China's short dramas are 95 per cent AI as compositional video guards fail
About 128,000 AI short dramas were published in China in the first quarter of 2026, triple 2025's full-year total, at 90 to 120 dollars a minute versus roughly ten times that for actors. An arXiv preprint the same week reports 11,024 compositional attacks where harmless inputs combine into harmful video. Current multimodal guards miss them systematically while the industry directly employs 690,000 people.
Artificial Intelligence··Night
128,000 AI short dramas triple last year's China output
Financial Times reporting carried by THE DECODER on 29 August puts first-quarter 2026 short-drama output at about 128,000 titles, three times the count for all of 2025. The China Netcasting Services Association's tally says 95 per cent were made with AI. A minute of AI video now costs 90 to 120 dollars, roughly a tenth of a minute shot with actors, and volume climbed after ByteDance released Seedance 2.0. The shift reaches an industry that directly employs 690,000 people, while livestreaming is the primary job for 15 million people in the country. Tsinghua University professor Shen Yang assesses the cost gap as the mechanism behind the surge in the reporting; the piece does not present that gap as established.[1]
Multi2AV-Safety finds 11,024 compositional video attacks
A preprint on arXiv dated 29 August introduces Multi2AV-Safety, built from 11,024 attack instances covering all 11 non-singleton combinations of text, image, audio and video conditioning for audio-video generation. Its twelve authors report systematic weaknesses in current multimodal safety guards. Harmful meaning can emerge when individually harmless inputs are combined, while explicitly harmful cues become harder to detect once mixed into harmless multimodal context. The authors describe this as a capability gap in compositional risk perception. The work has not been peer reviewed.[2]
Seedance 2.0 arrived as guardrails lagged compositional risk
THE DECODER's 29 August report ties the production surge to ByteDance's Seedance 2.0 release without presenting the cost gap as established. The same week, the Multi2AV-Safety authors warn that guards tuned to single inputs may not see harm that only appears after inputs are combined across modalities. The preprint has not been peer reviewed; the industry figures come from Financial Times reporting relayed by THE DECODER.[1], [2]
