We Asked A.I. To Create the Joker. It Generated a Copyrighted Image | Hacker News Reader

We Asked A.I. To Create the Joker. It Generated a Copyrighted Image

nytimes.com

5 pointsjeffwass8 comments

Wouldn't copyright infringement with AI only apply if you try and sell a product or make money with it? I would think OpenAI would just insert some disclaimer making the user responsible for the final use of the image or output, since anyone can make fan art and that, I don't think, is infringing any copyrights. I would see no difference between generating fan art, and actually making it, so it's on the user to understand how they are using the tool.

Nope, copyright strictly depend on whether the infringer makes money from it. By that logic, Neo-Nazis can print up pictures of Batman wearing a swastika and start giving them away for free, and there's nothing Warner Brothers / DC can do.

Right, but technically OpenAi (or Midjourney) isn't the one explicitly generating the copyrighted image. Their tool is, but not based on their own intentions, and it seems that responsibility is on the user who generated the prompt. I don't know much about the copyright laws, but it seems generative AI can be classified as an artistic tool.

Fan art and memes are tolerated despite their infringing nature. It's a laughable double-standard. You only need compare the contents of a typical Fandom wiki vs. Creative Commons repositories to see this in action.

If fan art increases brand recognition and promotes the properties, then it gets a free pass.

Seems like the NYT want to pick a fight even if it's not theirs. Kinda nice, that a publisher can push their own agenda and sell it as "news".

Uhhhhhh... yes, this is exactly what I would expect it to do.

> Uhhhhhh... yes, this is exactly what I would expect it to do.

So you expect MidJourney to not draw Mario when asked "Mario from Mario Bros" while you expect it to draw Mario when asked "Famous plumber from a video game"?

No?

If you ask it to "Create an image of Joaquin Phoenix Joker movie, 2019, screenshot from a movie, movie scene" I would expect it to possibly (and probably) make an image that resembles the copyrighted Warner Brothers image. That's example 1. I fully expect systems which are trained on large corpora to reproduce copyrighted images like this.

If you ask it to generate a "popular movie screencap" I would expect it to generate an image that possible (and probably) resembles a copyrighted image.

As for the mario example, where they didn't explicitly say Mario, just "italian video game character", given that mario has been ubiquitous in our culture for 43 years, and mario is the most common italian video game character, I wouldn't expect it to produce another character (such as luigi, or... are there other italian video game characters?).

The datasets used to train these models almost certainly contain lots of images that closely resemble the one it created. I know that OpenAI says their corpora is "fair use" but I would be absolutely completely surprised if it didn't contain large numbers of images that are extremely close to what it produced.

In the future I would expect that training companies will reduce the amount of copyrighted imagery in their data sets, and also improve their loss functions to avoid images that appear to be copyrighted.

Reply on news.ycombinator.com