Can someone with actual fundamental understanding of LLMs explain to me why they think it's perfectly legal to train models on copyrighted material? I don't know enough about this. Please don't answer by asking chatgpt.
AI developers will most likely rely on a Fair Use defense. I think this has a reasonable chance of success since, while the use of a given copyrighted work may affect the market for that work (in this case NYT's article), it can be argued to be highly transformative usage. As in Campbell v. Acuff-Rose Music: "The more transformative the new work, the less will be the significance of other factors", defined as "whether the new work merely 'supersede[s] the objects' of the original creation [...] or instead adds something new".
There's also potential for an "implied license", as in Field v. Google Inc for rehosting a snapshot of a site, where "Google reasonably interpreted absence of meta-tags as permission to present 'Cached' links to the pages of Field's site". As far as I can tell in this case, NYT's robots.txt of the time was obeyed, which permitted automated processing of all but one specific article for some reason.
Probably. The question for the courts to decide, then, is how much use is considered fair use.
I have a gadget that will, with some probability, steal your life's savings. It operates through a process that is analogous to a human chewing. When engineering it, we just say for simplicity that the gadget "chews". Of course, that's only a metaphor -- machines can't chew.
But (and here's where your argument gets ridiculous), unless you can quantify the fact that my gadget can't chew, then I will steal your savings. Good luck.
I can think of 2 instances of that machine already. the finance industry fees and an ex-wife.
Isn't this what Mistral AI did?
Not sure about the rest of the world, but at least for US content I don't think any company would publish that LLM.
That's like 40 years before the civil rights movement, and right about the time of the Tulsa massacre.
It's right around when women got the right to vote.
Trying to get it to not say anything horrible under modern standards seems fraught with issues. I don't know if it would even understand something like "don't be racist", given the context it was trained on.
1. Training an LLM is akin to human learning. It is legal to read a textbook about music to learn music, and later to write a book about music which likely includes some of the concepts you earlier learned.
2. Neither the LLM nor the output text contain sufficient elements of the copyrighted work to qualify for copyright protection. Just like if you turned old library books into compost and sold the compost, you wouldn't expect to pay authors of those books a royalty for the compost sales.
If you learn a little too hard though, and reproduce the original textbook in it's entirety, you'll get in trouble.
My guess is that courts will determine that the training itself will not be found illegal, but either the AI companies, or the users, will be found liable for reproducing copywrighted work in output, and no one will want to hold liability for that.
Who owns the copyright then ?
If you ask for Harry Potter and it gives you Bart Simpson it’s useless.
Technology that makes copyright violations easier/quicker have typically been found legal if "the technology in question had significant non-infringing uses".
"I say to you that the VCR is to the American film producer and the American public as the Boston strangler is to the woman home alone."
Then we invented from whole cloth reasons why they were perfectly OK because there was a ton of money to be made and everyone would actually be better off if the VCR was a thing and everyone knew it because it ended up argued after millions of VCRs were already in households.
They're just on a hunt for some extra money.