OpenAI tapped a million hours of YouTube audio for GPT-4

AI companies have a serious hunger problem, and they are running out of internet to eat.

According to an extensive report from The New York Times, OpenAI exhausted its supply of useful online text back in 2021. Desperate to feed GPT-4, the team reportedly built its Whisper audio transcription tool specifically to scrape and transcribe over one million hours of YouTube videos.

The Times reported that OpenAI knew this was legally questionable, but pushed forward anyway under the assumption that it qualified as fair use. OpenAI president Greg Brockman was even personally involved in gathering the video URLs.

Everyone is bending the rules

OpenAI spokesperson Lindsay Held told The Verge that the company curates unique datasets from public sources and non-public partnerships, while also exploring synthetic data. But before turning to YouTube, podcasts, and audiobooks, OpenAI had already hoovered up GitHub repositories, chess move databases, and Quizlet homework materials.

Google officially pushed back. A Google spokesperson stated that its robots.txt files and Terms of Service strictly forbid unauthorized scraping of YouTube videos, echoing recent warnings from YouTube CEO Neal Mohan regarding Sora.

Yet the Times revealed that Google was doing the exact same thing. Google reportedly transcribed YouTube videos to train its own models, though the company claims it only used content within creator agreements.

Google even went a step further. Its legal team reportedly pushed the privacy group to alter policy language to harvest consumer data from apps like Google Docs. To avoid public scrutiny, the update was quietly rolled out on July 1st, right before the Independence Day holiday weekend.

Meanwhile, Meta was caught in its own scramble. Internal audio obtained by the Times showed AI researchers discussing the unauthorized use of copyrighted works to catch up with OpenAI. After burning through virtually every available English book, essay, poem, and news article online, Meta leadership even floated the idea of outright buying a major publishing house.

Why it matters

We are watching the frontier AI labs collide with the physical limits of human knowledge. The Wall Street Journal recently noted that AI demand for high-quality text could outpace new human-generated content by 2028.

For everyday creators and web users, this is an uncomfortable reality check. Anything you post online, from video essays to private drafts sitting in cloud office suites, is being eyed as high-octane fuel for the next frontier model. Companies are betting on quiet terms-of-service rewrites and fair-use defenses to shield themselves from legal blowback.

In our view, this data frenzy exposes a fragile foundation beneath the current generative AI boom. Transcribing a million hours of video or slipping privacy changes past users over a holiday weekend are not signs of sustainable technological breakthroughs; they look like genuine desperation. Brute-force scaling is hitting a wall, and until synthetic data or alternative architectures prove they can reliably bridge the gap, the AI race will remain a legally treacherous land grab.