Current US copyright law is not clearly in a place to view model training as infringement. Courts have a long history of permissiveness in the face of copyright challenges to new tech (e.g. the image search engine cases, Google v. Oracle and smartphones, Sony v. Universal and VCRs) and I predict it will happen again with AI. The cat is out of the bag and judges know that finding training to be infringement of each training example will have a negative impact on a new product category. If training was more obviously infringement then that permissiveness would be harder to sell, but in my opinion it's really difficult to argue that a "copy" of an example has been made during training (aside from the copy made to process the example).
I'm not so sure about this. I'm not saying courts in the US will rule one way or the other, I'm just saying it's certainly not a forgone conclusion that training is fair use. Even if it is, the companies might not have sone their due diligence.
Lots of data they trained on is available for purchase (e.g. artists often sell prints or reproduction rights, the books in books3 are widely available, etc). It's my understanding that companies like Stability and OpenAI did not attempt to determine if the data they trained on was available for purchase and then buy a legally purchased copy for training. That might cause them to run afoul of fair use doctrine in the US (not sure of other jurisdictions).
See these excerpts describing fair use for copying library materials [1] (many of these collections are being released by groups referring to themselves as libraries):
> Copying a complete work from the library collection is prohibited unless the work is not available at a “fair price.” This is generally the case when the work is out of print and used copies are not available at a reasonable price. If a work, located within the library’s collection, is available at a reasonable price, the library may reproduce one article or other contribution to a copyrighted collection or periodical issue, or a small part of any other copyrighted work, for example, a chapter from a book. This right to copy does not apply if the library is aware that the copying of a work (available at a fair price) is systematic. For example, if 30 different members of one class are requesting a copy of the same article, the library has reason to believe that the instructor is trying to avoid seeking permission for 30 copies.
> The copying, whether performed by the library or whether unsupervised by the library patron, cannot be for a commercial advantage. This means that the library (or a copying service hired by the library) cannot profit from the copying. In addition, the copying for the patron must be done for purposes of private study, scholarship, or research.
The availability of the copyrighted works is not determinative. Fair use in the US takes (at minimum) four factors into account, listed in the federal copyright statute: https://www.law.cornell.edu/uscode/text/17/107.
That quote from Stanford's library is not discussing fair use doctrine in general, but rather is stating what is permitted in those specific circumstances. There are plenty of instances of fair use where the underlying work used was available at a fair price. That's the whole point of fair use law: some use of a work that is facially infringement escapes liability because the particular use is considered fair.
That's a pretty low bar though, no? Everyone just updates their T&Cs and laughs their way to the bank with free training data - the real meat of the claims was what was dismissed.
Yeah; headline leaves out the pretty salient bit that the plaintiffs can amend their claims to be more accurate, which will only make the case more logistically sound. Headline makes it seem like the case is over, when it's not.
It's hard to remember exactly what my problem with it was, but if you look up the original filing and read what the artists claim is being done with their images, you can clearly seem some invalid assertions on their part. That's why I even took an interest in the case, in the first place. I like Anderson and I think what she's aiming for is a good goal, but when I read the case I was like "oh shit, this is doomed to fail."
Again, I don't remember what the exact description was, but I remember it jumping out as obviously wrong (a layman's interpretation of a neural network). Feel free to go over that and it should be pretty obvious how the claims could be amended. Barring that, just wait for the filing because I'm sure the plaintiff lawyers will be jumping all over this.
The IP lawyers I've talked to about this at length—and who don't have a particular axe to grind—are pretty skeptical in general outside of a specific output being a derivative work of a specific copyrighted work. The argument against generative AI seems (IANAL) to rest on there being some sort of collective copyright (across many different creators and even including works that are not actually copyrighted at all) which somehow carries over to an output that, had a human created it would probably be seen as an original work.
I get the emotional appeal to companies are using my stuff without paying me or even crediting me. But, as I understand it, works are normally considered derivative of specific works—not some large corpus.
Collective copyright-like arguments do look strange but they were the core argument in the pirate bay trial. The founders were found guilty of assisting in infringement of copyright where the specific case of infringement and the specific copyrighted work is both unknown and deemed unnecessary to define.
The argument as it went went like this. Is the theory that no infringement has occurred believable, or can it be said as a forgone conclusion that over the course of the operation that some specific copyrighted work has been infringed on at some point in time and that the operator knowingly were aware that such cases was likely to have happened.
