My last post tested the latest Stable Diffusion XL against Stable Diffusion 2.1. I concluded that, from a technical standpoint, SDXL represents a huge improvement.
However, I haven’t seen any indication that SDXL is an improvement from a legal or ethical standpoint. For example, I don’t see any indication that SDXL was created with more attention to copyright than SD21 was. (The image above was generated using SDXL via the prompt, “evil smiling AI, grinning mischievously while creating images in violation of ethics and copyright.” I used SDXL to generate the other illustrations in this article, as well.)
Which means that we still have a lot to discuss on these matters.
NOBODY knows what will happen next as far as legal challenges. But all of us have guesses on the subject, including those of us who aren’t even lawyers, so here’s mine.
My first claim: An AI that can generate stuff (art, text, code, music, etc) is a fundamentally different thing than the stuff itself. Images are data. SDXL is a model that summarizes the data. The algorithm manipulates data, using the model. Models, algorithms, and data are different things. That’s why they are different words.
Therefore, I anticipate that the American circuit and Supreme courts will generally rule that creating an AI is transformative. As a result, they will rule that creating an AI is fair use and not illegal under copyright law. And the suit by artists against AI creators will fail or settle. Other countries have different laws, which will definitely lead to varying results.
I still see 2 big legal issues.
First, even if using data for training doesn’t break copyright law, how the AI creators obtained that data might violate civil law. For example, Microsoft trained an AI using a dataset that included some code whose license explicitly required attribution. Is this a license violation? Seems so to me. Do we suddenly not care about licenses or terms and conditions? If a website specifically says, “Don’t scrape me, bro,” does this constrain use even when copyright law does not?
The upshot of this point is that I think AI creators will have to pay for rights to scrape from other platforms such as ArtStation, which will increase the cost of using AI but not prevent it. The reason for my saying this is that platforms have the money to fight the AI-creators on this one, and they will be fighting for their life.
Nonetheless, I seriously doubt that the likes of ArtStation will pass along much of that new revenue to artists. I predict this because, well, has any online platform ever substantially compensated content creators?
Second, even if it’s totally legal to create an AI, that doesn’t automatically mean it’s legal to use an AI. Intellectual property law prevents me from using a pencil and paper to depict Yoda, then trying to sell that image online. That’s very clear and well-established. So why should I be able to have an AI make a Yoda image, then sell it as part of a game about weird creatures? The law allows us to do a lot of stuff as long as doing so doesn’t directly generate profits (e.g., illustrating free blogs, I’m thinking). But there are plenty of cases where somebody took umbrage that somebody else commercialized a specific piece of intellectual property, particularly if not for commentary or satire.
The gray area has to do with emulating (“stealing”?) an artist’s style. You’d think this is new with AI, but it’s not. To me, the really interesting case to watch here is Roberts’ suit against Edwards for stealing her style. No AI is involved at all. It’s just one artist apparently (?) emulating another artist’s style in a way that hurt her business.
I predict that Roberts will lose, though, as long as Edwards isn’t claiming his work was by Roberts. Why? Because artists emulate one another a lot. Maybe she’ll be able to settle for some compensation, though, due to the economic harm claims. I feel sad for her about how Edwards is clearly cutting in on her turf. Yet, feelings aren’t the same thing as precedent and law.
Bottom line: I think the only suits likely to prevail are those by the big IP powerhouses and, with limitations, by the art platforms. I think the legal decisions will limit use of AI-generated art that commercializes existing IP without permission, and the decisions will increase the cost of using AI tools. Artists will probably not profit much from these decisions. Sad for the artists, actually. I think the artists are going to get hung out to dry.
Having driven off my audience, it’s time to turn to ethics. Everybody comes at this from a different standpoint. Mine is Christian and, in the context of AI art, amounts to, “Don’t steal stuff.” Let me try reasoning through the 3 legal issues above from this ethical standpoint.
Is it ethical to train an AI using data in a way incompatible with the license, terms or conditions? I don’t see how this could be ok, ethically speaking. I mean, if I put a pie on my windowsill with a note, “This is for the homeless,” and you lick it, then I’m going to be annoyed. By licking it, you’ve sullied the pie, and thus stolen an aspect of its quality from me as well as the intended consumer.
