This article claims generative AI systems are in a precarious legal position because you can ask them to generate images of copyrighted characters like Spider-Man or Mario & Luigi.

I think this concern isn’t valid. A tool being able to draw Spider-Man because you asked it too isn’t copyright infringement until you do something with that image. Otherwise drawing Spider-Man in Photoshop as a personal sketch would be illegal without a license.

Of course, I’m not a lawyer.
https://garymarcus.substack.com/p/things-are-about-to-get-a-lot-worse

Things are about to get a lot worse for Generative AI

A full of spectrum of infringment

Marcus on AI
@carnage4life Isn't OpenAI generating the picture, and then giving to me (for money, in the case I am a subscriber)?
@carnage4life I think the issue is that generally, people actually do want to do things with those images. The idea that the investors behind AI want to make a return on their investments, but if no one they sell to can make any $ from the generated images, they are screwed.
@carnage4life I'm no lawyer either, but I think copyright covers even just the creation of derivative work. So fan art is technically illegal at creation and usually gets caught when it's distributed. And most of the time the copyright holders don't care/isn't worth it to go after them. But going after OpenAI or Microsoft might be worth it, or prolific enough to want to stop them

@carnage4life yeah, I think you're wrong on a couple counts.

It's not just "draw Mario" that's gonna get them in trouble (though that is problematic).

The obvious ease with which prompts can generate copyrighted work is evidence of what we already know: the image gen llms have been "trained" on copyrighted works that the AI companies do not have a license to use.

In both the chatgpt case and this one the AI acts as a Rube Goldberg copyright violation machine.

@tob There hasn’t been a legal resolution to the question as to whether a license is required to train LLMs on copyrighted works. I expect we’ll get clarity here withon the year given this is the biggest open question facing LLMs.

@carnage4life if I ask a tool to draw Spider-Man, and it does… isn’t that because the tool has been trained on copyrighted images of the character? How can a model reproduce copyrighted work without copyrighted input?

I think this is fundamentally different from a person grabbing a notebook and sketching the character. Mead doesn’t profit from what someone sketches in their notebooks. Meanwhile, whoever maintains the model is collecting revenue based on what it generates.

@carnage4life

Maybe the problem is, that the fact that it can draw copyrighted characters when prompted is an indication for the fact, that it was trained on images of those characters.