Apple, Anthropic and Other AI Firms Have Reportedly Trained AI Models on Thousands of YouTube Videos
Apple, Anthropic, and other major artificial intelligence (AI) firms have reportedly trained AI models on data from hundreds of thousands of YouTube videos. A new report claims that multiple AI companies used a publicly available dataset called Pile which contained the plain text of videos’ subtitles without any video imagery. The data was collected from popular YouTube creators such as MrBeast, Marques Brownlee, and PewDiePie as well as Indian YouTube creators such as CarryMinati, BB ki Vines, and Ashish Chanchlani.
Multiple AI Models Reportedly Trained on YouTube Videos
Proof News conducted an investigation to find that subtitles data from as many as 1,73,536 YouTube videos were taken from more than 48,000 channels. As per the report, EleutherAI, a non-profit AI research lab, curated this dataset. Later, it was used by companies such as Apple, Anthropic, Nvidia, Salesforce, and more. Notably, the AI lab published a research paper highlighting the details of the dataset.
EleutherAI created a data repository of 800GB dubbed Pile and made it publicly available for those who wanted to train AI models but could not afford large datasets. The majority of the dataset was taken from publicly available sources such as English Wikipedia, e-books, and more. However, it also contained the subtitles from all the videos compiled in a dataset called YouTube Subtitles.
The report claimed that the Pile was used to train Apple’s OpenELM AI model, on the basis of the research paper’s description. Salesforce, Nvidia, and Anthropic’s AI models’ research papers also reportedly mention the usage of the dataset.
Anthropic spokesperson Jennifer Martinez told the publication in a statement, “The Pile includes a very small subset of YouTube subtitles. YouTube’s terms cover direct use of its platform, which is distinct from use of the Pile dataset. On the point about potential violations of YouTube’s terms of service, we’d have to refer you to the Pile authors.”
Notably, YouTube’s terms of service prohibit anyone from accessing the videos on the platform using automated means such as robots, botnets or scrapers. YouTube Subtitles will fall under the scraping category. A Google spokesperson told Proof News in an email response that the tech giant has taken “action over the years to prevent abusive, unauthorised scraping.” However, no comments were made about AI firms’ usage of the data.
In a post on X (formerly known as Twitter), Marques Brownlee called out Apple for sourcing data from companies that included his videos’ transcripts, but he also highlighted that it was not the iPhone maker’s fault since they did not collect the data.
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
While this dataset was collected and distributed publicly, there could be other instances of data scraping on platforms such as YouTube. With AI firms scrambling to find more data to train their large language models (LLMs), data procurement might continue to enter similar legally grey areas.