NewsNation

Major tech firms trained AI using stolen YouTube content: Report

FILE - The YouTube app is displayed on an iPad in Baltimore. YouTube has blocked access to videos of a protest song in Hong Kong, days after court approved an injunction banning the song in the city. (AP Photo/Patrick Semansky, File)

Tech giants Apple, Anthropic, Nvidia and Salesforce pilfered data from tens of thousands of content creators on YouTube to train AI.

An investigation by Proof News, copublished with Wired, found that the Silicon Valley luminaries fed subtitles from 173,536 YouTube videos across more than 48,000 channels to AI programs.


The “YouTube Subtitles” dataset was allegedly ripped from the platform and its users without permission, an act that violates YouTube’s guidelines.

“No one came to me and said, ‘We would like to use this,’” David Pakman, host of “The David Pakman Show,” told Proof News.


Best Prime Day Deals for 2024:

Products still on sale after Prime Big Deal Days

Beats headphones and tablets remain marked down

Make sure you’re stocked with Apple products

BestReviews is reader-supported and may earn an affiliate commission.


“This is my livelihood, and I put time, resources, money and staff time into creating this content,” Pakman added. “There’s really no shortage of work.” 

“Apple has sourced data for their AI from several companies,” Marques Brownlee, known by his handle MKBHD, shared on X. “One of them scraped tons of data/transcripts from YouTube videos, including mine. Apple technically avoids ‘fault’ here because they’re not the ones scraping. But this is going to be an evolving problem for a long time.”

YouTubers aren’t the only ones whose content was used for the dataset. The report also found transcripts from online learning resources, including Khan Academy, Harvard and MIT videos, as well as news media outlets such as NPR, BBC, ““The Late Show With Stephen Colbert” and “Last Week Tonight With John Oliver.”

According to research published by EleutherAI, the “YouTube Subtitles” dataset is part of a larger body of language modeling data called the Pile, which mega companies dialed in to for training their AI.

Click here to access a tool that shows which YouTube channels and videos were siphoned into the dataset.