If training AI models on these images is fair game, then artists publishing images that are adversarial to the model is also fair game.

Regardless of your stance on the AI art/IP debate, it seems enormously hypocritical for AI builders to say that they not only have the moral right to use images against the artists’ wishes, but that the artists also have a moral obligation to provide these AI builders with clear, denoised versions of their work.

Edit: if you use a tool to thwart internet data collection systems, thus making your data less useful for advertisers, are the data collectors justified in bitterly calling you a “malicious user”?

I'm not saying that artists shouldn't be allowed to perform this attack. I'm just asking how it should be defended against. E.g., in football, asking "how do I defend my team's quarterback from a blitz?" is not the same as saying "I think the rules of football should be changed to prohibit blitzing".

It is the word “malicious” that added venom to your first question.

If I asked how artists can thwart “malicious” tools like Midjourney, you could reasonably assume that I disapprove of Midjourney and think they ought to stop doing what they do.

That is the way artists often talk about tools like Midjourney. I chose that word exactly to turn the tables and call attention to that implicit venom on their part.

What are best ways for burglars to defend against malicious tools like alarms, locks, and security patrols?

What are the best ways for crackers to defend against malicious tools like copy protection?

Burglary is bad whether a human or some kind of a robot does it. There's nothing wrong with an art school taking students to a museum to look at established artists' work. Isn't training AI on existing art the equivalent of that? If so, why does it being AI instead of human make it bad? If not, what's the difference?

No they are not the same. An artist taking time and effort to study and learn a new art style is not the same as an company writing a program that gobbles up every image on the internet. There is a fundamental difference between when a human does it and a machine does it. I feel like this is obvious and yet people keep making this argument for some reason.

> I feel like this is obvious and yet people keep making this argument for some reason.

It's because scores of people coming from non-art domains are suddenly being exposed to the art domain without much real knowledge of working in that domain, find the diffusion tech convincing enough that it entrances them, and makes them believe that their original notions of art are the most important thing that matters.

This debate has made me think, if there's some inability to automate a process, only the people that have spent months/years working in those domains will understand certain things about that process. The rest will be operating on only the beliefs they form from limited exposure. For example, I imagine for a lot of people, "art" is something that magically appears in their Twitter timeline within 5 seconds of opening the app, and years of opening Twitter and seeing new art almost instantly would probably condition some to believe that's all art really is, a thing that's plentiful and can be found near-instantly. So it's not controversial to them when they come across a piece of software that can give them thousands of new pictures near-instantly.

But obviously, prior to around October of 2022, there was nearly always a human behind the art on those timelines. To a human artist, their own work is probably not as plentiful and takes way longer to produce than to consume. Those consumers never needed to interact with the artist or understand the process to appreciate the art privately. Up to now that was usually not a problem.

Maybe that idea will undergo changes in the following years, that the tech is so advanced that a level of respect for the specific process of human artists is necessary for cordial communication of art to succeed. Otherwise it will be disparaged with the equivalent of "opinion discarded" from the nature of the thing.

It makes me wonder, what are the implications of this for domains that haven't been automated to the extent of SD yet? SD has proven that many, many people wanted this level of "art" automation from the beginning. The singular event that made this fact clear was the release of SD. There could be millions of people out there with their own preconceived notions, correct or not, just waiting for the ability to cut out the middleman in their own domains. It's like a powder keg waiting for the tiniest spark of technological progress to be ignited.

For the "respect of artists" angle, I'm reminded of this article[1] about a modern art installation from 12 years ago where people's art contest entries were trampled upon by visitors. People were upset by this.

As Marshall McLuhan once stated, "art is anything you can get away with."

[1] https://en.wikipedia.org/wiki/Pixiv#Chaos_Lounge

It's not obvious. Like the parent asked, what's the difference when an AI does it versus when a human does? If you're a reductivist like me, you'd see that humans are simply biological computers who are doing the same as an AI. If you think humans are somehow special, then of course you'd think there is some magic in learning a new style of art, magic that an AI doesn't have.

You’re saying it’s not obvious that AI and humans are different, but that’s the default assumption. What’s the affirmative argument for how we’re the same?

Physicalism, or more accurately that we're both made of matter and that the structure of brains gives rise to intelligent behavior in such a way as to be analogous to computers. Note that I did not say that computers or neural networks are like the brain's neurons, merely that matter gives rise to computational abilities when organized in particular ways. This is the physicalist approach in philosophy, that there is nothing beyond the physical. If this weren't true, then we wouldn't be able to identify parts of the brain that do certain things, and that damage to the brain in those areas wouldn't affect the ability to do said things.

