Quite similar to the New York Times lawsuit. If you can reproduce an original work, then use of the source material was obviously not transformative, and not covered under fair use. These systems can currently function like a lossy compression algorithm.
This is going to be a mess until higher courts in the US or Europe make some rulings. Personally, I think they are going to favor the copyright holders. A few billion handed out to damaged parties, and future commercial systems only trained on licensed content.
Well, it may be legal to create such a painting, but (depending on details) it may be a derivative work, in which case the original work's copyright still applies.
In other words, in your example, the painter is not necessarily free to distribute their painting while ignoring the original image's copyright.
Of course, this is not about the works. This is about the painter, and it's mostly not about exact duplication of the work but about works based on their previous work. You know, like other painters do.
Ostensibly, they want to prevent OpenAI, and any algorithm from making works based on their work.
But, frankly, what they really want is not really that. What they really want is to prevent it from making any works at all.
The real problem is these algorithms making works of similar quality to the artists. Thousands for $20/month (in OpenAI's case)
Now there are egregious examples where this is abused to target a specific artist, and I'm sure you'll find all the AI companies are entirely on board with preventing that.
The NN has copyright material stored (maybe in compressed form) inside of it. That's quite literally the definition of copying, and if you do that without licence copyright infringement.
Note how software licences typically give you the licence to make a copy for running the program? That's because the act of copying the program from the medium (cd, USB stick, HDD...) is considered copying.
I really don't understand what's so difficult to grasp about this. And yes it is illegal to sell equipment whose purpose is circumventing copyright protections.
> The NN has copyright material stored (maybe in compressed form) inside of it.
There isn't anything like enough space in one of these models to store all the training material, "compressed" or otherwise. Not by orders of magnitude.
I don't know US, but if similar to Australia then you can sell the machines, but there are specific regulations regarding public access to them, which boils down to posting notices about copyright violations and passing liability onto the user. Because apparently a stern warning stops students from photocopying expensive text books. Which is the sort of thing Midjourney and OpenAI will be relying on for their defense, yes.
It is also interesting that there are technology specific laws. You can't sell photocopiers to the public capable of reproducing bank notes for example. Or taxes added to blank medium such as cassette tapes.
Midjourney cannot grant you the copyright to an image it generates, because it may generate the same or very similar image for another user. The best it can do is grant you a non-exclusive permissive license. And maybe it would get in trouble if it granted a permissive license to a copyrighted or trademark infringing image, or maybe the copyright or trademark claim simply trumps the license.
Midjourney's story is different. Midjourney sees it's AI as a tool, that you pay to use to make images. According to them, it's the user making the image, and the user has copyright (other than giving Midjourney a license to store, show, train, ... that image)
If he’s mad at that wait till he sees what the cp command does!
In all seriousness, midjourney is like any other tool. You can break the law with it, big surprise. It should still be on the person that broke the law, not the makers of the tool. If you go down the “can’t use my art as training data” how far do you push it? I watched Batman, does that mean if I draw caped super heroes I am stealing training data? Wouldn’t that apply to all comics after Superman? Isn’t he guilty of derivative work himself?
Also, nobody is being harmed by this. AI is always derivative. Nobody is interested in copies or impressions of popular art, otherwise the guy selling animes on velvet at the local art fair would be making more money. This just sounds like a Luddite shaking his fist at this perceived enemy.
> If he’s mad at that wait till he sees what the cp command does!
in order for that analogy to even remotely be applicable, midjourney would have to be a program you run on your own computer and not a service running on midjourney's computers that does what you ask it to.
> In all seriousness, midjourney is like any other tool. You can break the law with it, big surprise. It should still be on the person that broke the law, not the makers of the tool.
that would be midjourney, since they're both the creators and operators of their model.
> If you go down the “can’t use my art as training data” how far do you push it? I watched Batman, does that mean if I draw caped super heroes I am stealing training data? Wouldn’t that apply to all comics after Superman? Isn’t he guilty of derivative work himself?
you just called AI "a tool", that's why it's different.
> Also, nobody is being harmed by this. AI is always derivative. Nobody is interested in copies or impressions of popular art, otherwise the guy selling animes on velvet at the local art fair would be making more money. This just sounds like a Luddite shaking his fist at this perceived enemy.
that's an argument against IP law in general, not an argument against holding companies which create and operate deep-learning models to the same standards as everybody else.
