A key aspect folks should understand about US Copyright Law is that it much more severely penalizes infringement with
distributing, or an intent to distribute, unauthorized copies than to just consume privately. Distributing unauthorized copies is a federal crime (which can escalate to a felony based on various factors) whereas doing whatever for private use is usually a much milder civil liability.
If you look at all the much publicized piracy trials around the Napster era, they invariably came down to the defendant distributing the works they downloaded.
This explains why the AI giants are legally fine doing what they're doing. They are explicitly not distributing copies of the work they acquired (although Meta apparently torrented some stuff, which seeds i.e. distributes data by default, and so might be in trouble to that extent) and training AI has so far been ruled to be fair use. Which, technically, makes sense; encoding patterns gleaned from billions of texts into weights is clearly different from distributing exact copies, and if you look at the cases where LLMs can be made to regurgitate content verbatim, it only really works for a handful of extremely popular books.
This also explains how whatever happened with Aaron Swartz, as tragic as it was, is qualitatively different from what is happening with LLMs. He very explicitly scraped all that content with the specific intention of distributing it. On the other hand, AI companies are literally destroying mountains of books to comply with the law.
You could call this tragic on multiple levels, but not hypocritical.
If you look at all the much publicized piracy trials around the Napster era, they invariably came down to the defendant distributing the works they downloaded.
This explains why the AI giants are legally fine doing what they're doing. They are explicitly not distributing copies of the work they acquired (although Meta apparently torrented some stuff, which seeds i.e. distributes data by default, and so might be in trouble to that extent) and training AI has so far been ruled to be fair use. Which, technically, makes sense; encoding patterns gleaned from billions of texts into weights is clearly different from distributing exact copies, and if you look at the cases where LLMs can be made to regurgitate content verbatim, it only really works for a handful of extremely popular books.
This also explains how whatever happened with Aaron Swartz, as tragic as it was, is qualitatively different from what is happening with LLMs. He very explicitly scraped all that content with the specific intention of distributing it. On the other hand, AI companies are literally destroying mountains of books to comply with the law.
You could call this tragic on multiple levels, but not hypocritical.