power becomes clearer, that could soon change.

The Chinese Communist Party (CCP) has long aimed to balance the potential against the threat posed by the power that comes with cutting edge technology. When Jack Ma’s Ant Group grew too powerful, for instance, regulators swiftly crushed its IPO and forced a restructuring, while Ma in effect went into hiding. 

The move against Ma came amid a vast expansion of regulation in the sector, resulting in the collapse of Chinese startups from over 51,000 in 2018 to a projected 260 in 2024. Venture capital funding has similarly imploded, with yuan-denominated funds raising just $5.38 billion this year compared to a peak of $125 billion in 2017. “The whole industry has just died before our eyes,” one Beijing-based executive told the Financial Times. “The entrepreneurial spirit is dead. It is very sad to see.”

Despite this, China’s tech titans like Tencent and Alibaba – and smaller labs like ZhipuAI and 01 AI – have blazed ahead in AI research. The government has largely played a coordinating and regulatory role, including the organization of an open source repository of models. Making Chinese models available to other Chinese researchers seemed an essential part of catching up, ensuring all Chinese labs benefit from the advancements of other labs.

And China seemed to view itself as far behind, evidenced by the PLA use of an U.S. open source model in Llama over Chinese equivalents. Everyday users agree: on ChatbotArena, where users choose between side-by-side answers from two unnamed models, Chinese-language users preferred answers from Google and OpenAI to those from 01 AI, China’s top-performing firm.

Chinese regulations have therefore been domestically focused. While the Cyberspace Administration of China’s guidelines demand that models “[a]dhere to socialist values” and do not generate content that “divides the nation… or spreads harmful information,” the only reference to other countries is that firms must ensure foreign investors adhere to other laws regulating foreign investment. 

A Chinese model on top could change that equation, especially as it appears that Xi Jinping has growing concerns about AI. The most recent plenum of the CCP listed rogue AI among the most serious threats to the country’s future, a position hard to square with allowing powerful models to fall into the hands of anyone with access to a server, especially adversaries.

On a recent visit to China, Benjamin Todd, a leading AI commentator, noted that many researchers felt Beijing had yet to “wake up” on the importance of advances in AI. Greater investment from the government, he suggested, could close much of the remaining gap with the United States, but no one was sure what would make that happen. China “is very much not acting like the stakes are as high as [many U.S. AI researchers] think the stakes are about to be,” wrote Zvi Mowshowitz, another prominent AI commentator.

The same week as DeepSeek’s R1 release, China announced an export ban on rare earth metals critical to the fabrication of semiconductors, by far its most aggressive move to slow U.S. progress. Six days later, OpenAI released a model that unseated DeepSeek as the world’s top mathematical performer. As advances and dangers accelerate in the field, expect more intervention from Beijing – both to progress the next DeepSeek and to slow the next OpenAI.