Of course, if you are a mind-body dualist or support idealism, you will think that humans are somehow different than other arrangements of matter that give rise to computation.

To be clear, I’m not asking for the argument that it’s possible for a computer to be somehow “the same” as humans. I’m asking for the argument that AI as it presently exists is “the same”. Because you’re not making this argument about dogs or parrots or Markov chains or Microsoft Clippy, but you are making it about generative neural networks.

Humans and machines are just different.

As someone who does creative work (writing and animation mostly), posts it publicly for free on the Internet, and tries to actively help people learn (by, e.g, answering questions when somebody asked how I did something or what my thought process was), I'm very much in favor of humans learning from something I made, and I'm also very much against a model like ChatGPT or Stable Diffusion learning from something I made. Let me explain my stance:

There's a lot of reasons to help humans that don't exist for machines. The biggest one (and I suspect this to be the case for a lot of artists) is that anybody that's trying to learn an artistic endeavor is automatically somebody that I relate to. Learning to do art well is hard. I think any good artist is in a constant state of struggling to improve, so they're very going to be very much cognizant of how it feels to struggle to improve, and it's highly likely (or at least my perception is that it's highly likely) that anybody seriously studying somebody else's art and trying to learn from it is experiencing something similar, and has similar goals and motivations. In that situation, it's only human to want to help the other person out.

That doesn't apply to a machine. Stable Diffusion is not living with three roommates with no health insurance and only one pair of pants with no holes in them, earning rent and food off spotty freelance work for random people on Twitter. ChatGPT isn't burning the candle at both ends working a day job and spending long nights trying to piece together a short story good enough to get published in an open call fiction magazine with a 1.1% acceptance rate that pays ten cents a word. There is no toiling human making sacrifices because they were born with an irrepressible urge to create; instead, it's arguable that the primary reason these things exist is to provide some return on investment for the huge companies that own them. Microsoft's dumped billions of dollars into OpenAI. It must be expecting to make that money back somehow, so helping ChatGPT out feels like I'm just making some faceless suit even richer than they already are, which is sort of the exact opposite of helping out a fellow creator.

(There's an argument to be made that things like ChatGPT are providing a net positive to society, so I should feel happy to help train them on that count, but in their current state I don't think they are. That's a separate ramble though, and a tangent I don't really feel like getting into while I'm explaining my stance here).

The outcome I expect from helping a human is also different from the outcome I expect from helping a machine. Humans are different from machines. For a human to copy somebody's style, I think they have to fully understand the process that the other human took to arrive at that result. To do that takes solid fundamentals, and to develop those fundamentals you have to spend a lot of time practicing, which means exposing yourself to new influences. I really do think it's pretty much impossible for somebody to completely copy somebody else's style. When you create something, it's a deeply personal experience. What resonates with you? What image, or theme, or emotion, or experience, spoke to you on such a deep level that it drove you to spend forty minutes trying to animate a foot so it slaps against the ground in just the right way? Is it really likely that any other individual is going to see what you made, and get out exactly what you put into it, without putting any of their own spin on it? In my experience, it doesn't happen at all. So when somebody copies you (as long as the copy was effortful), it's a joyous experience to see what they made. It's interesting and illuminating and fun to see an earnest expression of how somebody else sees your art, to the point where people do stuff like this: https://static.wikia.nocookie.net/85799efa-2901-4cf7-a05d-a5... (I don't have any special connection to WoF, it just happened to pop up when I googled 'draw other artist style meme')

OK, but somebody with solid enough fundamentals could get /close enough/ that it doesn't matter, and then sell to less discerning members of your audience, right?

Well, I guess in-principle that's possible, but it just really doesn't happen. Anybody that had the drive to get to the point where they could copy somebody else and have it look good probably already has their own style they're proud of. I don't know a single artist that would be happy with a client coming to them and saying "hey can you draw me something and make it look exactly like <other artist> drew it?". In the worst case they'd tell you to fuck off, in the best case they'd begrudgingly accept your money but probably be insulted and not put as much effort as they would be working on something personal and exciting for them. So another person learning from you isn't infringing on your audience and livelihood. Or if it does, the impact is so small that the positive feeling of helping someone similar to you out outweighs that by quite a bit.

