How a Searchable Database is Exposing YouTube's Role in Training AI Video Tools

Journalist Alex Reisner, writing for The Atlantic, has documented millions of YouTube videos that were used without creators' permission to train AI video generation tools — a disclosure published as part of The Atlantic's AI Watchdog project, which investigates where generative AI systems get their training data.

Reisner has built a track record in this kind of work. In 2023, he obtained a dataset showing that Meta had used over 191,000 books without permission to train its generative AI systems — a finding that became central to copyright lawsuits against AI companies. The YouTube investigation follows the same approach: locate the training data, identify who owns the rights, and make the results publicly searchable.

That searchability piece matters in practical terms. The Atlantic's AI Watchdog tool lets you search directly through the datasets Reisner has uncovered. Creators, authors, and publishers can type in their name or work and find out whether it appears in datasets used to train commercial AI models. This has real teeth in courtrooms: copyright lawsuits against AI developers increasingly need named plaintiffs who can point to specific instances of their work being used without permission.

That YouTube videos ended up in AI training datasets is not surprising to anyone following the generative video space. When teams build tools that create video using techniques like diffusion (a method of generating images by starting with noise and gradually refining it) or transformers (a type of AI architecture that learns patterns in sequences), they face a straightforward choice: scrape video from the internet at scale. YouTube offers an unusual concentration of labeled, captioned video spanning virtually every visual subject matter. What Reisner's reporting adds is not speculation but documented evidence — which specific videos, how many of them, and which products used them.

The consequences ripple across several groups. For individual creators, the issue is whether their content generated commercial value for an AI model they received nothing from. For the AI developers involved, the legal question is whether grabbing publicly available video counts as fair use — a doctrine that U.S. courts have never clearly defined for text-based AI, and even less for video, which carries additional legal layers like synchronization rights for music and terms of service that explicitly forbid scraping for machine learning. YouTube itself has prohibited this kind of data collection for years; whether that rule actually binds the companies that build models from the stolen data is still a matter of active legal debate.

The books case from 2023 offers a comparison, though not a perfect one. The dataset Reisner exposed then consisted of plain text extracted from a shadow library. YouTube video is more complicated: the legal claims are stacked higher, the file sizes are vastly larger, and the technical relationship between the training video and the final model is harder to pin down. A video diffusion model doesn't save and replay the original frames the way a search engine might cache a page — that technical detail could matter in court, but judges haven't yet drawn a clear line around it.

There is a shift worth noting in how this affects the broader power balance. For years, creators had no practical way of finding out if their work had been fed into a training dataset. Discovering this through the legal system is slow and expensive. A free, searchable public database that answers the question instantly changes who has leverage in these disputes — and it likely means more lawsuits will be filed sooner.

Over the past few years, we've watched training data secrecy start to crack. Reisner's 2023 reporting on books broke that silence for text-based AI. This YouTube investigation does the same for video. Each new disclosure makes the next one easier to uncover, and each searchable database The Atlantic publishes creates a public record that courts, regulators, and lawmakers can draw on. The industry spent its early years betting that scale and speed would keep scrutiny at arm's length. That calculation appears to be shifting.