I thought this was already a settled argument, though. It is not reproducing exact copies of copyrighted work, it is creating simulacra based on data that it has trained on. Obviously it trained on copyrighted material, as all large AI models had, and the newest version has gotten better at being able to create output resembling its training material if the user specifically requests a concept that fits the description (and likely has it regenerate 20+ times until it gives you the “infringing” work you’re looking for).

If Midjourney is breaking copyright, then so is every artist who has ever sketched a scene from a movie.

“I asked the AI to give me an exact copy of an image I selected and was surprised when it got pretty close” eyeroll without listing the prompts this article is a meaningless hitpiece.

Edit: I missed the prompts on the image itself. I will admit the prompts are pretty simple and output pretty similar to the film itself. That said, still not identical, equivalent to fan drawings of the film in my mind as Stovetop pointed out.

It also suggests that Midjourney has been trained on high-resolution copyright images, to which they may or may not have a license.

This will always be a stupid question. Especially when people don’t know how to conjugate “copyrighted.”

Yes, the draw-anything machine was fed with all the images they could find. What else was it going to use?

No, none of it is licensed. And it never will be. And that’s fine. It’s supposed to be trained on basically all images ever, and a film is one hundred thousand images in a row. The more images you shovel in, the less any specific one matters.

Ask for a popular character in a scene like the movie they’re from and of course you might get something vaguely resembling some frame from the movie. There’s a lot! They’re kinda formulaic. The only one here that’s not an egregious hand-wave is the Joker example, and it’s exactly the kind of output these things love to generate for any prompt: centered figure, looking at the camera, blurry background.

Ask a human artist to sketch any of these prompts from memory and you’d get similar results. You wanna say they’re violating copyright law, just by remembering a movie?

You wanna pretend a truly from-scratch AI would be incapable of violating copyright law? Like if it came up with something that already exists, we’d have to shrug and say, welp, guess that AI owns Batman now. Because that’s totally how it works when humans draw bespoke new frames of existing characters. Or brand-new variations on those characters. Or genuinely novel characters that happen to resemble those characters.

Basically - if you expect anything coming out of these image-generators to be immune to copyright and trademark laws, you don’t understand those laws. Ditto if you expect everything coming out to be illegal. But if you think being able to draw Batman on a tricycle means the AI contains a copy of exactly that, you don’t understand these AIs.