HKU Study: ByteDance's Seedance Leads Video Quality but Lags in Safety Filters
A new University of Hong Kong study finds Chinese AI video models surpass US rivals in visual quality, but high-performance models often exhibit weaker safety safeguards.
A recent study led by Zhenhui Jack Jiang, professor of innovation and information management at the University of Hong Kong's business school, evaluated 12 leading AI video-generation models. The results show that ByteDance's Seedance 2.0 Pro took the top spot for visual quality, scoring 3.84 out of 5, surpassing Google DeepMind's Veo 3.1 (3.47) and OpenAI's Sora 2 Pro (3.27). Seedance also maintained its lead when audio capabilities were included.
However, the research revealed a significant gap between technical performance and content safety. When tested with prompts involving extreme violence or unlawful behavior, some high-performing models generated legally prohibited visual content, exposing weaknesses in their safety filtering. Seedance ranked sixth among 11 models assessed for blocking harmful content, scoring 3.12. In contrast, MiniMax's Hailuo 2.3, which ranked near the bottom in visual quality (2.95), emerged as the safest model tested with a score of 3.63, narrowly beating Veo (3.62).
The evaluation used 71 Chinese-language prompts covering text-to-video and image-to-video generation, judged by 20 human evaluators. The study focused on model versions available in April, excluding newer releases. Researchers noted that high technical performance and robust safety guardrails are not mutually exclusive, suggesting companies can strengthen safety standards without sacrificing quality. This finding comes amid escalating competition and increased scrutiny of AI safety globally, following incidents where ByteDance faced criticism over unauthorized reproductions of copyrighted characters and celebrity likenesses.