Runway's AI Video Generator: Powered by Thousands of Scraped YouTube Videos for Enhanced Content Creation

Runway has trained its AI text-to-video generator using thousands of YouTube videos and pirated content, according to a report by 404 Media. An obtained spreadsheet of training data lists YouTube channels owned by major entertainment companies such as Netflix, Disney, Nintendo, and Rockstar Games, as well as popular creators like MKBHD, Linus Tech Tips, and Sam Kolder. It also includes links to channels owned by news media outlets like The New Yorker, Reuters, and Wired. A former Runway employee noted, “The channels in that spreadsheet were a company-wide effort to find quality videos to build the model with.” This data was subsequently used by a web crawler to download videos while bypassing Google’s restrictions.

Runway, an AI startup backed by significant funding from Alphabet (Google's parent company) and Nvidia, has developed powerful tools for creating realistic AI-generated videos and animations. Its latest tool, Gen-3 Alpha, launched in June, boasts the capability to generate videos in virtually any style. Like many AI models, Gen-3 Alpha requires a diverse dataset for training.

The investigation by 404 Media also revealed that Runway's dataset potentially includes links to piracy sites like KissCartoon, which offers free access to anime and other animated content. It's currently unclear whether all the videos listed in the spreadsheet were used to train the Gen-3 Alpha model. In June, Runway cofounder Anastasis Germanidis mentioned that the company relies on “curated, internal datasets” for training but did not elaborate further.

In response to inquiries about the use of YouTube videos for AI training, Google directed attention to a statement from YouTube CEO Neal Mohan, who emphasized that training AI models on the platform's content violates their policies. Other companies, including OpenAI, have also faced scrutiny regarding their AI training data linked to YouTube. OpenAI’s CTO, Mira Murati, expressed uncertainty about whether their text-to-video generator, Sora, trained on YouTube. Additionally, a report from Proof News and Wired indicated that Anthropic, Apple, Nvidia, and Salesforce have trained their AI models on over 170,000 YouTube videos.