Is Generative AI Original Art?

I’ve found one of the best explanations of generative AI from Liberty RPF. He was recently on the Infinite Loops podcast and laid out the case for why, not only is generative AI original work, but isn’t appropriating other artists' work the way many people think it is.

The typical artist view is that the AI model is being trained on their work and they should be compensated. Liberty notes that if you’re generating a Disney character and trying to sell it, then of course you should be stopped. However, he’s uncomfortable with the idea of restricting machine learning because it’s something we don’t do for humans. We don't tell humans you can’t look at any painting that’s copyrighted, and that they must only learn from non copyrighted works.

Liberty points out that human artists learn from what comes before them. Their creations are a remix of everything that’s in their heads. Most human artists start by copying, or at least imitating the style of great works. It’s the way they learn.

AI learns the same way. It’s being trained on billions of images and learns the parameters, the shapes, the edges, the colors, and the textures. What most people don’t understand is that what the inference model generates at the end doesn’t contain ANY portion of those images it trained on. And if an artist gripes someone took their image and put it out there, Liberty notes, their image isn’t even in the output.

AI is using what it learned from those patterns to generate something original; much like the painter who’s learned the patterns and style of only three or four different masters, but AI is learning from a billion or more images!

Another thing to consider, according to Liberty, is that AI will learn more and more from images generated by AI than from humans. Whether that’s good or bad remains to be seen.

---