Chinese Models Create 'Deepfake Tsunami'
New Chinese generative video models like C-Dance 2.0 are reportedly not just matching but surpassing the capabilities of OpenAI's Sora and Google's Veo. An analyst warns this has collapsed the barrier to entry for creating sophisticated, realistic deepfakes, forecasting a "deepfake tsunami" as the technology becomes widely and cheaply available.
- ByteDance's Seedance 2.0 distinguishes itself from competitors like Sora and Veo with a "quad-modal" input system that accepts text, images, video clips, and audio references simultaneously, allowing for a high degree of creative control. Another prominent Chinese model, Kuaishou's Kling, is noted for its ability to generate high-speed motion with strong realism and has already reached an estimated US$240 million in annual recurring revenue. - The cost to create synthetic media has plummeted, with some services now available for free, a stark drop from previous years when a minute of high-quality deepfake video could cost between $300 and $20,000. This accessibility is a key driver of the "deepfake tsunami," a term used to describe the anticipated flood of synthetic content. - The financial fallout from deepfake-driven fraud is projected to be substantial, with one forecast estimating that GenAI-related fraud losses in the United States could reach $40 billion by 2027. - Deepfakes have already been deployed in numerous recent elections globally. In the 2024 U.S. New Hampshire primary, an AI-generated audio clip of President Joe Biden was used in robocalls to discourage voting. Similarly, India's 2024 general election saw an estimated $50 million spent on AI-generated content, including deepfakes of politicians and celebrities. - Regulatory frameworks are emerging to address the threat. China has implemented rules requiring deepfake content to be clearly labeled and for service providers to register users with their real names. The EU's AI Act imposes strict transparency obligations, mandating that AI-generated content be disclosed. In the U.S., there is no federal law, but at least 28 states have enacted their own legislation, often requiring disclosures on deepfakes used in political ads near an election. - While Western models like Google's Veo are recognized for 4K cinematic quality and OpenAI's Sora for its physics simulation, Chinese models are competing on different strengths. Seedance 2.0 focuses on multimodal direction and editing flexibility, while Kling excels at animating static images and maintaining character consistency, making it effective for high-volume social media content.
