The Future. Investigations by Wired and Proof News have found that firms like Nvidia, Anthropic, Apple, and Salesforce have trained their systems on copyrighted YouTube videos. If YouTube decides to take any of these companies to court, it could radically alter the development of AI systems and put power back in the hands of creators.

Swipe softwareAI companies have been very busy scraping content off the internet under the veil of it being “publicly available.”

  • Wired and Proof News discovered that top AI firms have trained their systems on a dataset that includes the plain text subtitles of 173,536 YouTube videos from 48,000 channels in various languages.

  • Those channels include Khan Academy, MIT, Harvard, WSJ, NPR, The Late Show with Stephen Colbert, Jimmy Kimmel Live!, MrBeast, and PewDiePie.

  • The dataset, called “YouTube Subtitles,” was collated by EleutherAI and included in a release called “The Pile” — a dataset that Big Tech has admitted to using for AI training.

  • Additionally, OpenAI has been vague on whether its upcoming video generator, Sora, was trained on the video of scraped YouTube content.

Of course, all of this is strictly against YouTube’s protocol, and likely, copyright law (several lawsuits are making their way through the courts).

But, it looks like if your content is anywhere online, AI firms now consider it fair game.

The post AI firms have scraped thousands of YouTube videos appeared first on TheFutureParty.

Reply

or to participate