It is true that many lawyers and legal experts thought that the pirate founders were immune to such claims since usually one need to produce specific cases of specific copyrighted works in order to find someone guilty of assisting. The case however illustrated that such requirements are not always needed, and the Swedish supreme court did not feel it necessary to analyze it further.
>The argument as it went went like this. Is the theory that no infringement has occurred believable, or can it be said as a forgone conclusion that over the course of the operation that some specific copyrighted work has been infringed on at some point in time and that the operator knowingly were aware that such cases was likely to have happened.
Though in this case, I assume that a great many lawyers would argue that it's hardly a foregone conclusion in this case.
The likelihood that no output has ever infringed any copyright will be an interesting argument. People are constantly testing and prompting different AI's to see if they can get them to reproduce known content, and there are a few known examples from copilot where such test has shown large amount of reproduced exact copies of comments and other identifiable segments. Similar things has been done with images, like the getty watermark.
AI developers has also started to add filters and other techniques to remove outputs that are too similar to existing works, which is both a good thing and a bad thing. It shows that the tools do output such works (as otherwise they wouldn't need to filter it out), but also show that they are working to minimize it. Courts would have to look at it and decide if such efforts are an admittance of the issue or if the mitigations are enough.
In this specific case, I doubt such discussion will occur at this point since the main point that the judge brought up is that the copyrighted works need to be registered first at the copyright office before it can go any further (except for 12 works which already are registered).
If it turns out that passing a data set through PyTorch creates enough distance between the source material and the resulting data set to invalidate copyright protections, it will be interesting to see what happens when somebody does that to AI companies' assets.
With the "edge computing in AI" push (offline processing on phones and the like), we might see soon enough.
Seems like a good table-setting measure. Judge is saying "what you claimed isn't how it works; if you want to amend your claim to say how it actually works (and why that's infringing), you can do that now."
Judge is also letting the 'unauthorized use' charges move ahead.
To me, this feels like the right way to make sure the questions at the heart of the trial are satisfied by the results of the trial. SOP, for sure, but it definitely could have gone the other way, with the judge just throwing it out altogether, so I'm glad this is moving in this direction, instead.
As a non lawyer my understanding is that it is no longer necessary to copyright images to be protected as creator of a work, but it seems you need actually to copyright work if you want to sue.
No, you copyright the work if you want to claim all the damages like punitive and such. You can still sue to prevent usage of your work without having formally filed a copyright.
Headline is quite a bit more definitive than the actual ruling:
> The three artists — Sarah Andersen, Kelly McKernan, and Karla Ortiz — immediately ran into problems as two of them — McKernan and Ortiz — did not register their works with the U.S. Copyright Office.
Two of them got bounced, Anderson remains.
> Judge Orrick writes that it is “unclear” as to whether Stable Diffusion holds “compressed copies” of the images and points to the defense’s argument that the training dataset, which contains five billion images, can “not possibly be compressed into an active program.”
Once again, regulation that's needed is complicated by the fact that people in power do not understand how this shit works.
However:
> The judge has offered the plaintiffs an opportunity to amend and clarify their theory as to how Stable Diffusion operates its training data.
And:
> However, the judge allowed Andersen’s complaint that her 16 copyrighted works were used without her authorization to move forward.
So the case is still developing.
However, this is frustrating:
> The problem for the artists is that the training data for these programs is a black box. Outside of LAION, very little is known about what exactly went into training AI image generators but it is widely assumed that the companies did an almighty scrape of images on the internet which included taking copyrighted and copyrightable pictures.
I'm not sure if the publication stated "widely assumed" here because of legal reasons or because the author doesn't understand this either, but the entire crux of Anderson's case here is that if you include the proper prompts, you get comics out of these generators that strongly resemble hers, in both style and content, which are copywritten. That an generator can do that basically requires that a good amount, and I mean a LOT of her comics were pushed into the training data, and not only that, but were tagged explicitly as her work. Given how much of LAION is scraped from social media, this is not surprising.
In order for the generator to know what Sarah Anderson's style looks like, her comics must be in the data set. Period, paragraph.
This case involved a guy under NDA working with a competitor to build similar product and included testimony from oculus that they used ZeniMax's code, literally, including copy pasted segments.
