ByteDance's 1733 Commercial Space in Zhongkun Plaza.
ByteDance’s New Department: China’s Shifting Perspective on AI Safety
As the U.S. and China race for AI dominance, Chinese tech giant ByteDance has taken an unusual step: establishing a dedicated AI safety department. On October 14, the company published a recruitment post for a new “Model-Alignment & Preparedness” team, marking a rare public statement on AI safety risks by a major Chinese developer. With Chinese companies’ close ties to Beijing and recent reports of rogue behavior by Chinese AI agents, ByteDance’s new department sparks a change towards China’s growing willingness to engage seriously with AI safety concerns in an unprecedented way.
ByteDance, best known for creating TikTok, is one of China’s top AI powerhouses. Chinese AI developers have typically prioritized speed over safety, while U.S. firms like OpenAI have discussed the potential for AI safety risks since 2015. ByteDance CEO Liang Rubo has only ever broadly attested to AI safety, making no official statements on the topic. Similarly, Beijing officials have expressed interest in AI governance, but their messaging has not always been consistent. Last month, China’s minister for state security outlined AI safety as “the bottom line requirement” beneath a primary focus on innovation.
ByteDance’s move may reflect a broader shift in the government’s approach to AI safety. Beijing is much closer to its leading technology companies than Washington is with its private sector tech giants; under the 2023 generative AI measures, Chinese AI companies must align their content with CCP values and go through security assessments and algorithm filing before launching publicly. If the government had a staunch position against AI safety, ByteDance would be unlikely to have formed this department. ByteDance’s new safety department could also reflect Chinese AI developers’ growing concerns about the risks advanced AI systems pose, especially given recent issues with Chinese AI agents exhibiting rogue behaviors like those of U.S. frontier models.
If China’s industry and government align on the need for AI safety, it could open the door to increased cooperation between Beijing and Washington on guardrails. During last month’s U.S.-China summit in Washington DC, Xi Jinping even proposed moving towards further cooperation on AI. The first barrier to meaningful cooperation, however, is the Trump administration’s reluctance to impose restrictions on AI development. President Donald Trump explicitly opposes AI safety collaboration with China, since the U.S. already has a significant lead in the AI race.
The second barrier is U.S. antitrust law, which restricts private companies from entering into mutual agreements that could be seen as market-slowing behavior. While Anthropic CEO Dario Amodei has called for slowing down to address safety risks, leading U.S. AI developers could face significant antitrust barriers if they formally agree to suspend or slow down frontier AI development collectively.
The third barrier is that Washington and Beijing often interpret the same events through vastly different perspectives, leading to continuous tensions and distrust. This distrust means neither can be sure the other would honor an agreement, so each assumes the other will keep building to surpass it. With AI now being a fundamental part of the race for global hegemony, Washington and Beijing will both be reluctant to enter any agreements that risk slowing them down.
Even setting those barriers aside, a new safety department does not guarantee safer AI development. It’s unclear how much influence ByteDance’s safety unit will have over AI model development, testing, or deployment, and even on the U.S. side, safety efforts have also not yet proven effective either. For instance, the Hugging Face breach and U.S. government access incident both involved OpenAI agents, even though OpenAI is one of the labs more publicly focused on safety. Still, the risks of misaligned models acting in unintended ways are worth evaluating. Understanding these rogue behaviors is the first step towards preventing more incidents from continuing, which is why collaboration and data sharing between Washington and Beijing could be helpful in making progress on AI safety for both sides.
If more Chinese AI companies follow ByteDance’s lead, this could signal a shift in Beijing’s willingness to engage constructively with Washington on AI safety. But barriers like the Trump administration’s lack of interest in AI safety, U.S. antitrust law, and mutual distrust make actual progress on pacing the frontier highly unlikely between the two countries. ByteDance’s decision offers a potential opening for cooperation; whether Washington and Beijing seize it will help determine if AI safety can advance alongside competition rather than be sacrificed to it.
Image credit: N509FZ, via Wikimedia Commons, licensed under CC BY-SA 4.0.
