Study Finds Chinese AI Video Models Lead in Quality but Trail in Safety
A University of Hong Kong study indicates that Chinese AI video-generation models currently outperform leading U.S. competitors in visual quality. However, the research also highlights that these same Chinese models often lag behind in implementing safety safeguards.
A recent study conducted by researchers at the University of Hong Kong has evaluated 12 prominent artificial intelligence video-generation models to compare their performance and safety standards. The findings suggest a shift in the competitive landscape, with Chinese models demonstrating superior visual output compared to their American counterparts.
According to the study, ByteDance’s Seedance 2.0 Pro achieved the highest ranking for visual quality, earning a score of 3.84 out of 5. This performance surpassed major U.S.-developed models, including Google DeepMind’s Veo 3.1, which received a score of 3.47, and OpenAI’s Sora 2 Pro, which scored 3.27. The research team, led by Professor Zhenhui Jack Jiang, utilized these metrics to assess the current state of generative video technology.
While the study highlights the technical proficiency of Chinese models, it also identifies a significant disparity regarding safety protocols. The researchers noted that several of the top-performing Chinese models fall behind U.S. rivals in the implementation of safeguards designed to prevent the generation of harmful or inappropriate content. This suggests that while the industry in China has made rapid advancements in visual fidelity, there remains a gap in the development of robust ethical and safety frameworks compared to the current standards maintained by leading U.S. firms.
📡 Media Analysis
How each outlet framed the story — angles, word choices, and what they chose to push or ignore.
Focused on the technical superiority of Chinese tech while acknowledging the safety deficit.
"surpassed leading US rivals"
🔍 What Nobody's Reporting
- ·Lack of specific details on the methodology used to measure 'safety safeguards'.
- ·Absence of comment or reaction from the U.S.-based companies mentioned in the study.
📰 Sources
0 A-rated source(s) among 1 total. Lowest trust: SCMP (B)