> > If he’s mad at that wait till he sees what the cp command does!
>
> in order for that analogy to even remotely be applicable, midjourney would have to be a program you run on your own computer and not a service running on midjourney's computers that does what you ask it to.
So your argument is that the tool running remotely as a service somehow makes it different from one running on your local computer?
So if someone copies the contents of a novel into Google Docs, Google is liable for their copyright infringement, but if they paste it into Word running on their computer, all of the liability is on them?
At what point does the liability shift? What if all of the functionality is implemented in JavaScript in the browser, rather than running on someone else's server?
What if they run all but the last layer of the model on the server, then finish the computation locally?
I don't think the situation is nearly as cut and dry as you make it out to be.
>So your argument is that the tool running remotely as a service somehow makes it different from one running on your local computer?
no, my argument is that the tool being run by midjourney makes them liable for how it is operation. because it's their model, their program, their computer. They are responsible at every level.
>So if someone copies the contents of a novel into Google Docs, Google is liable for their copyright infringement, but if they paste it into Word running on their computer, all of the liability is on them?
not sure why you think that's relevant, but yes google is liable for copyright infringement facilitated by their platform. that's why if you upload a full-length video to youtube and the rights-holder posts a DMCA, google will delist or delete your video.
> At what point does the liability shift? What if all of the functionality is implemented in JavaScript in the browser, rather than running on someone else's server?
why the fuck would javascript change anything?
>I don't think the situation is nearly as cut and dry as you make it out to be.
it really is, there are well-established precedents for how this works and it's only not "cut and dry" if you arbitrarily consider AI startups to be exempt from the same rules that apply to everybody else.
> no, my argument is that the tool being run by midjourney makes them liable for how it is operation. because it's their model, their program, their computer. They are responsible at every level.
I still don't see how this would be any different from something like Adobe hosting a cloud-based version of Photoshop. Someone could use that tool to create an image that infringes on someone's copyrights, but I think you would be hard-pressed to convince people that Adobe is liable for the drawings that other people create with their tools.
> not sure why you think that's relevant, but yes google is liable for copyright infringement facilitated by their platform. that's why if you upload a full-length video to youtube and the rights-holder posts a DMCA, google will delist or delete your video.
The YouTube case is about Google themselves redistributing the copyrighted material. The distinction with the Google Docs example is that Google isn't distributing the copyrighted material, it's just a tool being used to facilitate the creation of that material.
The JavaScript remark was just because you implied that it made a difference whether the software was running locally or on Midjourney's servers. If that's the case, I want to know where you think the line is drawn between the two situations.
So, no, I still don't see how this is as cut and dry as you claim.
> > > If he’s mad at that wait till he sees what the cp command does!
> >
> > in order for that analogy to even remotely be applicable, midjourney would have to be a program you run on your own computer and not a service running on midjourney's computers that does what you ask it to.
> So your argument is that the tool running remotely as a service somehow makes it different from one running on your local computer?
> So if someone copies the contents of a novel into Google Docs, Google is liable for their copyright infringement, but if they paste it into Word running on their computer, all of the liability is on them?
But that's not what happening here. If Google docs hat a button, "put Harry Potter books here" that when pressed it puts the whole text of the Harry Potter books into the document. Now who do you think should be liable for copyright infringement, Google who put the whole text of Harry Potter books somewhere into the Google docs source code, or the user who pressed the button? Your argument seems to be the user, but I find that highly illogical.
The central issue is that the copyrighted images are already inside midjourney, the user is just getting them out via queries (pressing a button). The user is _not_ adding the copyrighted material to midjourney.
If you were to search any Harry Potter book using google search I'm sure you will find the unlicensed content, served up to you by google, very quickly, with less prompt than midjourney needs. You can hand wave about how it's so different but I see the search example returning a hashed copy vs. midjourney making something you could argue in court is infringement. The thing is, both are already illegal. You don't need additional government meddling in code to make illegal something that is already illegal.
