PirateBay vs AI Training: A Copyright Paradox
The Surface Difference
Look at the two acts mechanically. Pirating Mad Max copies a file, bit-for-bit, and redistributes it. The pirate makes no transformation, no new expression, just a perfect duplicate that competes directly with the original in the market. That's textbook copyright infringement: reproduce the work, distribute the work, done.
Scraping for AI training is different. You're not copying The Pirate Bay or Netflix's file. You're ingesting millions of texts, images, and videos, then squeezing them through a statistical blender to produce a model that outputs new sequences of tokens, pixel patterns, or audio wavelets. Most of those outputs do not match any single training example verbatim. So AI companies lean on a "transformative use" argument—the same fair-use doctrine that lets critics quote a book or artists make parodies.
That's where the clear-cut logic falls apart.
The Transformative Fiction
The key issue nobody wants to talk about: "transformative" is doing a lot of heavy lifting. Yes, the model weights are a mathematical transformation of the training data. But the outputs can be surprisingly close to the originals—style-wise, structure-wise, sometimes even near-verbatim for sufficiently obscure prompts. The model doesn't say "this is a quote from X"; it says "here's something I probabilistically produced." So the consumer gets a substitute for the original work without the original author being credited or compensated.
The pirate gives you a perfect copy and says "it's the movie." The AI gives you a collage of a thousand books and says "it's your prompt's answer." One is obviously theft under current law. The other is... arguable. And because it's arguable, it's litigable, and that's expensive. Anthropic, OpenAI, Meta, whoever—they can afford the litigation. The guy running a torrent website usually cannot.
Where the Real Difference Lies
It's not about legality, ultimately. It's about market substitution and market power.
ThePirateBay directly replaced the thing it copied—you get the movie instead of buying it, and the studio loses a sale. AI training can replace work too, but the harm is downstream and diffuse. A writer loses freelance gigs because clients generate drafts instead of hiring them. That's a real harm, but it isn't written into copyright law as a clean test of "did you copy a file?" And crucially, the AI companies didn't sell the scraped works; they sold the model trained on them. Copyright law was written before machine learning. It has an answer for "you copied my screenplay," but not for "you learned my narrative voice and now offer it as an API."
So the difference is less about morality and more about the legal fiction that transformation erases the original's presence. But if you scrape a million copyrighted books, the model is carrying those books with it, statistically. The question isn't whether the output matches the training data; it's whether the model is a derivative work. I'd argue it often is, but I also know that saying that gets you called a luddite by half the startup world.
The Boring Conclusion
The real answer is that copyright law is a business asset, not a moral compass. The pirate site undermines studios' revenue streams directly, so it's illegal. AI companies are building a massive new market and paying huge