My read on the case is that it happened, John made a big public outrage post, and then gradually after news interest died down, more facts were revealed that made it much more obvious that oculus was in the wrong but nobody cared anymore.
It's also not relevant to making an ai model. That is at best relevant to producing a similar output using an ai model
Nothing is analogous to making an ML model because making an ML model and training it are fundamentally different processes than both copying something (which is what copyright is built on which is why it probably isn't going to work to handle these issues) and from a person learning, because the most brilliant person on the planet cannot in timely matter digest and learn from over five thousand images and then reliably reproduce works based upon them.
This an entirely new schema and the endless slap fights between people trying to say "well the computer is just learning like a person does" is just, bafflingly, wildly wrong headed. No it's not. A person learns so they can accomplish a goal, and an ML algorithm by it's nature wants nothing, it is incapable of want, versus the people who are saying "well ML lets people copy other people" which is also wrong headed, because it isn't copies, it's remixes and re-imaginings of similar material are just not productive. On the one side people trivialize the act of learning how to art being just looking at shit and reproducing it, which is just wrong, completely wrong, and on the other side people think it lets the ML model copy artists, which it dooooooes, kind of? But that test also fails because they aren't strict copies and them being copies isn't the goddamn problem. The problem is the style, the vibe, the unique look of a given creative is trivialized into something you can get from stable diffusion after a few tries with the right prompt. This isn't a problem because Sarah Anderson lost a sale or whatever the fuck. It's a problem because it completely undermines the notion of the output of paid creatives having value, which is already a shaky proposition for most people and why tons and tons of prominent creatives cannot earn a living via their creations. This is going to make that situation even worse and no matter how many times this is explained to people, they just run back to this "well I wasn't going to buy things from them anyway" and the entire thing makes me want to fucking scream.
All that to say: AI is not AI, it's ML. ML is not learning like a person does, it's fundamentally different, stop equivocating. If you're so determined to cheer on as silicon valley knocks already struggling creative professions down a hill with systems that COULD NOT POSSIBLY EXIST without their pre-existing output to exploit, without permission, fundamentally unethically, if you just don't give a shit about that and are going to gleefully cheer that process on, then all I really ask is you be honest about that. Just say "I know this is fundamentally corrosive, unethical, and undermines an entire segment of the economy, but now I can get as many over-tuned generic shitty pieces of art as I want, and that's just more important to me." Just say it, and stand by it.
And the synthesis was created with material that the AI company did not license and did not have permission to use for this purpose.
I have ZERO issue with image generators on their own. They're quite cool technology. The only thing I and tons and tons of other people have requested is they not use images without permission to train them, that's literally it. That's my sole, single, solitary objection and yet every AI interested person, advocate, and company rep will do intellectual back-flips to explain why this is unreasonable.
It's not unreasonable to ask permission before you use something. It just isn't.
If I take a copy of your art to hang on my wall, I've violated your copyright.
But if I "copy" the experiential knowledge of your art into my brain by viewing it, I'm not violating your copyright. My brain doesn't contain a copy of the art, it's just been influenced by viewing it, and I might be more capable of producing art that mimics your style.
What these models are doing feels, to me, vastly more like the second case.
But you are not capable of reproducing from your mind something that is in that artists' style, after that viewing. If you show someone a painting and ask them to recreate it, even setting aside the skill gap, you do not get the same painting back. You get a different painting, with some overlap, with the "focus" of it being what the person paid the most attention to during their viewing.
In this way, ML training is just not the same as a person being inspired by or even being asked to recreate another person's creative work. Over the course of making something, an artists' "voice" would be best characterized I feel as the tiny choices they make along the way that all point to and reinforce a larger point or purpose to the piece. This "voice" shows up in all creative output, not just spoken word. That is what I feel people are feeling is lacking in generated art: because a machine-learning model does not have a voice, it does not have intent, it has a mandate from a third party from which it tries to draw from, and instead of making numerous, tiny but contributory choices, it instead decides on a weighted average of all the choices made in the art that the model was trained upon, which is simply not the same thing.
