3 ms·
> My point is that producing an AI image “in the style of a photograph” is less of an infringement on artists’ rights than producing an AI image “in the style o
by dahart 3y ago
> My point is that producing an AI image “in the style of a photograph” is less of an infringement on artists’ rights than producing an AI image “in the style of Pablo Picasso”. Surely you can agree with me on that point?
It depends entirely on what image was produced. The prompt isn’t the infringement, the output is. And photographs are no less copyrightable than any other art. So I can’t agree with your statement as written because the statement doesn’t make sense and is not true. If an AI’s output happens to look like one or two specific photos, then it violates copyright, regardless of what the prompt was. (The photo doesn’t even have to be in the training set!)
You might have gotten stuck or confused on the idea of whether or not users actually prompt for a specific artist or art style. I stopped talking about a prompt to copy a named artist several comments ago and started pointing out that AI inference can break copyright without a prompt to break copyright.
So, If you’re trying to give a hypothetical example of AI producing an image that looks like a photograph, but not one that looks like any existing photograph, then I can agree that this inference doesn’t violate copyright. But if the output does end up looking like an existing photograph for any reason, then it does violate copyright. This doesn’t depend on whether we’re talking photography or drawings, the standard is the same regardless of the medium. The balance of photos in the training set might change the likelihood of accidentally reproducing one of them, and that’s an interesting question, but not what we’ve been discussing.
> I'm not splitting hairs on whether an image is imitating "style of <artist1>" or imitating "style of <artist1> and <artist2>". A distinction like that is not meaningful. If you imitate artist1 and artist2 in the same image, that's still imitation of specific artist styles.
Good. You’re getting closer to what I’ve been trying to say. AI always imitates the style of some number of specific artists, because it is always trained on some number of specific artists, and it usually inferences from fewer than all the artists in the training set. The number might be high for any given run, and a particular style might not always be identifiable, but the design of the system is, in fact, to copy from multiple artists at all times. I agree with you that the distinction between one and many is not that meaningful, however in the courts there is precedent for declaring a copy of one artist to be a violation of the law while letting a remix of several artists off the hook. From the copyright perspective, AI inference might not violate copyrights as long as there is no identifiable infringement of a singular work.
Whether or not training on copyrighted materials in the first place infringes has not yet been litigated, but since humans get threatened and sued for it all the time, there’s reason to believe AI could have legal trouble in it’s future unless they manage to get ahead of it and change the law, which is exactly what OpenAI recognized and started trying to do.
Hopefully this is obvious to you, but following people on Civitai doesn’t prove anything about how often people try to prompt for specific artists, and given the growing exposure to copyright issues surrounding this, it’s entirely possible that the people using prompts that intend to break copyright are choosing not posting their output. You have absolutely no idea how many people are prompting for specific situations or styles or names. And again, the “majority” usage is completely irrelevant to this discussion.
- baobabKoodaa 3y ago> the “majority” usage is completely irrelevant to this discussion. You've shifted the goalposts so much that I genuinely don't know what you're arguing for anymore. Are you saying "some people use <technology> for <bad>, therefore <technology> is <bad>"? Because then we're back to banning knives, banning cars, banning internet, banning everything basically. Typically in these discussions it matters what the intended or majority use case for a technology is. If it doesn't matter, then what is it that you're arguing for exactly? I fully agree with you that some people use image generators to generate new images in specific artists' styles and I agree that there's some moral issues there.
- dahart 3y agoI’ve been suspecting for the entire thread that you’re really not understanding nor listening to my argument, and that might indeed make it seem like the goalposts are moving. I know you said you think you understand, but my points haven’t changed. Banning things is a straw man on your part and not something I ever discussed in this thread. > it matters what the intended or majority use case for a technology is. Okay, speaking of moving goal posts, you previously made unverifiable claims about what the majority type of prompt is, in order to contradict my points about specific art prompts being common. Until now, our discussion was not about the majority use case for the technology. The prompts and the tech are not the same things. The majority use case of AI technology has nothing to do with whether a user is seeking generic photographs or art in a specific style. The intended use case of Stable Diffusion, for example, is to generate images from a text description. All text prompts and all image outputs fit the stated majority use case of Stable Diffusion. The intent of it has - deliberately - never included any statements on what the prompt contains nor what the output looks like, and half of the examples on Stable Diffusion’s home page are art anyway. You claim accidental infringement is an unintended failure mode, but it’s a surprise to nobody that a sophisticated copy machine sometimes copies things. Plus they had to copy stuff in order to train it, which may well be illegal even before anyone starts inferencing. It’s fundamentally a copy machine they built, and they intended for it to copy things, so unless they design it to guarantee remixes of many, hoping it won’t copy enough of any single source to violate copyright is wishful thinking and nothing more. The copyright office may decide to view that as willful ignorance in violation of the law. Accidents and intentional infringements both are bound to happen, and the machine is effectively distributing copyrighted work and could be violating existing laws. So far you’ve ignored all of my points touching on the possibility that training AI on copyrighted material might be illegal. There might be legal issues with AI, unless the people building them curate legal training sets. You’ve also failed to engage with any points about OpenAI requesting copyright exemption, which I’ve mentioned multiple times and was my first point. Their actions have already demonstrated what I’m saying here, that they know the potential copyright problems with their AI go far beyond overfitting accidents. We haven’t even discussed ChatGPT, but it’s pretty obvious that they would like to have the right to disseminate information and excerpts from copyrighted sources. They want it to overfit on some things, otherwise they can’t improve it’s ability to tell the truth and not hallicinate false information.