579
Microsoft’s AI boss thinks it’s perfectly OK to steal content if it’s on the open web
(www.theverge.com)
This is a most excellent place for technology news and articles.
If the model isn't overfitted it's also not even copying. By their nature LLMs are transformative which is the whole point of fair use.
So I have a LLM read a book and paraphrase its contents, that's not stealing?
Again, even an exact copy is not stealing. It's copyright infringement. Theft is a different crime.
But paraphrasing is not copyright infringement either. It's no different than Wikipedia having a synopsis for every single episode of a TV series. Telling someone about what a work contains for informational purposes is perfectly fine.
copyright laws are broken. what seems ethical can be illegal and what seems unethical can be legal.
Have I just stolen The Hitchhikers Guide to the Galaxy and given it to you?
You've probably not infringed the copyright, only the court can decide though; if you were to be challenged by the rights holder.
I think there are lots of factors in your defence:
But add in some more quotes, flesh it out, and then try to sell it . . . each step weakens the 'fair use' defence.
This the the problem for the LLM, it can be used for many things, and if it has no filter or limit, then eventually the collective derived works might add up to commercial, substantial reuse, and might include enough to have copied a substantial portion of the original. Very hard to determine I'd think. Each individual use might be fair, but did the LLM itself go too far at some point?
Copyright holder probably struggles to challenge the LLM on the basis of all the things infinite mokeys might use it for in future.
I agree with pretty much everything before this but that particular comment was just talking about summaries, which imo is a lot more cut and dry. (SparkNotes, for example)
An LLM by itself is unlimited and unfiltered, but it's not impossible to limit one and sell it. For all the shit OpenAI deserves to get, I have to give them one thing, their copyright restriction system seems to be on par with YouTube. I paid for a month of it when GPT4 came out and tried my hardest to bypass it, but it won't even give me copyrighted texts when the words are all replaced with synonyms or jumbled around.
I think if someone's offering their LLM as a service and has a system like that in place, they aren't stealing any more than YouTube is stealing. Otherwise I agree that there's a strong argument for copyright infringement.