Forum

How You Collect Matters: YouTubers Test an Anti-Circumvention Theory

Close-up of a gray cast metal panel with a dense lattice of diamond and cross-shaped openings

Most lawsuits over AI training ask whether using creative work to train a model is fair. A new set of cases asks a different question: how was the work obtained in the first place?

On 3 April, YouTube creators filed proposed class actions against Apple, Amazon and OpenAI. The plaintiffs, who include the company behind the h3h3 Productions channels and two golf channels, allege that the companies circumvented YouTube’s technological protection measures to download millions of videos and use them to build text-to-video systems: Amazon’s Nova Reel, OpenAI’s Sora and Apple’s video models. The complaints describe automated downloading tools and rotating IP addresses used to avoid detection. The same creators sued Snap in January over similar claims involving its AI products.

The legal basis is Section 1201 of the US Digital Millennium Copyright Act, which prohibits getting around technical measures that control access to copyrighted works. The plaintiffs are seeking statutory damages and injunctions. The companies have not had their day in court, and the allegations remain allegations.

Why this theory matters

Fair use arguments turn on what happens after copying: whether a use is transformative, and whether it harms the market for the original. Courts have reached different conclusions, and the cases are slow.

Anti-circumvention claims look earlier in the process. If a court agrees that a platform’s restrictions are technological protection measures, and that a company deliberately got around them, the question of fair use may matter less. The way data was gathered becomes the issue.

That logic is familiar from other recent disputes. In the books settlement last year, how copies were acquired turned out to be central. These video cases apply the same instinct to a new medium.

What it means for visual creators

Designers, animators and filmmakers who publish on video and image platforms rarely think about the technical measures those platforms use. They may matter more than expected.

  • Platform protections are part of your rights picture. Where you publish affects how hard it is to collect your work legitimately, and what claims might exist if someone does it anyway.
  • Terms of service are not the same as law, but they shape it. Understanding what a platform permits, and what it prohibits, helps you understand your position.
  • Records help. Knowing where and when your work was published can matter if a dataset ever has to be traced.
  • Keep your masters. High-resolution originals and project files that never leave your control are both a creative and a legal asset, and they show what you made before anyone copied it.

A note for students and independent model builders

The same principle applies in reverse. Many students and independent artists now assemble their own datasets to fine-tune models. It is tempting to treat anything visible in a browser as available. These cases are a reminder that downloading tools, rotating addresses or ignoring access controls can turn an experiment into a legal problem, quite apart from the ethics of using other people’s work.

Building on your own material, licensed collections or clearly open sources is slower. It is also the only foundation that holds up when someone asks how the dataset was made.

How competitions see datasets

The International AI Award asks entrants to document the tools and models behind their work and, for code-based projects, to submit the source code and explain any modifications to open-source models. It also requires entrants to own, or have permission for, everything they submit. A clear account of where training material came from fits naturally into that kind of documentation, and juries are likely to read it closely.

The direction of travel

These cases will take time, and they may fail. But they signal a shift in how creators and their lawyers think about AI training. Instead of arguing only about what models do with creative work, they are asking how that work was taken. For anyone who builds with data, the answer to that question is becoming part of the work itself.