But does Reddit still control the legal rights? Who holds the copyright, and rights to use the data for various purposes?

If some new AI was trained on those backups, could Reddit take the creators to court for using that data without a license?

The users own their comments afaik, reddit only holds a license to them.

Are closed source products forced to reveal their training materials? Would be interesting to see a new company come out be closed source have a killer product then get forced to reveal training data in court only to reveal they pretty much only used stuff they didn't have rights to. Is it technically illegal yet to "learn" from data you don't have ownership of yet?

Take that one step further - what about learning on generated data from a model that used copyrighted data? It is exactly what MS "Orca" does. Does this step launder copyright?

If you tried a language model pre 2017, you'd see it was not even able to produce two lines of coherent text. Now we are past coherence to the point we complain it bullshits in an authoritative style. That's of course because we use writing style as a shortcut for writing quality, not the best idea.

If somebody uses it and achieves some success, you can bet the truth will come out. Whether whistle blowers or crafty people on the internet piecing it together, people are pretty good at figuring that sort of stuff out.

ahem OpenAI ahem

User-generated content belongs to the author. You need their permission if you want to use it. I don't know what the limit of a TOS is, I imagine misrepresentation for example is not allowed even if you write it inthere.

how would you prove what stuff was trained on? it's computationally infeasible to do so

Yep, a fuckton of it has been archived.