OpenAI sued for training models on scraped YouTube videos

AI labs are running out of text on the open internet, so they started raiding YouTube's audio tracks.

Now the bill might be coming due. On August 5, a YouTube creator based in Massachusetts, David Millette, filed a class-action lawsuit against OpenAI in the U.S. District Court for the Northern District of California. The suit alleges that OpenAI secretly transcribed millions of YouTube videos without creator consent or payment, using the data to train the models behind ChatGPT.

Represented by Bursor & Fisher, Millette is seeking a jury trial and over $5 million in damages for creators whose work ended up in OpenAI's training sets. The complaint argues that OpenAI profited heavily from creator content while infringing copyright and violating YouTube's terms of service, which ban third parties from scraping videos for independent apps.

Inside OpenAI's million-hour scramble

According to the complaint, OpenAI built products that command lucrative subscriptions on the backs of uncredited, uncompensated creators.

This isn't just speculation. In April, The New York Times reported that OpenAI developed its Whisper speech-to-text model partly to solve its looming data crunch. A team that included OpenAI president Greg Brockman transcribed more than one million hours of YouTube audio with Whisper, funneling the text straight into GPT-4. Internal staffers even debated whether the move breached YouTube's policies.

OpenAI is hardly alone. In July, Proof News revealed that Anthropic, Apple, Salesforce, and Nvidia trained models on The Pile, a popular dataset packed with subtitles from hundreds of thousands of YouTube videos. Apple later claimed it had no plans to use those specific models for consumer AI features. Even Google, YouTube's parent company, quietly amended its terms of service last year to clear the way for using user data in its AI pipelines.

Running out of road

Tech companies are turning to video transcripts because the public web is shutting its doors.

More than 35% of the world's top 1,000 websites now block OpenAI's web crawlers, according to Originality.AI. Research from MIT's Data Provenance Initiative shows that roughly 25% of data from high-quality sources has been pulled from major AI datasets. Research outfit Epoch AI estimates developers could exhaust the pool of usable training data between 2026 and 2032.

August kicked off on a rocky note for OpenAI across the board. The same Monday, Elon Musk refiled his lawsuit against OpenAI and Sam Altman, doubling down on claims the company abandoned its nonprofit mission and adding racketeering allegations.

Why it matters

For creators, this suit draws a line in the sand. Every video script, voiceover, and spoken tutorial is now vulnerable to being converted into commercial training material. A $5 million damages demand won't make anyone rich after legal fees, but a creator-friendly verdict would force AI companies to negotiate collective licensing deals instead of treating the web as a free buffet.

For the AI sector, the "fair use" defense is showing cracks. Scraping publicly available text was already controversial, but converting private audiovisual media into synthetic training text sits on even shakier legal ground.

OpenAI's decision to transcribe one million hours of YouTube audio was an undeniably clever short-term hack to power GPT-4. But we think that shortcut created a massive legal liability. As crawling gets blocked and lawsuits pile up, the next frontier in AI won't be about who has the biggest cluster, but who can secure legitimate access to real human knowledge without getting sued into oblivion.