Hacker News Reader

Comment by Earw0rm | Hacker News Reader

In a strict reading, the tool itself is a copyright violation. In that an LLM is the code and its weights, and the weights contain a lossy compressed copy of a copyrighted work.

Now you might say that's an overly rigid reading of copyright law, and I might agree, but that is nevertheless the reading of it applied to other domains - music, film and so on - where incorporating even a fractional, lossy and heavily processed portion of another work is enough to trigger copyright law. There is no blanket de-minimis that says "well, it's only 5%, so it's fair use".

View on HN

The legality of Copyright material being present in the latent space of a commercialized foundation model would definitely be an interesting case!

Though most of these cases aren't litigating this. They are charging that the ability to recreate copyrightable material is the fault of the tool and not the tool user[1].

[1]: https://www.darrow.ai/resources/ai-copyright-law

If I measure the length of a song? Not copyright infringement. If I take even a very short sequence of PCM samples, copyright infringement.

But the practical thresholds appear to be different, there's a bunch of ways you can extract kbs' worth of information from a song, some are considered copying (ripping the melody to midi with a converter, say), some are not.

And in the case of "AI", depending on the model it could be extracting and encoding almost anything, because it's a catch-all term encompassing dozens of architectures and methods.

If using the work as training data was lawful (which hasn't been always the case as we know today) I tend to agree with you, but it might depend: creating an alteration of the work might be prohibited altogether or you might need to give proper credit to the original authors and that might already be violated when you present certain output to a user. In general, the user is responsible to obey the license of the original work. The problem is that the LLM doed not know on which works its output (mostly) is based on. So in theory, no output for which copyright laws might apply should be used whatsoever.

It's not cut and dried in that way ("certain output"), in many models, any one piece of training data can influence any number of weights and potential outputs, and it's difficult or impossible to say which output was influenced to what extent by which input.

A drop of blood falls in a lake.. which bit of lake now contains blood?

> the user is responsible to obey the license of the original work

The user is in this case the "AI" company.

They already violated law almost certainly by copying stuff from the net in the first place. (Almost certainly as it by now does not look like this was fair use, and there is no other means to make legal what they did.)

That's not true.

Some people (like the NYT¹) are in fact demanding before curt the destruction of the trained models as they obviously contain copyrighted content.

¹ https://pressgazette.co.uk/media_law/new-york-times-open-ai-...

Reply on news.ycombinator.com