It makes all generated art have this very sterile, soulless feeling to it because these tiny choices that would otherwise be made by a person trying to illicit an effect are instead just the machine sort of shrugging and being like "well in most things I've seen where a woman is sitting this way, her hand is tilted this way" but it doesn't know why the hand is tilted or what that means for the subject, which means the hand-tilt might be applied to subjects for whom it makes absolutely no sense at all to tilt the hand.
> But you are not capable of reproducing from your mind something that is in that artists' style, after that viewing.
I am capable of doing this.
> If you show someone a painting and ask them to recreate it, even setting aside the skill gap, you do not get the same painting back. You get a different painting, with some overlap, with the "focus" of it being what the person paid the most attention to during their viewing.
same for ai models. in fact, more true for ai models. You'll have a much harder time recreating an image from popular models than from human memory.
> It makes all generated art have this very sterile, soulless feeling to it because these tiny choices that would otherwise be made by a person trying to illicit an effect are instead just the machine sort of shrugging and being like "well in most things I've seen where a woman is sitting this way, her hand is tilted this way" but it doesn't know why the hand is tilted or what that means for the subject, which means the hand-tilt might be applied to subjects for whom it makes absolutely no sense at all to tilt the hand.
People are already experimenting and/or deploying systems that run an LLM before doing text2image [1]. This helps a lot with 'understanding'.
The next low-hanging fruit would then be to go back and improve the labeling used for training the image-generation models in the first place (using new multi-modal LLMs).
[1] eg. the current GPT+ standard model will do Prompt -> GPT-4 -> DALL-E
Some forms of copying/uses for a work are permitted by law (eu) or permitted by fair use (us) and thus exempt from permission. Eg. it is generally accepted that you are allowed to download works if you're viewing them in a browser, storing them in a search engine database, or are taking statistical data for a research project.
AI companies argue (long story short) that their usage is closest to taking statistical data, and thus doesn't trigger the need for a license.
I do realize that some people find this to be a somewhat unfair outcome as is, but
A) It is the current law as is. (caveat: we still have to see if the courts actually confirm this particular interpretation)
B) I'm not actually sure we should let those people have their way in the first place.
Of course, ignorance of the law is no excuse.
I feel like some creative folks never actually bothered to learn the laws they were working under, or (more charitably) this turned out to be a corner case, and they were caught by surprise.
> It's not unreasonable to ask permission before you use something. It just isn't.
It can be unreasonable if it makes it harder to society in general to reuse and innovate on previous works. Intellectual property in general can be quite destructive. Heck, take a look at Permission Culture, which is actually a negative thing:
Permission culture is a term often employed by Lawrence Lessig and other copyright activists such as Luis Villa[1] and Nina Paley[2] to describe a society in which copyright restrictions are pervasive and enforced to the extent that any and all uses of copyrighted works need to be explicitly leased. This has both economic and social implications: in such a society, copyright holders could require payment for each use of a work and, perhaps more importantly, permission to make any sort of derivative work.
Lawrence Lessig describes permission culture in contrast with free culture. While permission culture describes a society in which previous creators or those with power must grant people permission to use material, free culture ensures that anyone is able to create without restrictions from the past.
...
An implication of permission culture is that creators are blocked by systemic procedures and this discourages innovation. Requiring permission in this sense means that creators will have to prove their usage of material is fair, even where legally unnecessary,[4] which is a process that some would decide not to continue.[5]
>> He wrote that the sheer size of the LAION database may protect the company because it is “simply not plausible that every training image used to train Stable Diffusion was copyrighted (as opposed to copyrightable) or that all DeviantArt users’ output images rely upon (theoretically) copyright training images.”
This is where I think the judge misspoke. "Copyrighted" does not mean registered. Copyrights attach at creation and does not require anything of the creator. What the judge is referencing is the fact that registered copyrighted works are given additional protections, notably the award of attorney fees in disputes. But the fact that a work was not registered, what the judge incorrectly describes as "copyrighted", does not mean it is outside copyright. All the same rules apply governing how the material may be used or not used. You just don't get as money when you sue over a non-registered work.
I would hazard that the vast majority of copyrighted works are not registered. Every holiday photo taken is copyrighted. Every word typed on HN is copyrighted. Virtually none are registered. That doesn't mean such things are public domain, free for use by whomever/whatever runs across them online.
> This is where I think the judge misspoke. "Copyrighted" does not mean registered. Copyrights attach at creation and does not require anything of the creator. What the judge is referencing is the fact that registered copyrighted works are given additional protections, notably the award of attorney fees in disputes.