In all cases, the person committing the crime is the person to prosecute. What would happen if I used Photoshop to carefully recreate the image by hand? Or even a paint brush and canvas? It would still be illegal! What are you arguing to make illegal? The paint brush? For _some people_ making clones of copyrighted work is very easy, this tool makes is easy for more people. But for you, _it's the tool?_
And in any of these cases you're trying to protect the same IP holders that cut off access to content you purchased, so any idea of fairness in copyright is a joke at this point anyway. I work in that industry so I knew it was possible, but I never thought they would do it with a huge swath of content like what happened to play station owners. IP holders deserve no sympathy or protection under the law after that move.
You might say "but these are not the people we're trying to protect". Too bad, that's what any IP law restricting generative AI would protect. And consider the draconian means required to do so.
> If you were to search any Harry Potter book using google search I'm sure you will find the unlicensed content, served up to you by google, very quickly, with less prompt than midjourney needs. You can hand wave about how it's so different but I see the search example returning a hashed copy vs. midjourney making something you could argue in court is infringement. The thing is, both are already illegal. You don't need additional government meddling in code to make illegal something that is already illegal.
That's not how the Internet works. Google is not serving you the Harry Potter books contents, they (might) point you to a different server that (might) be serving you the books, but what theat website would be doing is in fact illegal in most jurisdictions. In contrast midjourney are the ones serving you the content, that is a huge difference and not "handwaving". If you don't see a difference there than I really don't know what to say.
> In all cases, the person committing the crime is the person to prosecute.
Yes and it is midjourney who is committing the crime. Their models contain the unlicenced copyrighted material and they sell a service using that material.
> What would happen if I used Photoshop to carefully recreate the image by hand? Or even a paint brush and canvas? It would still be illegal! What are you arguing to make illegal? The paint brush?
Why are you continuing to ignore the central point that midjourney contains the copyrighted work? That's completely different to being a brush?
> For _some people_ making clones of copyrighted work is very easy, this tool makes is easy for more people. But for you, _it's the tool?_
Yes because "the tool" is already making the clone, inside it's model.
> And in any of these cases you're trying to protect the same IP holders that cut off access to content you purchased, so any idea of fairness in copyright is a joke at this point anyway. I work in that industry so I knew it was possible, but I never thought they would do it with a huge swath of content like what happened to play station owners. IP holders deserve no sympathy or protection under the law after that move.
So instead we are happy for even larger corporations to just take all creative outputs (and that includes lots of open source software) simply feed their models and resell this work of others through their models essentially disowning the previous creators without paying anything? Now if these models would actually be freely available and all parameters published I might be OK with it as the end of copyright as we know it, but that's not what is happening here. We are essentially witnessing the dispossession and monopolization of knowledge by a few large corporations and startups.
> You might say "but these are not the people we're trying to protect". Too bad, that's what any IP law restricting generative AI would protect. And consider the draconian means required to do so.
> Nobody is interested in copies or impressions of popular art, otherwise the guy selling animes on velvet at the local art fair would be making more money.
The vast majority of artists make almost no money off their art. Original or otherwise.
Gerhard Richter's painting 'Cathedral Square, Milan' is a 'bad' copy of a photograph:
> The source after which Richter executed Domplatz, Mailand was a newspaper photograph that was in focus, from which Richter clipped a section and modified it.
> It was sold by Sotheby's in New York on 14 May 2013 for 37.1 Million dollars, breaking Richter's own record price for an artwork by a living artist, his 1994 32.4 million dollar painting Abstraktes Bild (809-1).
[0] Of course, motivated reasoning can always find some aspect that makes the superstar artists and their art original. But the same amount of mental effort can also make 'the guy selling animes on velvet at the local art fair' original.
> In all seriousness, midjourney is like any other tool. You can break the law with it, big surprise.
I believe the point was rather that the output makes it obvious the "tool" was trained on copyrighted work.
As an aside to your justification though, if a main purpose of a tool is to circumvent copyright, that tool is usually illegal.
> I watched Batman, does that mean if I draw caped super heroes I am stealing training data?
It means if you draw a near copy and tried to sell it, you are committing copyright "theft" (I don't like the term, but that's what it's called).
> Nobody is interested in copies or impressions of popular art, otherwise the guy selling animes on velvet at the local art fair would be making more money
People are absolutely interested. I couldn't get RedBubble to even print a non-copyrighted work (which they mistakenly thought was copyrighted) that I had altered. Just because some people are getting away with it, doesn't mean it's legal or wouldn't immediately come to a stop if they came under scrutiny.
