A startup based in China, Sand AI, has introduced a video-generating AI model with an open license, which has received acclaim from figures such as Kai-Fu Lee, the founding director of Microsoft Research Asia. However, testing by TechCrunch suggests that Sand AI may be censoring images potentially contentious to Chinese regulators in the hosted version of the model.
Earlier in the week, Sand AI unveiled Magi-1, a model designed to produce videos by predicting sequences of frames autoregressively. The company asserts that this model can create high-quality, controllable footage capable of capturing physics more precisely than other open models.
Magi-1, with its size of 24 billion parameters, requires between four and eight Nvidia H100 GPUs, making it impractical for operation on most consumer hardware. Parameters are internal variables that models utilize to make predictions. As a result, many users, including the reporting team from TechCrunch, rely on Sand AI’s platform to experiment with Magi-1.
To initiate video generation, the platform requires a “prompt” image. However, not all prompts are allowed, as TechCrunch has discovered. Sand AI restricts the uploading of images related to Xi Jinping, Tiananmen Square, the “Tank Man,” the Taiwanese flag, and symbols supporting Hong Kong liberation. This filtering seems to occur at the image level since renaming image files does not bypass the restriction.
Sand AI is not the sole Chinese startup that blocks politically sensitive images on its video generation tool. Another platform, Hailuo AI from Shanghai-based MiniMax, also prohibits photos of Xi Jinping. Nonetheless, Sand AI’s filtering process appears to be more stringent, as Hailuo permits images of Tiananmen Square.
As previously highlighted by Wired, models in China must adhere to strict information regulations. A law enacted in 2023 prohibits models from creating content that could “damage the unity of the country and social harmony,” which essentially counters the government’s historical and political narratives. Consequently, Chinese startups frequently censor their models through prompt-level filters or by fine-tuning them.
Despite the tendency of Chinese models to restrict political speech, they often have fewer controls than American models concerning pornographic content. A report by 404 Media noted that several video generators launched by Chinese firms lack basic safeguards against generating nonconsensual nudity.
