So the real debate is whether training a model counts as copyright infringement. Training an LLM provides it with the means to create similar content. Does a prospective author reading around his topic of interest not do the same? The only real difference is the speed at which derived works can be created
An author taking someone else’s work, changing parts of it, and then selling it as new is plagiarism. The only difference is the speed at which the LLM can plagiarize its sources.
But didn’t you know that if I write in iambic pentameter and make up some flowery words to describe things that don’t already have a definition, I am literally Shakespeare?
It depends on the level of influence in the new content. There were a number of articles that clearly showed derivative (even stolen) work by using innocuous phrases. For instance, any prompt with “video game” and “plumber” will create an unmistakable clone of Mario.
I suspect the question will be whether these models could produce similar content without using the original content.
So the real debate is whether training a model counts as copyright infringement. Training an LLM provides it with the means to create similar content. Does a prospective author reading around his topic of interest not do the same? The only real difference is the speed at which derived works can be created
An author taking someone else’s work, changing parts of it, and then selling it as new is plagiarism. The only difference is the speed at which the LLM can plagiarize its sources.
Generating new content influenced by someone else is not plagiarism
But didn’t you know that if I write in iambic pentameter and make up some flowery words to describe things that don’t already have a definition, I am literally Shakespeare?
It depends on the level of influence in the new content. There were a number of articles that clearly showed derivative (even stolen) work by using innocuous phrases. For instance, any prompt with “video game” and “plumber” will create an unmistakable clone of Mario.
I suspect the question will be whether these models could produce similar content without using the original content.