|
Top-level
@ernie you know, I have to wonder if the inaction on prosecuting LLM training companies actually introduced a legal loophole for libraries. Consider that right now, the American legal standard is that GenAI output is considered a derivative work, even if it derived it from 30 billion works. I wonder if the Internet Archive "chunked" editions of books together into a specialized model, could they then "loan" the book out by inferencing a near exact but legally 'distinct' copy of that work? 9 comments
 @mitch Just thinking of how Aaron Swartz—how his good intentions got exploited by people looking to make maximum $$$ and an org actually working in his spirit is being threatened @ernie I miss Aaron a lot, we used to chat through Reddit comments way back when. I'm very glad that he didn't have to see all this, let alone what Reddit and Steve Huffman both turned into.    @mitch @ernie “Inaction” is inaccurate, publishers are exploring lawsuits, but you have to prove the use of copyrighted material in the training of LLM systems which isn’t straightforward unless the company straight-up admits it. It’s comparatively easier for Hachette to sue the Internet Archive since they explicitly reference individual books. |
Gregory's Server
@ernie after all, coaxing an LLM to reproduce a reference work basically in full is pretty established research at this point. We know it's possible — it's how the tech started, by being able to reproduce a ground truth image despite never having actually been exposed to the original file.
I dunno, I'm just some idiot online.