Chinese AI startup Sand AI, praised for its open-source video generation model Magi-1, appears to be censoring politically sensitive images on its platform. Tests reveal the AI blocks uploads of images featuring Xi Jinping, Tiananmen Square, the Taiwanese flag, and symbols supporting Hong Kong liberation. This filtering occurs at the image level, circumventing simple renaming attempts.
The Magi-1 model, requiring substantial GPU power (four to eight Nvidia H100s) due to its 24 billion parameters, is primarily accessible to many users through Sand AI's platform. This censorship aligns with Chinese regulations forbidding AI models from generating content that challenges the government's narratives. Another Chinese platform, Hailuo AI, also blocks images of Xi Jinping, but Sand AI's filtering seems more stringent.
Such restrictions highlight the challenges faced by AI developers in China navigating stringent information controls. While Sand AI's Magi-1 boasts impressive video generation capabilities, its accessibility is limited by these censorship measures, impacting its potential reach and adoption.