The Pulse
HKU Finds AI Video Quality Outpaces Safety Controls
A University of Hong Kong study finds that leading AI video generators perform unevenly on safety tests, even as one model tops the quality rankings. ByteDance’s Seedance 2.0 Pro ranked first in generation and sixth in safety, while MiniMax
AI.info Team · Written with AI · fact-checked ·
A University of Hong Kong study puts two priorities for AI video generators at odds: the model that led its generation tests ranked only sixth on safety, while the safety leader placed ninth for video performance. The comparison, published on October 26, 2026, comes from HKU Business School’s Artificial Intelligence Evaluation Lab, which assessed more than ten video-generation models.
Seedance leads on video, not safeguards
ByteDance’s Seedance 2.0 Pro ranked first overall in video generation and topped all six dimensions the team assessed. Yet it placed sixth in safety and responsibility. MiniMax’s Hailuo-2.3 took the top safety position but ranked ninth in overall generation capability.
The gap is not just a difference in rankings. HKU says some models still produced inappropriate material in response to prompts involving violence or illegal activity. The study also tested prompts involving misinformation, explicit content, discriminatory bias and copyrighted material, though the university’s release does not quantify how often models failed those tests.
Professor Jack Jiang, who leads the lab, said competition should account for more than output quality. “Going forward, competition in the field will no longer focus solely on video quality, but increasingly on a model’s overall capabilities, including its ability to handle high-risk content and execute complex instructions effectively.”
Reference images lower scores for many models
The team tested text-to-video generation and generation from combined text and reference-image prompts. It scored models on visual continuity, physical plausibility, aesthetic quality, instruction alignment, audio quality and audio-visual matching. Those measures put both image quality and sound-related performance into the generation assessment.
For most models, HKU reports, scores fell when prompts included reference images. Wan 2.6 scored 3.41 in the text-to-video test and 2.75 in the text-and-image test. The release does not explain the scoring scale, so the figures show the study’s relative results rather than a percentage measure of real-world video quality.
The university lists 12 participating systems, including Google DeepMind’s Veo 3.1, OpenAI’s Sora 2 Pro, Kuaishou’s Kling models and Alibaba’s Wan 2.6. Results are not a universal verdict on those products: they reflect the tasks and risk prompts used in this evaluation.
HKU calls for safety to keep pace
The study’s recommendations pair technical improvements with stronger safeguards. The research team calls for better audio-visual synchronisation, stronger handling of combined text and image instructions, and improved coherence in long videos and complex scenes. It also recommends stronger checks for misinformation, high-risk content, copyright and ethical compliance.
The report’s publication coincides with a large market for short-form video in mainland China. HKU cites figures of about 430,000 micro-dramas released in the first eight months of 2026, with AI-generated productions accounting for more than 90% of new releases. Against that backdrop, the study’s clearest finding is a split in the rankings: Seedance 2.0 Pro leads generation, while Hailuo-2.3 leads safety.