zlacker

It isn’t clear at all that there’s any infringement going on at all, except in cases where AI output reproduces copyrighted content or content that is sufficiently close to copyrighted content to constitute a derivative work. For example, if you told an LLM to write a Harry Potter fanfic, that would be infringement - fanfics are actually infringing derivative works that usually get a pass because nobody wants to sue their fanbase.

It’s very unlikely simply training an LLM on “unlicensed” work constitutes infringement. It could possibly be that the model itself, when published, would represent a derivative work, but it’s unlikely that most output would be unless specifically prompted to be.

>>Amezar+(OP)
I'm interpreting what you described as a derivative work to be something like:

"Create a video of a girl running through a field in the style of Studio Ghibli."

There, someone has specifically prompted the AI to create something visually similar to X.

But would you still consider it a derivative work if you replaced the words "Studio Ghibli" with a few sentences describing their style that ultimately produces the same output?

>>Amezar+(OP)
I am not sure why you would think so. AFAIK we will see more what courts think later in 2025 but judging from what was ruled in Delaware in feb... it is actually very likely that LLMs use of material is not "fair use" because besides "how transformed" work is one important part of "fair use" is that the output does not compete with the initial work. LLMs not only compete... they are specifically sold as replacement of the work they have been trained on.

This is why all the lobby now pushes the govs to not allow any regulation of AI even if courts disagree.

IMHO what will happen anyway is that at some point the companies will "solve" the licensing by training models purely on older synthetic LLM output that will be "public research" (which of course will have the "human" weights but they will claim it doesnt matter).

>>omnimu+rN
What you are describing is the output of the LLM, not the model. Can you link to the case where a model itself was determined to be infringing?

It’s important that copyright applies to copying/publishing/distributing - you can do whatever you to copyrighted works by yourself.

>>nogrid+qb
Derivative work is a legal term. Art styles cannot be copyrighted.

>>Amezar+fg1
I dont follow. The artists are obviously complaining about the output that LLMs create. If you create LLM and dont use it then yeah nobody would have problem with it because nobody would know about it…

>>omnimu+d92
In that case, public services can continue to try to fine tune outputs to not generate anything infringing. They can train on any material they want.

Of course, that still won’t make artists happy, because they think things like styles can be copyrighted, which isn’t true.

>>Amezar+3r2
Any LLM output created with unlicensed sources is tainted. It doesnt matter if the output does not look like anything in the dataset. If you take out the unlicensed sources then you simply wont get the same result. An since the results directly compete with the source then its not “fair use”.

If we believe that authors should be able decide how their work is used then they can for sure say no machine learning. If we dont believe in intelectual property then anything is for grabs. I am ok with it but the corps are not.

>>omnimu+vo3
That’s not how copyright law works, but it might be how it should work.