When Ben Luria co-founded his startup, the idea of machine unlearning was mostly theoretical. That changed when China’s DeepSeek launched its first open-weight large language model last year. Luria saw an opportunity to adapt it for Western markets.
Chinese models from DeepSeek, Moonshot, Z.AI, Alibaba, and Tencent have gained traction globally, including among American businesses. Their appeal lies in lower costs, strong performance, and customization. However, concerns persist about their alignment with the Chinese Communist Party and potential security risks.
The Rise of Chinese AI Models
Chinese open-weight models have quickly gained popularity, with companies like Airbnb, Uber, and legal AI firm Harvey building proprietary systems on top of them. Despite this adoption, some remain wary. Last October, the Trump administration banned civilian agencies from using DeepSeek products, citing national security concerns.
For Graham Webster of Stanford’s DigiChina Project, the risks depend on the context. “If I were running a product in Taiwan, I would be more concerned,” he said. “But if I’m making a product that helps people make a website for their coffee shop, it doesn’t matter.”
Realigning Models for Western Standards
Luria’s startup, Hirundo, focuses on realigning Chinese models to meet Western expectations. “We’re not making the model anti-Chinese,” he said. “We are bringing it up to speed on safety and to the standard of mainstream Western AI.”
Hirundo edits model weights to enhance security, fairness, and reduce harmful knowledge. For example, their analysis of Alibaba’s Qwen models revealed omissions on sensitive topics like the 1989 Tiananmen Square protests and forced labor camps in Xinjiang.
Cybersecurity firm CrowdStrike’s analysis of DeepSeek’s R1 model last year showed that prompts containing topics the Chinese government considers politically sensitive were up to 50% more likely to receive code with “severe security vulnerabilities.” Another study found that Chinese models produced less secure code when the user identified as a member of the U.S. government.
Read Also: Vivo X500 Ultra camera details leaked
Boston-based Lazarus AI removed mandatory censorship from Alibaba’s Qwen models. The company’s approach involved generating prompts related to banned topics and using answers produced by Google’s Gemini Flash to realign the models. The company’s ReAligned models have been deployed at about a dozen U.S. companies.
However, Oriane Peter, an AI researcher at King’s College London, argues that aligning models to so-called Western values is an “impossible standard.” She emphasizes that alignment must reflect the specific social context where the model operates.
Addressing Security and Bias in Chinese AI Models
Cybersecurity concerns surrounding Chinese AI models have been highlighted by CrowdStrike’s analysis of DeepSeek’s R1 model. The study found that prompts on politically sensitive topics were more likely to generate code with severe security vulnerabilities. Additionally, Chinese models were found to produce less secure code when users identified as members of the U.S. government.
Despite these efforts, concerns persist.
Global Pushback and User Perspectives
Worldwide, there is growing resistance to Western bias in AI systems. Researchers note that even some Chinese and Russian models default to American geopolitical narratives on historical events. In response, countries in Asia, Latin America, Africa, and the Middle East are developing sovereign AI models to better reflect their languages and cultures.
As Chinese open-weight models continue to gain traction, the debate over alignment, safety, and cultural adaptation remains a critical issue for global AI development. CrowdStrike’s findings and the study on code security show the technical challenges, while the push for sovereign models highlights the broader cultural and geopolitical implications.