Runway secretly trained Gen-3 on scraped YouTube videos

It turns out that Runway’s acclaimed Gen-3 video generator had some unauthorized help behind the scenes.

The secret YouTube spreadsheet

According to an internal training spreadsheet obtained by 404 Media, Runway secretly scraped thousands of videos from popular YouTube creators, major brands, and even pirated films to train its latest AI video generator.

The model, initially codenamed Jupiter and officially launched in June as Gen-3, received widespread praise across the AI industry. Runway previously raised $141 million from investors including Google and Nvidia, reaching a $1.5 billion valuation.

When TechCrunch asked Runway co-founder Anastasis Germanidis in June about where Gen-3’s data came from, he dodged specifics, claiming the company uses an in-house research team and "curated, internal datasets."

Why it matters

Having the receipts out in the open changes everything. While industry insiders have long assumed frontier video models rely on mass web scraping, concrete documentation turns whispers into real legal exposure.

It also creates an awkward situation for Runway’s own backers. Google poured millions into the startup, yet its crown jewel, YouTube, was essentially pillaged without permission. As creators and rights holders grow increasingly litigious, the "scrape first, apologize never" playbook is quickly running out of road.