AI training dataset used by tech giants allegedly created by scraping YouTube videos in violation of terms

Non-profit AI research group EleutherAI scraped YouTube subtitles to create a dataset in violation of YouTube’s terms of service, ProofNews said on July 16.

The dataset, called the Pile, allegedly includes subtitles of 173,536 YouTube videos from over 48,000 channels. About 12,000 deleted videos are part of the dataset.

Several top tech and AI firms, including Anthropic, have since used the Pile for training. Anthropic spokesperson Jennifer Martinez said the dataset includes “a very small subset of YouTube subtitles” but declined to comment on possible violations of YouTube’s terms of service.

Business software firm Salesforce also used the dataset. Salesforce VP of AI research Caiming Xiong said the dataset was “publicly available” and that Salesforce used it for academic and research purposes. ProofNews said Salesforce eventually released the same dataset publicly.

Apple used the Pile to train OpenELM, an efficient language model for on-device AI. Nvidia, Bloomberg, and Databricks also used the Pile for AI training.

ProofNews said its list of companies that used the dataset is not comprehensive, as companies do not always disclose which datasets they use in AI training.

Dataset contains crypto channels, more

ProofNews’ search tool indicates that Pile includes videos from crypto channels and creators, including Coinbase, Cointelegraph, Bitcoin Magazine, BitBoy Crypto, 99Bitcoins, Ivan On Tech, and Andreas Antonopolous.

ProofNews highlighted that the dataset includes transcripts from major news channels, education channels, late-night shows, popular YouTube hosts, and other categories. The Pile dataset extends beyond YouTube to other websites and online content.

ProofNews noted an earlier report from the New York Times, which said OpenAI and Google had previously harvested YouTube text. Google, which owns YouTube, said the action was permissible due to its agreement with users. OpenAI did not confirm or deny the report.

AI copyright disputes are far-reaching. Law firm Baker Hoestler lists at least fifteen lawsuits involving tech firms such as Anthropic, Meta, GitHub, Stability AI, Nvidia, and Google. OpenAI faces high-profile lawsuits from Mother Jones’ parent company and The New York Times.

Mentioned in this article