On being listed as an artist whose work was used to train Midjourney

So the somewhat necessary followup question: if Midjourney or ChatGPT includes a license in front of its model weights or access to its model that forbids using those weights or output from its model to create a competitor, is it fair use to ignore that license?

In both cases, we're relying on copyright. The companies are saying that you can access the models under a license. It's not too hard to circumvent that license or get access through a third party or separate service: but doing so would obviously be seen as deliberate circumvention. We can easily compare that to an artist putting up a click-through in front of a gallery that says that by clicking 'agree' you agree not to use their work for training. And in fact, it's literally the same restriction in both cases: OpenAI maintains that they can use copyright to enforce that people accessing their model avoid using their model to train AIs. Paradoxically, they also maintain that artists can not use copyright to block people accessing their images from using those images to train AIs.

Facebook has released model weights with licenses that restrict how those model weights are used. But I can get those models without clicking 'agree' on that license agreement. They're mirrored all over the place. We'll see what courts say, but Facebook's public argument is that their license doesn't stop applying if I download their software from a mirror. So I find it hard to believe that Facebook honestly believes in a fair use argument if they also claim that the same fair use argument doesn't apply to their own copyrighted material that they've released online (assuming that model weights can even be copyrighted in the first place).

This is one of my biggest issues with the fair use argument -- it's not that there's nothing convincing about it, in a vacuum I'd very likely agree with it. I don't think that licenses should allow completely arbitrary restrictions. But I also can't ignore that the vast majority of companies making this argument are making it extremely inconsistently. They know that their business models literally don't work if competitors figure out ways to use their services to rapidly produce competing models. If another company comes up with a training method to rapidly replicate model weights or to augment existing models using their output, none of these companies will be OK with that.

None of these companies believe that fair use invalidates license terms when it comes to their own IP.

People do whatever they want with their own property. You have no right to steal it just because they want to monetise it. What’s killing art is stealing it en masse using procedural generators.

>What would you buy? $10 H&M or $100 hand-made shirt? - (My guess, if you could afford the later.)

This is an interesting example because even in the $100 case you are still talking about machine-augmentation. You can have a seamstress or a tailor customize patterns, using off the shelf textiles, for that order of magnitude price - but if you want to use custom built, exotic materials or many kinds combined, the cost is on the orders of thousands not hundreds. Also there is a large industry of just printing designs on stock-shirts, that has a different point effort-scale equilibria.

Thinking about how how automation disintermediates is very important. For animation, often productions have key-frame artists in the animation pipeline that define scenes, and then others that take though to flush out all the details of that scene. GenAI can potentially automate that process. You could still have the artist producing a keyframe, and can render that into a video.

Another big factor is style. One hypothesized reason that more impressionism, absurdism or abstract art all become styles is photography. Once cheap machine-produced photography became available, there is less need for a portrait artist. But further, it also is no longer high-status and others push trends alternative directions.

All the experiments and innovation going right now will definitely settle into a different set of roles for artists, and trends that they will seek to satisfy. Art-style itself will change as a result of both what is technically possible and also what is _not_ easily automatable in order to gain prestige.

I think there is a case to be made when AI models do produce copies. For instance, I think the NYT have a right to have an issue with the near verbatim recall of NYT articles. It's not clear cut though, when these models produce copies, they are not functioning as intended. Legally that might produce a quagmire, is it fair use when you intend to be transformative but by accident it isn't? Does it matter if you have no control over which bits are not transformative? Does it matter if you know in advance that some bits will be non transformative but you don't know which ones.

I presume there are people working on research relating to how to prevent output of raw training data, what is the state of the art in this area? Would it be sufficient to prevent output of the training data or should the models be required to have no significant internal copies of training examples?

I feel like the issue here, is you are giving AIs agency.

AIs are not magic. They are tools. They are not alive, they do not have agency. They do not do things by themselves. Humans do things, some humans use AI to do those things. Agency always rests with a combination of the tool's creator and operator, never the tool itself.

Is there really a difference between a human flooding the market using AI and a human flooding the market using a printing press?

Even if human's can't compete (An obviously untrue premise from my perspective, but lets assume it for the sake of argument), is that a bad thing? The human endeavor is not meant to be a make work project. Humans should not be forced to pointlessly toil out of protectionism when they could be turning their attention to something that can't be automated.