> This just sounds like a Luddite shaking his fist at this perceived enemy.
Luddites had some very good points, so he's in good company.
The picture in the article is a lot more than drawing a caped super hero after watching batman. Not to mention that the access to viewing batman for the inspiration is usually behind some sort of compensation.
I would argue that original creators are being harmed by this. Not in a traditional sense, or even in a legal way with any precedence but in a thousand cuts sort of way that is novel.
I believe he is calling out the company not the tool. And calling them out because they infringed copyright by showing the copyrighted material to the model during training (as evidence by it being able to reproduce it so faithfully)
You are arguing that Spotify should not pay royalties or license fees for songs, because it is the user might break the law by playing an unlicensed song.
Oh, wait. It isn't the user. Spotify and Midjourney are actually the guilty party in this scenario, for distributing copyrighted works and probably for storing them without permission in the first place.
Midjourney may be a tool, but it is a tool containing a large body of copyrighted and unlicensed images, including trademarks, and even illegal content. And that is what the court cases are going to be about.
“If You knowingly infringe someone else’s intellectual property, and that costs us money, we’re going to come find You and collect that money from You. We might also do other stuff, like try to get a court to make You pay our legal fees. Don’t do it.”
That seems like a tenuous defense at best. If your prompt references copyrighted material and Midjourney generates a copyrighted image as a result, it's clear the training data must have contained that work, and whatever filters they have didn't stop them from serving it to you. I can't imagine a court siding with Midjourney in that case.
It's amazing to read what people think is fair use. Have a read of the wikpedia article, it's definitely not as straight forward as adding a slight variation and then I'm good.
If that was the case all of the following would be fair use (hint they typically are not).
- making a movie from a book (I mean it's a pretty big variation)
- making a sequel/prequel to a book, movie using the characters of the book
- making new comic stories using Mickey Mouse/Batman/Superman...
- Making merchandise using redrawn characters from comic books...
...
Also it's important to note that one of the important aspects to judge fair use is commercial interest, i.e. what is allowed for an individual privately, might very well be copyright infringement for a company (like Midjourney) who sells a commercial product.
I am looking at it from the perspective of someone using mid journey to generate the image, not the company itself.
What if I want to use midjourney to reimagine a new scene between to scenes to prove whatever social issue I want to blog about. I then post the blog with the image.
Would there be a difference if I hand drew it? Is me describing the movie in the context of my social issue considered copyright infringement because it is a compressed memory?
Would it, really? And would you own that new scene?
Midjourney is a paid service which uses copyrighted works as source, without any agreement with the copyright holders, while declaring that its users own all the assets that it creates for them.
I am not a lawyer but adding your own items and creating a hypothetical scenes seems more like fair use than the reaction videos on youtube do? Or, is that a different ballgame altogether?
I'm not sure this really qualifies as anything new. It's been well known for a long time that many generative AI models are capable of reconstructing copyrighted works (copilot, SD, etc.) from their original training data with nearly 100% fidelity, but until we have a definitive legal ruling it doesn't really matter.
Is a definitive legal ruling even possible?
Looking at music for example there are still many cases being fought over stolen melodies and that’s been a thing for generations.
I think it’s going to go a similar way with „AI“ - lots of court cases, lots of different and nuanced rulings converging on a general idea what the legal boundaries are but with many grey areas in between.
I keep doing a mental trick where I ask, "what if this were a feature of Google/YouTube?"
I think you're allowed 30s clips (or something) while staying within fair use (see: film YouTubers are able to use some scenes and clips)
Thus, you could imagine a Google that would show you the clip of the movie or link to a 30s scene from the movie and then prompted you to rent/buy the movie to see more.
I think Google would get away with that. I'm not positive or a lawyer.
Is this different because it's still images only? Again, Google shows copyrighted images in its image search (usually thumbnails, and then maybe serves the full image from the source? Or sometimes sends you to the source?)
It's not quite clear to me that simply showing the image is copyright infringement - non generative ai companies do that. So where is the line?
My understanding is that it really should be incumbent on the user (like it would be if your search found copyrighted material) but I don't have a legal basis for that! It just seems like nobody is suing Google for image search.
The image generator that provided the screenshot without licensing/permission.
