How YouTube became a gold mine for AI Chatbots.

From American Public Media, this is Marketplace Tech.

I'm Lily Jramale. companies competing in the chatbot wars have consistently turned to something known in the industry as the pile

to train their large language models.

The pile is a massive trove of open source data, about 800 gigs worth, that's made up of several smaller data sets. data in Parliament's website, even a collection of scandalous emails from employees at the now

defunct Energy Company Enron.

Investigative reporter Annie Gilbertson of Proof News recently took a deep dive into the pile and discovered something else.

A data set called YouTube subtitles.

Text from videos are now being used by Silicon Valley heavyweights like Apple,

and Vidia, and Anthropic.

Gilbertson says it's happening without the consent of creators.

What we found was more than 170,000 YouTube videos had been swiped to train these artificial intelligence

models. That includes channels labeled as education, Khan Academy, MIT, Harvard, and also news publishers, Wall Street Journal,

NPR, BBC, entertainers, and then of course some of YouTube's biggest stars, Marquez Brown

Lee, Mr Beast,

Puy, for example.

Mm-hmm.

So Annie, how did you and your team go about figuring out

whether companies were using YouTube

and how they were using it.

Yeah, so tech companies are just very secretive about the type of

training data that they're using.

...