Aren’t there visual art in the public domain?
INAL but the judge’s ruling seems technically correct, so what am I missing?
Yes, photos taken by people working for the government for example, and anything shared with Wikipedia should have a license for free use (typically CC0) [0] That’s an important, valuable part of all images, but a very small segment overall.
The vast majority of images (holiday photos, posts on social media, Deviant Art, etc.) are not in the public domain, even if they are online.
Most images found on Wikimedia sites actually require attribution to the original author, including for derivative works. So if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output. (SVM is one machine learning method that can easily do this, but I'm not sure if it's ever been applied to image generation.)
> if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output.
This is a huge leap that would change copyright law dramatically, even for humans.
Youre asserting that copyright doesn’t just protect from reproduction, but from use as an example to learn from.
I think this would break copyright altogether. What musician or artist can say a work is 100% original and does not incorporate any details or concepts from work they’ve seen / listened to / read?
I know people hold that position, but it’s important to be transparent that applying copyright to training/learning (in addition to output) would be an overhaul of copyrigh, not a clear and obvious application of the law as it stands.
You only have to make that leap if you suppose that machine learning algorithms actually learn and then become capable of creating their own original works, rather than producing mechanical transformations of their inputs.
Which, I don’t think that’s an inherently absurd statement. I mean, at some point humans physically exist, so a sufficiently complex artificial being could exactly replicate our physically processes, and it would be hard (imo) to argue that such a thing isn’t a person.
But I don’t see the non-big-leap option. Either we admit that machines can actually perform creative works which are their own ideas, or we call the thing they do a derivative work, and we admit that we can’t just eyeball the difference between an original work and a derivative work anymore. That there’s something intrinsically special about having a human in the loop.
> You only have to make that leap if you suppose that machine learning algorithms actually learn and then become capable of creating their own original works, rather than producing mechanical transformations of their inputs.
Why do you think humans can do that, and that what we call “original works” of humans are not just mechanical transformations of the human’s inputs.
How, in a universe of deterministic physical laws, is it possible for a human to do anything else?
This presumes that an input having influence on an output means that the output qualifies as a derivative work of that input. A work being in some way derived from another work does not mean that it is a "derivative work" for the purposes of copyright law. Many artists want to make this presumption, but IMO it's extremely dubious as a general claim. It is likely that when this all shakes out, specific works will be able to be established to be derivative works of specific other works, but simply being an input will not cause that. (caveat: IANAL, but I do have a law degree and IP experience)
> So if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output.
That would be like requiring human artists to enumerate exactly every single influence on each piece of art they make. Did they see a painting by so and so at six years old that had a 0.1% influence on them? Better include it too.
Publicly-licensed works are 100% copyrighted. That's why they need licenses. A true public work doesn't need or get any licensing. Linux is copyrighted, and licensed, to become f/oss. Shakespeare is actually public domain, no license required.
Copyright does at least in theory lapse, depending on the nature of the work. I don't think I've seen any AI companies claiming all their training dataset is pre-copyright however.
> Those who more or less think copyright shouldn't exist in any case
I still believe a little bit in Jefferson's vision. It's just that that has so little to do with copyright that we're probably not even talking about the same thing. He envisioned copyright as being an incentive to create and do new things, to learn new things.
And here, as it would be used, it would shut down new technologies and quash new ideas. Clearly, such people are doing copyright wrong. I am not religious, I do not believe in "sin", but if that word did have any real use, it might be strong enough to describe those who would misuse copyright.
In any event, Jefferson was granting a very limited privilege and not describing a fundamental human right. This privilege was granted by the people, and the people do have the right to revoke that privilege if it continues to be misused. This, I think, is all the warning the misusers deserve. A comment that will be read by hardly anyone, that they will surely never see, and which will give them no headsup at all.
Copyright is a human right only in that it's the right to stop your stuff from being stolen. Which really isn't a right, but a set of rules that are needed for a developed nation to prosper by ensuring people make productive work in pursuit of selfish riches. Otherwise, you could make something, and someone could come along, take your idea/work, and undercut you slightly to put you out of business.
In the way in which you literally must make a copy of the dataset to use it in training, and that literal copy if not excused by fair use analysis including consideration of the downstream purpose, is a violation of copyright.