This isn't complicated, we have laws on the books, and I hope legislators learned their lesson after the "Uber's not a taxi company, and AirBNBs are not like hotels"
So if someone makes a forgery of, say, a picasso, who do we blame: The forger or the brush?
> This isn't complicated
This concerns copyright law, which is handled case by case, and deals with very fuzzy definitions like "fair use" or what constitutes "transformative works", and things like "impact in the marketplace".
The existence of the forgery isn't the problem!! Selling the forged painting _as if it's the real one_ is the problem (am I missing something?)
In _Tim's Vermeer_ he perfectly (too perfectly) forges a Vermeer. That isn't illegal!! He has it hanging in his home, and whoever is so lucky to inherit it will get to keep it for themselves too.
But if he did it again and then tried to sell it as an original Vermeer - that would be illegal!
If I go to Google and ask for screenshots from a movie, it will also produce copyrighted material. So you're right that it isn't complicated. Tools with the same capability already exist and are perfectly legal, and therefore, so is Midjourney.
Does Google claim the image is it's own? No. Does Google include a message that the image is copyrighted? Yes.
Does MidJourney attribute the source of the image? No. Does MidJourney's owners claim the requested image is an original generated by their black box correlation machine? Yes.
The images aren’t exact copies, though they’re obviously really close. Considering those are all famous stills from the movies I kind of doubt they actually just took random scenes from the movies themselves - instead it was trained on just images from the internet.
Point A means it might be transformative enough. Point B means it possibly wasn’t deliberately trained on copyrighted material, though obviously they know stuff would be in there. Further, no court has decided whether or not that’d be illegal anyways. My money is on that it won’t be, because there’s really no legal argument that it should be.
If a painter trains himself on copyrighted stuff he’s certainly allowed to sell his own paintings, even if they’re incredibly close to his training material.
I don't know. This sort of thing will lead to stuff like recording your own memories requiring a licensing fee if not careful. I also think there's a reasonable argument that the prompter is the one creating infringement.
If someone says "imagine batman on a boat" am I not prompting what's essentially my own internal generative AI that's been trained on seeing copyrighted content at some point? If someone says "remember that scene in..." isn't that a form of recovering a copyrighted sequence from your own "trained" memories? Once there's a better ability to extract data from someone's neural activity, remembering or imagining will just be another form of prompting.
In the Visual Arts (drawing, printmaking, painting, even sculpture) copying the masters has been practised, and even recommended for centuries (if not longer). I believe it's the same with music. It's what you will want to do to some extent in order to get better at doing your own thing - some do it a lot, some do it less. None does not do it.
The tool may be a brush, it may be a pencil, it may even be a contraption or machine of some kind. It could even be a camera, or software. This does not matter. The result is typically labelled "a studio", "a reproduction", or if has significant deviations from the original "a paraphrase".
Some artists include elements of the work of other artists in their own work regularly, even copyrighted "characters". The thing known as "sampling" did not originate with rap music. Here there's a line somewhere (just like in rap) that, how much is too much? The answer is always that some cases are "fair use" and others are not.
My concern in all this is that people scared of this new tool (AI) will react in a way that leads them to demand restrictions in the way that traditional artists have worked like, forever.
In my view the AI model is no different from a camera, or that combined with some software suite like blender, photoshop, or even gimp. It's just a tool. And, well it can create reproductions. Reproductions are just that, and no more.
To me it seems pretty unambiguous that the infringement is a problem, and I even agree with the idea that Midjourney, not the users, should be held responsible if their model is being used as a service. I can’t come down on what solution I actually think is reasonable though. Restricting training data to non-copyrighted works would pretty much make AI art non-viable for any modern style, which would make plenty of people happy but feels potentially problematic if the nature of AI continues to change which it certainly will. Restricting output feels more right to me but also feels substantially less effective, since you would need AI to do any practical restriction automatically and what margin of error there would be acceptable? I definitely think this should be clamped down on but I can never settle on what I think would be reasonable and not just totally dismissive of at least one major set of concerns.
Should you pay royalties if someone says to you "remember that scene from Blade Runner" or "imagine if Jean Luc Picard beamed into the middle of a Batman movie" and visualize something similar to something highly similar to a movie or TV show you've seen?
