Also, it's worth noting in the example of ChatGPT emitting LGPL code without attribution or license, the code is actually different [1]. Is the difference enough to circumvent a copyright violation claim? I don't know but a big part of determining whether it does is now muddled because of the way the system was designed. Even if we could get an entropy distribution on which training data was used to generate the text, it's not even clear the courts could use it in any meaningful way.
[0] https://ansuz.sooke.bc.ca/entry/23
[1] https://twitter.com/DocSparse/status/1581461734665367554