Chinese AI Developers Emphasize Cultural Nuances in Safety Amidst Global Scrutiny
Chinese artificial intelligence developers are increasingly highlighting the importance of cultural context in their approach to AI safety, asserting that their domestic models are rapidly closing the performance gap with their US counterparts. This perspective challenges a Western-centric view of AI risk mitigation, with industry insiders suggesting that Chinese companies should not be judged by the same standards as their global rivals.
The upcoming period is anticipated to be a significant one for Chinese AI firms, with the release of several major new models expected around the Lunar New Year. This follows a year where concerns about the potential risks associated with Chinese AI models led to adoption barriers for some international users. Notably, the prominent AI company DeepSeek faced bans or restrictions in over ten countries, including the United States, Italy, and India.
Recent third-party evaluations indicate that by the start of 2026, Chinese AI models have achieved a performance level closer to US competitors than ever before. This progress has, in turn, spurred calls for Chinese companies to dedicate greater attention to critical AI risks such as misuse and misalignment.
A former researcher at DeepSeek, Tu Jinhao, suggested in a recent podcast that the intense focus on matching US technological advancements may have inadvertently diverted resources and attention away from domestic AI safety initiatives. “All the computational resources are being spent training AI models, with little left to spend on safety work,” Tu commented. DeepSeek has not yet responded to requests for comment on this matter.
In December, a report published by the US non-profit organization Future of Life Institute (FLI) expressed concerns regarding the transparency of safety practices among Chinese AI companies, including DeepSeek, especially in light of their rapid model improvements. Max Tegmark, president of FLI and a physics professor at the Massachusetts Institute of Technology, stated, “As firms across [US and China] converge at similar capability levels, they should be held to similarly high safety expectations.”
However, proponents of the Chinese AI industry argue that a perceived lack of extensive public communication does not equate to a disregard for AI risks. Li Zixuan, head of global operations at Zhipu AI, a developer of the GLM model, stated that the company employs a “systematic evaluation suite” for assessing model safety and incorporates “refusal mechanisms” for sensitive content, though specific details were not provided.
Zhipu AI, known internationally as Z.ai, received a “D” grade from the FLI’s expert panel, which included AI professor Zeng Yi from the Chinese Academy of Sciences, for its AI safety practices. Despite this, the Beijing-based firm was the first Chinese company to commit to voluntary global AI safety pledges following the Seoul AI Summit in May 2024, promising to release a comprehensive AI safety framework by February 2025. This framework has yet to be published.
Li explained that Zhipu AI’s internal safety efforts are “based on social norms and cultural context of the society we operate in, which may not resonate with all global audiences.” He elaborated, “Values are not universal, they carry strong cultural and regional nuances. That said, we acknowledge that more than half of core values are shared across different regions and cultures, and we are committed to upholding these universal principles in our AI development.”
Navigating International Collaboration and Regulatory Frameworks
Shanghai-based MiniMax, another significant Chinese foundational model developer, also signed the voluntary commitments in February 2025 but has not yet released its safety framework. The company has, however, actively participated in international AI governance discussions, including a private Track 2 dialogue held in November 2024 in Tokyo, which brought together US and Chinese AI companies.
Track 2 dialogues are informal discussions among non-governmental actors designed to foster collaboration between nations on critical issues, such as global AI governance. Sources familiar with the Tokyo meeting indicated that participants also included representatives from OpenAI and ByteDance. Neither OpenAI, ByteDance, nor MiniMax have commented on their involvement.
Fang Liang, head of AI safety in China at the Beijing-based consultancy Concordia AI, suggests that the emphasis on voluntary disclosures can reflect a “Western-centric perspective” that disproportionately benefits US companies. He noted that while US firms often disclose their safety practices in line with voluntary commitments, Chinese companies are frequently subject to more stringent mandatory regulations and industry standards.
Chinese authorities have implemented a series of regulations targeting AI services over the past year. These include requirements for labeling AI-generated content and draft rules concerning AI companions.
Recently, China’s internet regulator announced that major AI service providers, such as ByteDance and DeepSeek, had applied AI labels to over 150 billion pieces of AI-generated content in the four months since the new rules became effective. This proactive approach to regulation underscores a different, and perhaps more prescriptive, path to AI safety being forged within China, distinct from the voluntary commitments often favored in Western contexts. The ongoing development and deployment of advanced AI models by Chinese companies, coupled with their unique approach to safety and regulation, will undoubtedly continue to shape the global AI landscape.
