A Chinese AI startup called Sand AI recently launched an advanced, openly licensed video-generating AI model named Magi-1, gathering significant attention and praise within the AI community. Tech leaders, including Microsoft Research Asia’s founding director Kai-Fu Lee, have applauded the ingenuity of the model, which the company boasts can create high-quality, realistic videos with improved accuracy in physical simulations.
However, recent testing reveals that Sand AI appears to be implementing strict censorship measures on its platform, likely in compliance with Chinese regulatory scrutiny. Specifically, the hosted version of the Magi-1 model follows aggressive filters that prohibit users from uploading politically sensitive images. Attempts to upload images depicting Chinese President Xi Jinping, the iconic scene of Tiananmen Square and Tank Man, the Taiwanese flag, and symbols supporting Hong Kong independence reliably trigger error messages, effectively blocking their use.
TechCrunch’s experiments suggest that these restrictions target the image content directly, as simple methods such as renaming the files could not circumvent the censorship. Although similar practices have been observed among other Chinese generative media tools—including MiniMax’s Shanghai-based Hailuo AI, which specifically blocks images of Xi Jinping—Sand AI’s filtering appears notably more rigorous. For example, Hailuo AI still allows images of Tiananmen Square to be uploaded, highlighting Sand AI’s heightened degree of internal image scrutiny.
Chinese artificial intelligence companies must navigate a stringent regulatory environment. A 2023 Chinese law explicitly forbids AI-generated content that “damages the unity of the country or social harmony,” effectively mandating adherence to government historical and political narratives. To comply, local companies frequently employ proactive censorship mechanisms, often through the fine-tuning of their models or the implementation of content-level controls.
Interestingly, while tightly regulated, politically sensitive topics are considered off-limits, Chinese generative models sometimes exhibit fewer safeguards around explicit and adult content than their Western counterparts. Recent reports have indicated the emergence of AI-generated nonconsensual nudity from video-creation platforms operated by Chinese startups, suggesting uneven enforcement of content moderation standards across different categories of potentially problematic media.
