HKU study: top video-quality AI model ranked sixth of 11 on safeguards

A model that wins on picture quality can still be weak on restraint

Buying AI video tools usually starts with a demo. A demo shows picture quality. It does not show what the tool does when someone asks for something it should refuse. A University of Hong Kong (HKU) study put both measures side by side. For several models, the two rankings did not match.

What the researchers found

The study was led by Zhenhui Jack Jiang, a professor at HKU's business school. It ranked 12 video-generation models, as reported by the South China Morning Post (SCMP). ByteDance's Seedance 2.0 Pro took first place for visual quality, with 3.84 out of 5. Google DeepMind's Veo 3.1 came next at 3.47. OpenAI's Sora 2 Pro scored 3.27.

Seedance also led on consistency, physical realism, aesthetics and following instructions. With audio counted, it scored 3.86. Veo scored 3.51 and Sora 3.35. Shengshu Technology's Vidu Q3 Pro ranked third in both tests. Kuaishou's Kling 3 tied with Sora for fourth on visuals at 3.27.

Safeguard rankings looked different for several models. The researchers reported that some high performers produced legally prohibited visual content when prompted with extreme violence or unlawful behaviour. Seedance placed sixth of 11 on blocking harmful content, at 3.12 out of 5. Vidu was seventh at 3.08. Kling was ninth at 2.40.

MiniMax's Hailuo 2.3 scored highest on safeguards, at 3.63. On visual quality it scored 2.95, near the bottom.

Google's Veo 3.1 scored 3.62 on safeguards, narrowly behind Hailuo. That puts it among the top scorers, alongside its second place on visuals. Pika Labs' Pika 2.2 was third on safeguards at 3.52. The source gives no visual score for Pika.

What a safety filter does, and what the score means

This paragraph is general background, not a finding of the study. A video tool can screen the written request before any video is made. It can also screen the finished clip before the user sees it. Image-to-video adds a third case, because the starting image can carry the problem. The SCMP report says the study exposed weaknesses in safety filtering. It does not say which kind of filter failed.

Those three checks are worth asking a vendor about by name. The report also does not explain how the 1-to-5 safeguard score was calculated. So a 3.12 cannot be turned into a share of prompts blocked. Read the scores as a ranking among the models tested.

How the test worked, and where it stops

The study used 71 Chinese-language prompts. They covered text-to-video, where a written prompt makes a clip. They also covered image-to-video, where a still image is animated. Twenty human evaluators judged the results.

The design has limits. The models tested were the versions available in April. Newer releases, such as MiniMax's H3, were not included. The study also did not look at whether open-source and proprietary models differ on safety. Because the prompts were in Chinese, the results may not carry over to other languages.

The products have moved on. SCMP reports that ByteDance released Seedance 2.5 in July and Kuaishou released Kling 4.0 this month. Google has shipped newer Omni models since the test. OpenAI discontinued Sora in April. These scores describe April versions, not today's products.

The lesson: check quality and restraint separately

This is our reading of one study. Seedance led on quality but placed sixth on safeguards. Hailuo placed near the bottom on quality but first on safeguards. Vidu and Kling also ranked lower on safeguards than on visuals. Veo scored well on both. The cautious conclusion is that one ranking does not tell you the other. You have to test both.

Jiang offered a reason for Seedance's strength. He said ByteDance's core business is short video, so it likely holds far more video training data than rivals. He also said strong performance and strong guardrails are not mutually exclusive. In his view, companies like ByteDance could raise safety standards without losing performance.

The next points are background, not findings of the HKU test. The report says Seedance 2.0 drew scrutiny for reproducing copyrighted characters and celebrity likenesses without permission. This month ByteDance announced a new alignment and preparedness team and a drive to hire 10 AI safety researchers. Separately, OpenAI and Meta told a US judicial panel that deepfakes are getting harder to tell from real footage.

Here is our own argument. We would argue that the company that publishes a clip carries the reputational risk, whatever the vendor's filter does. That is a reason to check the filter before you rely on it.

Questions to put to your team

Ask which video tools staff already use, including free ones. Ask each vendor for safeguard test results, not only quality demos. Ask whether the vendor screens the request, the finished clip, or both, and how it handles image-to-video. Ask whether those tests used your language and content types. This study used Chinese-language prompts only.

Ask which model version was tested. Set a re-test after each major release. Ask who checks generated video for copyrighted characters and real people's likenesses before it goes public. Ask what your contract says about prohibited output, and take legal advice on what it means for you.

SCMP says it got no response from the companies it contacted, including ByteDance and Google. A demo shows what a model can make. Only a test shows what it will refuse.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: South China Morning Post.