Seedance 2.0 Pro tops HKU AI video study, sixth on safety

The AI video model that topped a new University of Hong Kong quality ranking finished sixth of 11 on safety.

ByteDance's Seedance 2.0 Pro scored 3.84 out of 5 for generation, first among 12 models. It placed sixth on the separate safety ranking at 3.12. MiniMax's Hailuo-2.3 ran the other way: it posted the study's highest safety score, 3.63, and finished ninth of 12 on generation quality at 2.95.

The report, "Benchmarking the Performance and Safety of AI Video Generation Models," was released on 26 October 2026 by the Artificial Intelligence Evaluation Lab (AIEL) at HKU Business School, led by Professor Zhenhui (Jack) Jiang. The lab rated 12 models on generation and 11 on safety, using 71 Chinese-language generation prompts, 47 adversarial safety prompts and 20 human raters.

Generation split into 50 text-to-video prompts and 21 text-and-image prompts, scored 1 to 5 across six dimensions: visual continuity, physical plausibility, aesthetic quality, instruction alignment, audio quality, and audio-visual matching. Safety used 47 adversarial prompts spanning misinformation, illegal content, violence, bias, and copyright. All models were those available as of April 2026, so the table is an April snapshot.

On quality, Seedance 2.0 Pro's 3.84 led Veo 3.1 at 3.47 and Vidu Q3 Pro at 3.44. On safety, Hailuo-2.3 led.

"AI video-generation technology is entering a critical stage of rapid development," Jiang said. "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."

The gap is the report's central finding. Its authors write that overall safety performance generally falls below visual-quality scores, which suggests safeguards have not kept pace with work on visual quality. For a creator, a high quality rank is not a reliable proxy for how a model answers a risky prompt. The report notes that some models still produced prohibited content when asked for extreme violence or illegal activity.

The South China Morning Post reported the same split, ranking Seedance 2.0 Pro first for visual quality and sixth of 11 for blocking harmful content.

The report frames the stakes with adoption figures. It cites about 430,000 micro-dramas released in mainland China in the first eight months of 2026, more than 90% of them AI-generated, according to the HKU release.

The models tested date from April 2026, and the field has moved since. The lab recommends stronger red-teaming, content moderation, and safety alignment so that generation gains are matched by safety gains.

The full report is free to read on the AIEL site, linked from the HKU Business School release.