The comment section is going to be filled with people saying that this is fair use, and I'm sympathetic to that argument (even though I do feel like we should be mature enough thinkers to have the capability to tell the difference between the morality of building on top of a commons and plundering it). I'm very openly not a copyright maximalist; I think there's real danger in equating training to copyright infringement, and I think it's an argument that might come back to bite artists in the future -- I do worry that the current lawsuits are misguided.
However, in every single conversation about fair use, two followup questions need to be asked:
1. If you're making the argument that AI's "learning" from existing pieces is just like a human learning from existing pieces, then you're indirectly comparing that AI to a human being. So let's be consistent about that analogy: commissioning or working with a human artist to produce a piece of art does not grant you copyright over that art.
To people who are saying that AIs training are just like humans learning from art, do you support the idea that AI-generated works can not be copyrighted? Because that's the consistent view of AI as analogues for human training: commissioning humans doesn't grant you copyright.
2. If we're viewing AI training as fair use, then is it fair use to circumvent the licenses on these models in order to train other models on them? The majority of these commercial tools have explicit clauses in their terms of service to prevent their use when training other AIs. Is it fair use to circumvent that TOS?
If the TOS is in a different category that somehow magically gets rid of the fair use question, then if an artist puts up a TOS in front of viewing a piece of artwork that says it may not be used to train AIs, does training on that art stop being fair use?
It doesn't make sense that one shrinkwrap license would be OK and one wouldn't be. If Midjourney has a legal case that circumventing its TOS and gaining access isn't fair use, then artists should be able to put up a click-through in front of images that includes an agreement not to train and that should be legally enforceable.
If artists can't do that, if artists can't legally enforce terms of use in front of artwork about how that artwork can be used, then commercial companies building these models shouldn't be able to either.
----
It's not that the fair use argument doesn't hold any water, it's that people tend to be very selective about how they want that argument applied. And I think that's revealing. Taking a step back from those questions:
It's somewhat frustrating to have debates on this because the debates are very abstract and are disconnected from the completely bad-faith way in which these arguments are often proposed. You're forced to debate in a very dry way, and what you really want to do is step back and say, "none of these companies actually believe in open culture, all of them are copyright maximalists when it's convenient, none of them care about accessibility of art tools, they just want to plunder the culture and insert themselves as middlepeople between art and the people who make art."
We all know this is a grift. We know what these companies actually believe, but we have to pretend we're having some kind of emotionless conversation about "Open" models and to look at their uses as if the primary intended audience is people using image generators to make printouts for their local D&D campaigns.
But come on, that's not the reality that's ever going to happen.
That kind of indie, low-cost usage and Open access can't exist in the long-term if these models are intended to be the commercial powerhouses that companies want. Monetization of their output and access necessarily requires restriction of their output and access. The truth that we all kind of know deep down is that these models are about inserting a monetary layer between human beings and art, a layer that is owned by private corporations and is deliberately used (regardless of its quality or suitability) to devalue labor and to poison the resources that these models were built on in order to reduce competition and to give gatekeepers more leverage over their workforces and their markets. It's a tragedy of the commons being played out in real time: the creation of a service that just like every other privatized platform will become crappier and worse over time as more control over model use is taken away and as prices increase and as restrictions over usage are increasingly layered on top normal individuals who will increasingly need to ask permission in order to create. At the same time, businesses hope that model usage will shrink and dilute the fields that are necessary to sustain those models, fields that are now treated as competitors, and who's destruction is desirable because the elimination of that resource makes it harder to replicate or compete with established AI models.
As is very often the case, the goal is to get up a ladder and then pull the ladder up.
That doesn't mean that this isn't fair use (although as someone who advocates very strongly for expansion of fair use and heavy reduction of copyright, I will say that fair use and morality are two entirely different subjects). But we can advocate for fair use without pretending that the companies and grifters making these arguments actually believe them. Actual ubiquitous open access to AI art tools would be toxic to a company like Midjourney; Midjourney's entire business model relies on them constructing an artificial moat in front of the production of art that forces creators to give them money. And the existence of that moat relies on Midjourney being able to restrict what people can do with their model, it can not exist without strong legal protections around copyright and licensing agreements. If Midjourney actually believed in free culture and free use, they wouldn't be launching a private model behind Discord with a bunch of ridiculous license terms telling people what does and doesn't count as fair use when using Midjourney. They wouldn't be pretending that model outputs are covered by copyright. As far as I'm concerned, every single "license" in front of model weights or model output is an admission that the companies behind those models don't actually believe in the fair use argument. They're just saying it because it's convenient.
I would love to have a purely intellectual debate about AI as a genuinely open tool that can fit into artists' belts alongside other tools. I would love to ignore the hype cycle and the spammers and the grifters who all couldn't care less about making creation accessible or expanding the scope of what artists create. But unfortunately, I can't; there's a real world that exists alongside that hypothetical debate, and we are allowed to look at that real world for context. And I can be sympathetic to the fair use argument while still recognizing the utter hypocrisy behind how companies actually use that argument, and while still recognizing that the direction that large companies are pushing AI and the way they want to use AI is toxic to real artists.
----
That doesn't mean I'm panicking about it any more than I'm panicking about GPT trying to program. I understand the limitations of these tools as well, and a little bit of perspective goes a long way. But I still understand why artists are upset. I don't always agree with the specific arguments but it's not just them being sore losers or something, they have a legitimate grievance with the AI market even if they don't always perfectly know how to put it into words.
And in a lot of ways, this doesn't really have anything to do with the quality of the output or about what's "better" or "fair" anyway; that's not really what the debate is about. Midjourney would be perfectly happy to live in a world where all art gets kind of crappy and repetitive as long as everybody uses Midjourney to produce it. They don't care, they just want you to pay them every time you generate a picture. So the social dynamics around the market are very different from the technical dynamics or even the philosophical arguments that people make.