A report from 404Media claims that AI-based video generation software Jupiter has been using a database of videos scraped extensively from YouTube and other sources to train its models. Jupiter is run by RunwayAI, a $1.5 billion startup backed by industry giants like Google and Nvidia.
The data collected by 404 includes a spreadsheet listing the YouTube channels of some of the world's largest media groups and content creators, as well as websites hosting pirated content and links to specific videos.
An anonymous source told the publication that the spreadsheet was part of a "company-wide" effort to scrape content from the internet, and the report is the latest in a series of controversies over companies using creator data for model training without paying them anything.
Spreadsheets were used to feed data to crawlers that downloaded videos through proxies, sources said. Details show that the so-called spreadsheet contains links to YouTube channels of Netflix, Disney, Sony, Pixar, ViceNews and other well-known media. It also focuses on content creators such as popular Apple vlogger Marques Brownlee and lifestyle blogger Casey Neistat.
In addition to the spreadsheet, 404 Media also contacted a source who claimed to have worked at "Runway". While the person's identity is confidential, they provided key details about how spreadsheets are used within the company to train video-generating AI models.
According to them, Runway used spreadsheets to feed content to an open source software that scraped content from YouTube. In addition to channels, the spreadsheet also links to specific videos as part of a "company-wide effort to find quality videos to build models on."
Runway did not respond to 404's request for comment, and Google referred the publication to its statement from earlier this year, when it commented in April that OpenAI's Sora video generator violated YouTube's rules by using YouTube videos to train its model.
Runway also allegedly had employees filter videos through keywords that focused on specific types of content, according to the source. Videos are also categorized by subject, such as animated shorts and student films. The scope of the search is not limited to YouTube, but websites that operate pirated content are also targeted.
Jupiter is the internal codename for the RunwayGen-3 model, and 404's Gen-3 test prompts generated content similar to videos allegedly scraped from YouTube. The publication added that after contacting Runway for comment, the model stopped generating these videos.
Runway is one of the more popular artificial intelligence video generation companies. Its latest round of financing was Series C financing in June last year, valuing the company at US$1.5 billion. In the same month, it was also selected into the "100 Most Influential Companies" by Time magazine in the United States. There are more than 12 artificial intelligence companies on the list.