logoalt Hacker News

thephyber • yesterday at 11:44 PM • 0 replies • view on HN

I'm not going to address or defend your parent's comment here, but yours has a lot to be desired.

You are wrong on the legality of how LLMs train. US District Courts have ruled that it is fair use to train LLMs on copyrighted materials (like books)[1]. It is not "stolen" when you read a copyrighted book that you borrow from a library (which is the appropriate analogy for the vast majority of LLM training content). Your use of the word "stolen" here is either disingenuous or hyperbole. We know that Meta did train their LLM on some stolen / torrented content, but that is far from the majority of the training corpus.

The "inspiration" topic isn't "legal gaslighting". Humans learn via inspiration. It is the foundation of culture and the arts. You can't use a single English idiom without having heard and copied it before. But we don't sue for copyright infringement of idioms, for example. We have lots of "fair use" concepts established by law because the law recognizes that much of content produced everyday is derivative but mundane.

The important question then becomes: if you prompt an LLM for a unique work (making an attempt to avoid it regurgitating a single source verbatim) and it still outputs a verbatim copy of one of the training materials, what responsibility does the prompting user have to verify that the LLM's output content isn't violating IP (and also not covered by fair use)? This is the point of my comment above.

There are LLMs which are being trained entirely on permissively licensed content. But the quality of the output is bad because there simply isn't enough training data to train those models sufficiently. Hence there is pressure to make models usable by training on ALL data, not just permissively licensed data.

[1] https://www.linkedin.com/posts/andrewyng_judge-rules-trainin...