Recently, it has been noted that the video generation artificial intelligence (AI) model developed by Chinese startup Sand AI is actively filtering out politically sensitive images.
TechCrunch revealed that Sand AI’s newly released MAGI-1 model prevents users from uploading images related to contentious topics such as Xi Jinping, Tiananmen Square, the Taiwanese flag, and symbols supporting Hong Kong’s liberation movement. Notably, attempts to circumvent this restriction by altering file names proved ineffective, suggesting that the filtering mechanism operates at the image content level rather than relying on metadata.
This practice of political censorship is not unprecedented among Chinese AI firms. For instance, Shanghai-based MiniMax’s generative media platform, Hailuo AI, similarly blocks uploads of images depicting Xi Jinping. These measures align with legislation enacted by the Chinese government in 2023, which prohibits creating content harmful to national unity or social harmony. In response, AI startups have implemented various strategies, including prompt-level filtering and model fine-tuning, to comply with these regulations.
When Sand AI initially unveiled MAGI-1, the company touted its ability to generate high-quality videos with enhanced physical accuracy. However, the model’s impressive capabilities come at a cost: it requires a staggering 24 billion parameters, rendering it impractical for use on standard consumer hardware. Reports indicate that running Magi-1 requires 4 to 8 Nvidia graphics processing units (GPUs) H100s, highlighting the substantial computational resources necessary for cutting-edge AI video generation.