Is it ethical to sell AI-generated images about specific IP? Money went into creating Yoda, who is as specific of a thing as the concept of a claw hammer or a PNP transistor. I see no reason why it wouldn’t be as wrong to commercialize specific patented ideas (or specific copyrighted representations) as it is to steal specific physical things. In fact, making Yoda sketches even just for my own non-commercial enjoyment feels a little bit like stealing. It’s too specific. If I want a Yoda for more than disposable purposes, maybe I should buy a Yoda. This is a grey area, and I could see how other people might easily disagree on the issue of copying IP for durable personal use, even if we agree that it’s wrong to sell unauthorized copies of IP.
Is it ethical to sell AI-generated images that emulate an artist’s style? Just because artists study each other, and even emulate each other, doesn’t make it right. If I intentionally emulate somebody to take away their market, then I’m taking food off their table. In my own pursuit to be like others (Andrew Bosley, the Miko, and many others), I intend to be like them, not exactly the same as them. And I certainly have no plans of cultivating their style, then trying to take business from them; such a behavior would fill me with regret. I feel bad for Roberts, above, largely because of this point. The other person’s work is so similar to hers.
That said, I wouldn’t feel bad for Roberts if Edwards were doing this unintentionally. If I accidentally take something from a store without paying, then discover it in the car, I haven’t committed an ethical fault unless I fail to go back and pay for it (or return it). Now that Edwards knows he’s cutting in on Roberts, the least he could do (if it was unintentional) would be to say, “Oh, yeah, I can see that this is an issue. I’ll make a conscious effort in the future to be more distinctive and, frankly, innovative.” (Anyway, that’s what I’d do. Edwards is responsible for running his own life.)
And I wouldn’t feel bad for Roberts if Edwards’ work looked less specifically like the work of Roberts. They both make collages of black kids. At that level, the comparison is pretty vague and general. But his specific works have very specific similarities to her specific works in terms of composition, content, rendering, and other elements. On that basis, I have some hope that they’ll settle for cash.
Bottom line: Selling AI-generated art would trouble me ethically if it intentionally emulated specific existing IP or emulated styles in a competitive way — which is what some people seem to be doing, from what I gather. Note that a lot of this comes down to the prompts and the use of the generated art. Selling AI generated with “Yoda in a bathing suit” or “Sunset in the style of Andrew Bosley” would bother me somewhat, and it would bother me even more if I then started hawking that art to the customers of Andrew. But selling games containing generated with any of my SDXL test prompts wouldn’t bother my conscience: I didn’t create it intentionally to emulate specific existing IP or specific artist styles. And I have no intent to sell it to publishers in ways that would compete with other artists.
Room for discussing other dimensions and perspectives
Another ethical issue, which I never touched on above, is that we’ve got to leave room for people to benefit from others’ work, including through the use of tools. Why? Because that’s how civilization works. You can’t just say nobody has a right to get a leg up and join your profession or industry. That’s exclusionary. And it flies in the face of 5000+ years of civilization. There are reasons why the law places limits on ownership of ideas and representations of ideas. Building on one another is a common good.
Thus, you have to draw a line where some things are in and others are out. I’ve tried to draw a respectful line above. I might change my mind on some aspects of it, if anybody wants to have a reasoned discussion.
A final word
I asked Bing’s AI chatbot for its “opinion” on these issues. After all, the chatbot was trained on a lot of data without the permission of the data creators, quite possibly in violation of terms of service. I asked, “is it ok to use content from a website in violation of the terms and conditions, even if doing so doesn’t violate copyright law? it seems like this is what microsoft did when training an ai from code on github.”
The short answer to this question is “no,” unless you’ve obtained the author’s permission. In fact, virtually all digital content enjoys the same copyright protections as non-digital, “offline” content.
Regarding your question about Microsoft training an AI from code on GitHub, it seems that GitHub programmers have filed a class-action lawsuit against GitHub, its parent Microsoft, and its technology partner, OpenAI, for allegedly violating their open-source licenses and using their code to train Microsoft’s latest AI tool, called Copilot1. However, Microsoft Corp, Microsoft’s GitHub Inc and OpenAI Inc have told a San Francisco federal court that a proposed class-action lawsuit for improperly monetizing open-source code to train their artificial-intelligence systems cannot be sustained2.
I replied, “If it was wrong to train an AI from code on GitHub, then was it also wrong to create the Bing Search AI?” Its reply…
Microsoft Bing search is a product of Microsoft Corporation. It is a search engine that provides users with relevant information based on their search queries. The AI behind Bing Search is designed to provide users with the most accurate and relevant results possible. Microsoft has not been accused of any wrongdoing in relation to the creation of Bing Search AI.
It’s only a matter of time, Bing. Only a matter of time.