Ah sorry, I was trying to say two things at once referring to two different contexts. In both cases the only 'temporary copying' takes place.
- In the context of specifically training an AI-model under EU law: -
Article 5(1) of the Directive 2001/29/EC [1] is argued to apply. Of course, if sued, AI companies still need to show that their use is otherwise lawful (which should be fairly easy, since they actually do not retain anything remotely resembling a copy at all)
- In the context of specifically training an AI-model under US law: -
The purpose of the copy can be argued to be non-exploitative, incidental, for the purpose of enabling technology, and temporary.
By contrast: Google Books even argued that permanently retaining entire copies of books wholesale was fair use [2], provided they didn't provide copies of those books to 3rd parties.
OpenAI argues that there's no way they're doing anything even remotely close to that. [3]
You browser loads a website. That is copying. Your browser saves a copy on a hard drive: copying. You integrate an image into a database: copying. Your CD player reads a song from a CD into its memory buffer: copying. All of these debates have already been hashed out in various courtrooms. There is a huge body of law that carefully defines was is and isn't copying, what is actionable copying and what is de minimus or ephemeral.
I'm not arguing whether or not this is copying. I"m pointing out that a seller of AI services could market their model as being trained on source material that is exclusively from outside the window that copyright currently applies to.
I don’t really understand the argument here anyway (although, as your typical commenter, I am reading this before actually reading the article).
Every day, I commit a huge number of non-criminal actions. So as long as I have some non-crimes, and I guess it doesn’t even need to be the majority under this standard, I’m fine to commit a bunch of crimes?
No, that is not the precedent set here. A legal claim was brought against the defendant, but (the judge found) that some of the evidence didn't have a specific property (copyright). Also the case was done in civil court, which has different rules and procedures than criminal court. This case is not about what one person did during the day.
That doesn’t seem to fit with the quote, though. The phrasing “not plausible that every training image used to train Stable Diffusion was copyrighted” makes it pretty explicit, I think, that what is protecting the data-set is the presence of non-copyright images, not the absence of copyright ones.
Actually, just to follow up, I think the crucial element here is:
> The judge has offered the plaintiffs an opportunity to amend and clarify their theory as to how Stable Diffusion operates its training data.
Since they are allowed to amend part of their claims, it seems more like the judge is just asking them to be a little more specific about which copyright works were included.
In short, the judge ruled that McKernan and Ortiz hadn't followed proper procedure for a lawsuit of this type, because they did not register their works with the U.S. Copyright Office before filing the lawsuit.
The terminology can be confusing, but "copyrighted" is often used interchangeably with "registered". Copyright in a work exists from the moment it's the work is created, but it's not "copyrighted" in the sense the judge is using until the work is registered with the Copyright Office. Registration is a requirement to sue for infringement. So a work is protected by copyright as soon as it's created, but it needs to be registered before one can sue for infringement. That can be done after the infringement occurs. For example, if a work is created in 2020 and someone infringes on the copyright beginning in 2021, the work can be registered in 2023 and the creator can sue for the past infringement (though not all of the same remedies are available).
The issue here is that the artists are making broad claims that DeviantArt infringed all of their works (by letting them be used as training data), even though they didn't register all of their works. Only some of the artists that are suing registered their works, and they may not have registered all of the works they are suing over. That's what the first sentence in the next paragraph is about: "[E]ven if plaintiffs narrow their allegations to limit them to Output Images that draw upon Training Images based upon copyrighted images..." The judge didn't misspeak here.
You do have recourse available, but registering the copyright is a required step in seeking that recourse, and must be done before suing.
> Title 17 U. S. C. §411(a) states that “no civil action for infringement of the copyright in any United States work shall be instituted until ... registration of the copyright claim has been made in accordance with this title.”
> ... however, a copyright owner can recover for infringement that occurred both before and after registration
It's possibly an interesting wrinkle given that, especially for certain types of works, the vast bulk of material is likely copyrighted but not registered.
Far more then that, it's an absolute show-stopper for any potential lawsuits, as it's really unlikely any potential litigant will be able to prove that even a large majority of the images have been registered, and thus eligible for a lawsuit.
All they have to do is register their works that they are claiming were specifically infringed, and they are fine.