Machines aren't like that. Machines will gladly slurp up your entire body of work, churn through it in training, and then shamelessly crap out a limitless supply of stuff that's rehashes of things you made for somebody that wrote 'by <your name>' in the prompt. You don't even get the warm fuzzies of knowing you helped somebody out, or the cool experience of seeing somebody else's interpretation of your art. It's just an uncanny Frankenstein of your work that's a sordid reminder that a good portion of people out there don't see artists as people, don't engage with art the way you do, and don't want to respect your wishes that the stuff you make not contribute to a cash-and-power shift that's decidedly not in your favor.

I cannot upvote this enough.

I see what you’re doing but the parallel doesn’t really work. This isn’t an equal struggle between AI builders and artists.

What happened is that AI builders took artists’ images, baked them in GPUs for a few months, and are now selling the output. Again, regardless of your opinion of the IP question, you have to admit that generative AI leverages artists’ own work against their economic interests. If someone’s livelihood is affected by this, of course they will feel venom. That is not parallel to your feeling of venom toward their efforts to stop this from happening to them.

I know the most about stable diffusion. Stable diffusion (based) models are typically open source, you can download them for free, and they easily fit on a cheap USB stick.

They're not actually supposed to have anyone else's pictures in them as such [1].

As far as I can tell, the training is more along the lines of "this is what a dog looks like", "this is what a cat looks like", "this is what a stick looks like" ... etc.

At this point in time, unaltered output is automatically public domain, because it is not made by a human.

[1] some early bugs notwithstanding.

There is a fundamental difference between a human taking time to learn a thing and a machine that gobbles up everything it can find in order to replicate it.

Good luck getting these people to even acknowledge that, though. Even though they'll acknowledge it in literally every other circumstance of ML.

Of course there are loads of differences between how humans learn and how AI models learn, from substrate to architecture. I don't see why any of those differences should make a difference to the morality or legality of learning from copyrighted material.

So, wouldn't licensing image with clause "can not be used for AI learning" solve that ?

Asking permission before building upon someone else's licensed work.

Make models that work without freeloading.

> trying to stop me from stealing your art is malicious

Yep, this is exactly the kind of take I'd expect from the AI crowd.

The thing about theft is that it involves deprivation of an asset or resource. Calling it stealing seems overwrought.

Call it piracy, then. Except instead of lone crackers breaking the protection on products of big corporations, it's corporations hoovering up the works of zillions of individuals, running it through their copyright-washing algorithm, and selling the results.

The AI learned about copyrighted characters largely from marketing materials. So why should corpos be able to force me to watch the ad for, say, the latest Iron Man movie, but at the same time say that I'm stealing from them when I use that ad to train the AI?

Guess what my stance on 'piracy' is.

Ah, so, the tragedy of the commons: you'll happily pirate every digital good you want, relying on others to make it worth its creators' time and resources to make it. Awesome.

You guessed.. poorly. I buy shit - piracy represents a service problem most typically, and in many cases does not in any way represent a lost sale.

The obvious conclusion is that I'm not tricked into thinking that some one is somehow deprived just because you bought into the 'piracy' propaganda. Side note to consider; why is it called piracy? Would many of the arguments offered by people trying to maintain their dragons horde of IP and the monies from that IP be weakened if we called it what it was - downloading a copy?

Most people are like this - they will gladly buy the things they enjoy when they can afford to. The scale gets tipped when the cracked version works better than the official version, or when some one lacks the disposable income. Its just really hard to demonstrate harm here because there is no harm.

Calling this "malicious" is a bit telling. It sounds like a thief calling a lock malicious. I know this isn't the typical response here (I hope not) but it does sound like you particularly do think that training on art inputs is theft at some level.

Provide a cogent argument for why you are entitled to art that a given artist does not want in training sets.

How is this tool "clearly communicating" that intent?

The tool changes the image. Okay. So does cropping, changing color composition and a gazillion other things.

Wanna clearly communicate that intent? robots.txt. Exif data. Attributes in `<img>` elements.

> What are the best defenses for AI builders against malicious tools like this one?

Technically? Incorporate Glaze into the training process.

Glaze takes advantage of an ML model's fragility towards noise and uses that against the model.

The simplest way of incorporation is to run Glaze on all of the training data, whilst duplicating

Glaze, reductively speaking, is the adversarial part of a GAN, wherein MidJourney/SD is the generator part of the whole system. What's happening now is that the adversarial part of the GAN is purposefully producing noise to throw the generator off of its rocking horse, and now the generator needs to learn to "remove" said noise when training.