There’s a long history of this and you might find some helpful information in looking at “transformative use” of copyrighted materials. Google Books is a famous case where the technology company won the lawsuit.
The real problem is that LLMs constantly spit out copyrighted material verbatim. That’s not transformative. And it’s a near-impossible problem to solve while maintaining the utility. Because these things aren’t actually AI, they’re just monstrous statistical correlation databases generated from an enormous data set.
Much of the utility from them will become targeted applications where the training comes from public/owned datasets. I don’t think the copyright case is going to end well for these companies…or at least they’re going to have to gradually chisel away parts of their training data, which will have an outsized impact as more and more AI generated material finds its way into the training data sets.
There’s more and more research starting to happen on it, but I’ve seen anywhere from 20% to 60% of responses. Here’s a recent study where they explicitly try to coerce LLMs to break copyright: https://www.patronus.ai/blog/introducing-copyright-catcher
I don’t have the time to grab them right now, but in many of the lawsuits brought forward against companies developing LLMs, their openings contain some statistics gathered on how frequently they infringed by returning copyrighted material.
There’s a long history of this and you might find some helpful information in looking at “transformative use” of copyrighted materials. Google Books is a famous case where the technology company won the lawsuit.
The real problem is that LLMs constantly spit out copyrighted material verbatim. That’s not transformative. And it’s a near-impossible problem to solve while maintaining the utility. Because these things aren’t actually AI, they’re just monstrous statistical correlation databases generated from an enormous data set.
Much of the utility from them will become targeted applications where the training comes from public/owned datasets. I don’t think the copyright case is going to end well for these companies…or at least they’re going to have to gradually chisel away parts of their training data, which will have an outsized impact as more and more AI generated material finds its way into the training data sets.
How constantly does it spit out copyrighted material? Is there data on that?
There’s more and more research starting to happen on it, but I’ve seen anywhere from 20% to 60% of responses. Here’s a recent study where they explicitly try to coerce LLMs to break copyright: https://www.patronus.ai/blog/introducing-copyright-catcher
I don’t have the time to grab them right now, but in many of the lawsuits brought forward against companies developing LLMs, their openings contain some statistics gathered on how frequently they infringed by returning copyrighted material.