Well, with the registration issue. They still have to prove us of an exclusive right and overcome any argument that the other side puts up that the use involved fell under fair use, but they'll have dealt with the threshold issue of “are we allowed to sue over the works we are alleging were used".
That is another hair to split, between "registered" at time of infringement and "registered" at time of lawsuit.
You need to register prior to bringing a claim. That isn't a big deal and can be done easily after any alleged infringement. But to get the attorney fees you need to have registered before the infringement occurred, something you cannot fix after you decision to sue.
Then to split hairs yet again, was the work registered at the time it was copied by the AI, or are you suing over infringement that occurred after a late registration? Are you claiming it was registered when the AI was trained on it, or are you claiming registration during the work's subsequent use by the AI after the training process? Maybe you can get attorney fees despite late registration if you can prove ongoing infringement after the pre-registration training process.
This might be a dumb question... I decide to try my hand at art. So, I take courses at the local college, where I'm exposed to copyrighted art. Anything I produce will, in some small way, be based on the totality of my life experience, including any copyrighted material I've encountered along the way?
How is training an AI any different than "training" a human? In the human case, isn't any court case really based on the output (my art looks like previous work), not the input (I saw some art somewhere)?
If a line appears on the page, it’s because that’s your intent. For an AI, when a line appears, that is always someone else’s intent.
AI can only copy.
I mean. Do we really need to go over all the reason why an art scanner that copies art and maybe paints it differently is fundamentally not the same as a human learning to create art?
That's just plain wrong. AI can generate things that are not in the training set, hence it can not "only copy" (on the other hand, I could understand trying to argue that AI can only create pastiche, or that it "uses" training examples in a fundamentally different way than we do).
Fundamentally, AI image creators stitch together images, then try to apply an art style.
You might call it “unique” because it stitches things together wrong (hello 10 fingers on one hand), but it is still fundamentally just copying, pasting, and then deforming lines together.
People claiming that AIs understand what they’re doing and that what AIs do is not fundamentally copying and pasting are the ones who have the burden of proof here.
>but it is still fundamentally just copying, pasting, and then deforming lines together
That's called "photobashing", very common to see used by concept artists. Also, it's fair use.
So now you've got a bigger problem in that almost all concept artists are engaged in theft I suppose? Since they are "only" copying, pasting, and then deforming things together?
Funny how you declare photo bashing as “fair use” but a cursory glance at the actual legality shows that you’re lying and that it is also in a somewhat ambiguous grey area at the moment, with lawsuits pending.
It is also massively frowned upon in art sharing circles.
>but a cursory glance at the actual legality shows that you’re lying
[[citation needed]]
>a somewhat ambiguous grey area at the moment
That's a funny way of admitting I'm right.
>It is also massively frowned upon in art sharing circles.
Who gives a shit? What part of "commonly used in concept art", and among professionals in drafting and conceptual stages in general, do you deliberately seek to not understand?
You mentally unwell or are you just an internet troll?
Ad hominem usually demonstrates you know that you’re wrong.
I am not the one arguing that a computer has thoughts, feelings and intent as you are here. It is you that is deliberately attempting to misunderstand through ignorance, lies, and insult.
Your “carte Blanche fair use” claim is nonsense. Copyright owners can, in fact, today stop derivative works that they know about, including photobashing.
By your own admission, AI is barely doing photobashing, so is very obviously not creating original works and is subject to copyright claims.
It gets tiring listening to HN AI enthusiasts pretending that AI is an actual comparison to human intelligence when it’s actually demonstrably a specialized search engine that does little more than copy and paste. If this were not the case, then:
-An AI would be able to feed itself without becoming shit tier nonsense
-An ai would be able to create materials not discernible from its training material
Neither case is true for any AI, including diffusion models.
Since you seem to not be aware of this: Fair use is an affirmative defense. By definition nothing I have written is therefore "carte blanche" as no defense of fair use can be ever said to be "carte blanche". This is pure sophistry on your part and transparently so.
>Here’s an article specifically on the legality of photo bashing
And mashups, photobashing, and many things likewise, remain legal and with fair use defenses in the courts as precedent. You continue to reach to any dishonesty or deflection including projection to avoid dealing with any of the points anyone has brought up.
That’s not a sound argument. AI doesn’t copy, but that doesn’t mean it must know what it’s doing in any human analogue of the “knowing” process.
