I think we just need to rethink copyright for language models. I'm okay just licensing 1 copy of a work to any LLM model throughout its various generations. Just don't pirate it if no special license is available, buying the ebook should suffice. It should be no different from a human buying a copy. The rule should only be that it does not leak the entire work.
I'm not OK with that, though... and here we have the nut of the problem. There is no agreement as to what's acceptable and what's not.
I personally think that the odds of me me being able to both publicly publish my words and code and be able to keep them out of training data is pretty close to zero. Since that's unacceptable to me, my only option is not to publish that stuff at all.
