Chinese AI Video Startup Implements Censorship on Politically Sensitive Content

Sand AI, a startup based in China, has launched an innovative video-generating AI model called Magi-1 that is capturing attention for its advanced capabilities. This model has received commendations from notable figures, including Kai-Fu Lee, the founding director of Microsoft Research Asia. However, it appears that Sand AI is implementing censorship measures on certain images that may provoke the scrutiny of Chinese regulators, as revealed by testing conducted by TechCrunch.

Introducing Magi-1: A Revolutionary Video AI Model

This week, Sand AI introduced Magi-1, a cutting-edge model that generates videos by predicting sequences of frames in an autoregressive manner. According to the company, Magi-1 is capable of producing high-quality, controllable footage that captures physical dynamics more accurately than other open models available in the market.

• Impressive AI Capabilities: Magi-1 can generate video content that is indistinguishable from professionally produced footage.
• Open Source and Free: Users can access the model without any cost, making it an attractive option for entrepreneurs and developers alike.

Technical Requirements for Magi-1

Despite its impressive features, running Magi-1 is not feasible on standard consumer hardware. The model boasts 24 billion parameters, necessitating the use of four to eight Nvidia H100 GPUs for operation. For many users, including this reporter, Sand AI’s platform remains the only accessible option to experiment with Magi-1.

Censorship Practices of Sand AI

To initiate video generation, the platform requires a prompt image. However, not all images are permitted. TechCrunch discovered that Sand AI blocks the upload of several politically sensitive images, including:

• Images of Xi Jinping
• Tiananmen Square and the Tank Man
• The Taiwanese flag
• Insignias supporting Hong Kong liberation

The filtering mechanism appears to operate at the image level; renaming files does not bypass the restrictions. Users attempting to upload such images encounter error messages from Sand AI’s platform.

Comparative Censorship Practices in Chinese AI Startups

Sand AI is not alone in implementing these censorship measures. Other Chinese startups, such as Hailuo AI and Shanghai-based MiniMax, also restrict politically sensitive images in their video generation tools. However, Sand AI’s censorship seems to be particularly stringent, as Hailuo allows images of Tiananmen Square.

As discussed in a Wired article from January, AI models in China must adhere to strict information controls. A law enacted in 2023 prohibits the generation of content that could “damage the unity of the country and social harmony,” effectively censoring any material that contradicts the government’s established narratives. Consequently, many Chinese startups implement various forms of censorship, including prompt-level filters and fine-tuning.

Interesting Observations on Content Filters

While Chinese AI models often block political content, they appear to have fewer restrictions concerning pornographic material compared to their American counterparts. A 404 Media report highlighted that several video generators from Chinese companies lack essential safeguards against generating nonconsensual nudity, raising questions about content moderation practices.

As the landscape of AI continues to evolve, the implications of censorship in technology remain a critical issue worth exploring.