Huawei builds AI model that’s “nearly 100%” effective at censoring sensitive content

Huawei has co-developed a safety-focused version of DeepSeek’s AI model that it claims is “nearly 100% successful” at preventing discussion of politically sensitive topics. The collaboration with Zhejiang University demonstrates how Chinese companies are adapting open-source AI models to comply with domestic regulations requiring AI systems to reflect “socialist values” and avoid sensitive political discussions.

What you should know: Huawei used 1,000 of its own Ascend AI chips to train the modified model, called DeepSeek-R1-Safe, which was built from DeepSeek’s open-source R1 model.
• The model achieved “nearly 100% successful” defense against “common harmful issues … including toxic and harmful speech, politically sensitive content, and incitement to illegal activities,” according to Huawei.
• However, that success rate dropped to 40% when behaviors were disguised through scenario-based challenges, role-playing scenarios, and encrypted coding.
• DeepSeek’s founder Liang Wenfeng is an alumnus of Zhejiang University, though DeepSeek and Liang had no apparent involvement in this safety-focused project.

The big picture: Chinese regulators require domestic AI models to reflect China’s “socialist values” before public release, forcing companies to build censorship capabilities into their AI systems.
• Chinese AI chatbots like Baidu’s Ernie Bot already refuse to answer questions about Chinese domestic politics or topics considered sensitive by the ruling Communist Party.
• DeepSeek’s original models shocked Silicon Valley and triggered a selloff of Western AI stocks in January due to their advanced capabilities, and they have since been embraced and modified across Chinese industry.

Key performance metrics: Huawei’s modified model maintained strong performance while adding safety features.
• The comprehensive security defense capability reached 83%, outperforming competing models like Alibaba’s Qwen-235B and the original DeepSeek-R1-671B by 8% to 15% under identical testing conditions.
• DeepSeek-R1-Safe exhibited less than 1% performance degradation compared to the original DeepSeek-R1 model.

Why this matters: The development highlights how Chinese companies are adapting cutting-edge AI technology to meet strict domestic content controls while maintaining competitive performance levels.
• It demonstrates the technical feasibility of building effective censorship into AI models without significantly compromising their capabilities.
• The project represents part of a broader trend of Chinese firms modifying and deploying DeepSeek’s technology across various applications while ensuring regulatory compliance.