Nobody that's familiar with copyright precedent seriously thinks training on copyrighted data is illegal. What's illegal is simply reproducing the copyrighted c

Nobody that's familiar with copyright precedent seriously thinks training on copyrighted data is illegal. What's illegal is simply reproducing the copyrighted content for others. See Google Books for an interesting precedent. Google still scans every book ever written without paying licensing fees. They just can't legally make all those scans freely available to all. Clearly, some of the responses GPT gives to users have infringed training data copyright, but the majority of their responses do not. They basically have 3 options: 1. Figure out how to engineer an LLM so that it reliably avoids "unfair" use of copyrighted content. "Fair use" is a legal doctrine with no rigorous definition, so this would be very difficult even if they had a clue where to start. I wouldn't hold my breath for this. 2. They can continue without any licensing, and field copyright lawsuits on a case-by-case basis for each individual prompt and response. That would be a logistical nightmare for the courts, plaintiffs, and defendant alike. It would certainly stress test the whole system, possibly result in knee jerk legislation that OAI may not like, or simply bury them in legal fees if infringement is common enough (which is not entirely clear yet). 3. They can strike deals, eliminating legal uncertainty and allowing them to plow ahead with reproducing copyrighted content without worries, while also getting other goodies like exclusivity deals at the same time. Seems like a no brainer to me