From the course: The State of AI and Copyright

What's the legal difference between a human creating derivative work and an AI creating derivative work?

From the course: The State of AI and Copyright

What's the legal difference between a human creating derivative work and an AI creating derivative work?

- You'll often hear musicians say, oh, you know I totally ripped off that bridge from this artist here. Not to say that they've ripped it up note for note but they're saying they were influenced and, you know, they they liked that content, they were influenced and that's how they ended up writing their version of that song which is kind of what certain AI models are doing. They're taking something and making it into something else, but being influenced. So in that way, it seems like these AI models are being human. But again, on a, on a larger, faster scale - I, I would say legally there's a difference until someone brings a case that storing something in your brain is a copy (laughing) because they actually work on silicon and the, like, that you are in fact making a copy for the purposes of copyright law. And so right or wrong, philosophically potentially there are infringements by doing exactly the same thing as you would've done in your head. I would however, say that there is a difference of scale. So whereas I suspect although our brains are amazing things and and store vast amounts of things, you tend to be being influenced by a relatively small number of let's say copyright works at any particular time. Whereas a large language model or an image generator is literally thinking across everything in its training set every time, or or the, or what's extracted from them. And it's guided by the prompt, obviously. I think all things being equal with an arbitrary prompt it is likely that the influence of any particular copyright work is extraordinarily small. That said, I think you do get what you ask for. So if you prompt and mention specific artists or specific works, particularly if you prompt to mention a specific brand or, or a a character from a movie, you are, I think quite likely to end up with an infringing work. - You know, Matt, we know so little about some of these L Lms and I've seen concepts like memorization but they can memorize specific examples that you can point out. It's such a fascinating issue. I fully support Matt's statement that the difference between a musician, a human musician listening to a song and playing it themselves there's no copy being made of that song. Whereas with computer hardware and the the techniques we have to use, the the issue is for making a copy and maybe restoring them. And I think maybe the technological question of the day is the model itself is that derivative work of all of the training data on which it was trained on. And we could think of all sorts of analogies, probably none of which are fully technologically correct. But, you know, one, one that I heard is, is, you know it's one of these models just a compressed version of all the training data that it's seen. And maybe there are ways to get that compressed training data out of it. And, you know, unfortunately I think there are a lot of maybe unknowns or, or answers that are difficult to to get out of these systems which might be why these issues are so difficult to answer. - Yeah. - [Jenny Maisel] Under kind of laws. - I think about stuff like what if a musician is using, say a rhyming dictionary is that different than asking, you know, an AI to say gimme 20 rhymes for this word, or, you know make sure this many syllables or something like that then, I know that's not a legal question. More of a philosophical question. - That's a great question though because one thing I do want to point out under copyright law there's the idea expression dichotomy. So copyright law doesn't protect everything. It doesn't protect facts, it doesn't protect, you know broad narrative themes and things like that. It doesn't protect genres. It protects a very specific copy of a work. And when you think about it a lot of things fall outside of that protection from the work itself. So thinking about rhyming are we just extracting facts about what words rhyme With other words? And is that kind of factual matter that we're extracting from text or are we extracting more of the expression of, oh, these two words sound really cool when they're combined together in this live, in this film. You know, there, there is a line there to be drawn.

Contents