Apple, NVIDIA and Anthropic reportedly used YouTube transcripts without permission to train AI models

Some of the world’s largest tech companies trained their AI models on a dataset that included transcripts of more than 173,000 YouTube videos without permission, a new investigation from Proof News has found. The dataset, which was created by a nonprofit company called EleutherAI, contains transcripts of YouTube videos from more than 48,000 channels and was used by Apple, NVIDIA and Anthropic among other companies. The findings of the investigation spotlight AI’s uncomfortable truth: the technology is largely built on the backs of data siphoned from creators without their consent or compensation.

The dataset doesn’t include any videos or images from YouTube, but contains video transcripts from the platform's biggest creators including Marques Brownlee and MrBeast, as well as large news publishers like The New York Times, the BBC, and ABC News. Subtitles from videos belonging to Engadget are also part of the dataset.

“Apple has sourced data for their AI from several companies,” Brownlee posted on X. “One of them scraped tons of data/transcripts from YouTube videos, including mine,” he added. “This is going to be an evolving problem for a long time.”

Apple has sourced data for their AI from several companies

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 https://t.co/U93riaeSlY

— Marques Brownlee (@MKBHD) July 16, 2024

A Google spokesperson told Engadget that previous comments made by YouTube CEO Neal Mohan saying that companies using YouTube's data to train AI models would violate the paltform's terms and service still stand. Apple, NVIDIA, Anthropic and EleutherAI did not respond to a request for comment from Engadget.

So far, AI companies haven’t been transparent about the data used to train their models. Earlier this month, artists and photographers criticized Apple for failing to reveal the source of training data for Apple Intelligence, the company own spin on generative AI coming to millions of Apple devices this year.

However, Apple told 9to5Mac on July 17 that its OpenELM model doesn't power any of its AI or machine learning features, including Apple Intelligence. Rather, the company said that the model was created strictly for research purposes. Previously, Apple has stated that its Apple Intelligence models were trained on "licensed data, including data selected to enhance specific features, as well as publicly available data collected by our web-crawler," as 9to5Mac noted.

YouTube, the world’s largest repository of videos, in particular, is a goldmine of not only transcripts but also audio, video, and images, making it an attractive dataset for training AI models. Earlier this year, OpenAI’s chief technology officer, Mira Murati, evaded questions from The Wall Street Journal about whether the company used YouTube videos to train Sora, OpenAI’s upcoming AI video generation tool. “I’m not going to go into the details of the data that was used, but it was publicly available or licensed data,” Murati said at the time. Alphabet CEO Sundar Pichai has also said that companies using data from YouTube to train their AI models would violate of the platform’s terms of service.

If you want to see if subtitles from your YouTube videos or from your favorite channels are part of the dataset, head over to the Proof News' lookup tool.

Update, July 16 2024, 3:17 PM PT: This story has been updated to include a statement from Google.

Update July 19, 2024, 2:32 AM ET: Story updated with comment from Apple that its OpenELM model doesn't power Apple Intelligence or any of its other AI or machine learning features.