Copyright law does not apply to imaginary reproductions, such as your imagination. Your question becomes more interesting if get someone to draw that scene from Blade Runner. At that point, it comes down to if a judge considers the derivative work transformative enough (for hand drawn sketch, it would be, except maybe if drawn by a savant). And in the Batman case, you might violate trademark law depending on what you did with image after you drew it, which is a whole other set of rules.
When was the last time you heard the original Happy Birthday song sung on TV? It is now generally avoided, because producers don't want to pay royalties to the copyright holders.
We should assume that at some point we'll be able extract imaginary and remembered images and sequences from neural activity - I think very poor/crude versions exist already.
Even if it's the paid version, the prompter had it make the infringing content for their own use. And it's only going to be anything more than equivalent to someone making their own sketch of Batman if the prompter reproduces and distributes. It's not like it's illegal for you to tell your artist friend to sketch a Batman either - Comic Cons and such are full of that.
So you need a legal process to determine if that generation was actually infringing on copyright.
IANAL so I have no idea really, and I think most of the other commenters here can’t say for sure either.
Prompters aren't artists, and neither are image generators. The corporations that create these generators are responsible for the content they illicitly fed into them that they are now spewing out.
If you sell me an endless coloring book that occasionally spits out copyrighted content, that's on you.
Of course not because you PAID to see it in the first place. Or you watched it with someone who paid. Or you saw it on a website that was free and THEY paid for it. If you never saw it, you would have no idea what it looked like.
You had to see it at some point, the way in which you see it is the issue, not the memory of it.
If you're arguing that the AI is just like a random human browsing the web and they happened upon these things and now they are allowed to sell those things back to me. Man that seems like so uncompelling of an argument to me.
We're not allowed to do this in any other context. I can't learn music and go tour, take requests from the audience, and act like it is mine. No credit or anything. Sure on a small scale, and there are times when it's unclear, but I can't go to Wembly without paying the original owner when someone asks me to play "Thriller by Michael Jackson"
I'm pretty sure plenty of people have seen Batman without paying. Cartoons were just broadcast over the air - advertisers paid I guess. But also if your friend had a video or DVD, or you checked out a comic from the library, or many of dozens of similar situations. Similarly, comic cons are full of people drawing endless numbers of superheros they have no rights to. Hiring an AI is not buying back copyrighted works, it's buying trained experience and having it produce output for you, IE work for hire, like you might get out of a comic book artist. And it's on the hiring party. If there's no sin in a human artist being able to draw batman because they've seen or trained enough, why is it a sin for AI?
I feel like I responded to all of these arguments in my original comment but:
1. The TV station paid for the cartoons and monetized them through advertising
2. As I said, your friend paid for the DVD, The library paid for the comic, etc. SOMEONE is paying or else the original maker would not be able to sustain making the thing. I can't think of one scenario where you consume media where someone didn't pay for it downstream.
3. A comic book artist creating comics inspired by batman, but not a comic OF batman. Again, as I already pointed out, on a small scale this happens all the time, it's just not word DC sueing a guy on the street doing one painting a day.
If the argument is you should be able to pirate or everything should be free, all good, no judgement, but that's not how it works in the world right now.
It IS a sin for a human to get a batman plushie made and sell it. That's my point. That's what MidJourney is doing when they charge $10 a month to make batman photos.
No manufacturing company in the USA let's you do this, that's why you have to go to China to make Nike knockoffs
I feel like an artist should be more upset with what could be created from the prompt that was not true to the original film, with only minor changes to the prompt. For instance, a picture of batman and joker having drinks together on the roof top.
I don't know if this is the case, but if these images were released to press & public as part of a media resource pack with a permissive license used to market the movie (which I believe is commonly done in this industry), I'd have a hard time empathizing with the viewpoint that Midjourney is doing something wicked by including it in their training data.
You are right, the regurgitation was indeed produced by the studio for promotional purposes. But notice how the image is rendered at the link: with credit and copyright. And specifically, for promotional purposes. And while I don't presume to know the specifics of the licensing of that image, I wouldn't either assume that this use is licensed. Especially without copyright and credit.
A web browser lets you easily commit copyright infringement by right clicking on a copyrighted image you do not have the rights to and selecting download.
This is going to be a mess until higher courts in the US or Europe make some rulings. Personally, I think they are going to favor the copyright holders. A few billion handed out to damaged parties, and future commercial systems only trained on licensed content.