That it doesn’t copy it’s very easy to prove: ask for something that is not in the training set. Does it generate it? Then it’s not a copy.
“Copy-pasting and deforming lines” is such a generic description (and to a certain extent a misrepresentation of what is actually happening) that I don’t see how it’s useful, or relevant. Even if we supposed this was true, you would be claiming that many of Duchamp’s works or potentially Arcimboldo’s or mosaics or collages are “just copies”.
If we could copyright every stroke in our paintings or every line in our artwork, nobody would be able to make anything new. Humans copy painting techniques, photography tricks, shot setups (angles, margins, etc) and make new pieces all the time that advance the world of art.
The intent is whatever the human puts into the prompt. If a human downloads an image, copies and pastes a solid black line thousands of times to make a mediocre drawing, is that not just copy and paste?
What is your actual intention here? Where does the "intent" in humans come from that AI doesn't, even when commanded by a human to perform something?
Where it comes from is that AI cannot create beyond its training scope.
If you ask an AI to train itself from its own images, it’ll always actually become worse using its own evaluation.
If you ask a human to train themselves, the human will notable become better of their evaluation.
AIs are fundamentally limited to their training set. Humans are not.
The fact that you can tell an AI to draw a specific thing, and it’ll steal content from other people doesn’t result in the AI having uniquely created an image. It has searched and blended strings together. That’s it.
Diffusion models are exposed to a blank canvas and then instructed to produce something. At no point are they looking into a database. They simply produce an image that will score highly in their aesthetic estimations.
Train an AI just on a few techniques. Then a couple fundamental stock photos with nothing but the subject. Then have it generate things outside of the realm of the stock photos recombined in a general style. Let us see your completely unique and unpredictable results.
Can you explain in more detail what you want? If you undertrain a diffusion model it will produce meaningless gibberish. Which is still 'creative', just not what you expected. Training for a diffusion model is not 'inserting' images into a dataset. Rather you should view it as a 'taste' alignment.
But it's not copy paste. At no point does it not 'create' something from thin air. With training it produces things along the lines of what you want. That's the difference. It doesn't copy. It just notices what you like, and generates that.
That doesn't make any sense to me as a line, as the human set up the machine to generate some way to generate images. The human still did "it", the "it" is just one step removed, and we've tons of things where some "it" is one step removed from the human where your line doesn't apply in law or people's concepts.
Furthermore, but even if a human creates something directly, they've still "trained" on all sorts of things that include copyrighted images--and, yes, if they create something too close to one particular copyrighted image, they can be found to have infringed copyright.
It's not a dumb question at all, that is one of the most central questions in this whole debate. I think there's a reasonable argument that humans add their own "stamp" on every artistic output (that isn't a direct 1:1 copy) which would make it a unique work even if closely inspired by another artist. A machine would not add its own "stamp," it's purely a mechanical derivation of others' works.
I don't know where I fall on the issue myself, I think there's many good arguments on both side, both in pure IP terms and economic ethics terms.
The thing with generative models is, it's hard to make them output anything from their training set verbatim. I don't know of a single case of that succeeding with SD or DALL-E or others. So it's not necessarily false to say that the models too always add their own "stamp".
Yes, but I'd counter that by saying the difference is in its reproducibility. Given the same inputs to the machine (AI model, [copyrighted] training data, prompt, seed), you will get out the exact same output. It's a mechanical derivation of the original works, and therefore, a copyright violation. On the other hand, humans do not work that way, you can't point humans at the same works and get identical outputs. I'm not 1000% sold on this interpretation, but I think it is a reasonable one.
A good counter to that is that the training data that was chosen as an input, and the prompt, are sufficiently original to count as their own work. Kind of like how an artist might use a copyrighted stamp to create their own work (Campbell soup cans).
And a good counter to that is that a translation of a book into another language is still considered a derivation of the original work, not its own work, despite requiring a whole lot of creative human input on the part of the translator.
>And a good counter to that is that a translation of a book into another language is still considered a derivation of the original work, not its own work, despite requiring a whole lot of creative human input on the part of the translator.
At the end of the day, it's all about legislation and case law--and judges making calls on the borderline cases, which creates new case law. It even plays into whether models themselves are copyrightable which, in turn, depends partly upon whether certain doctrines (like "sweat of the brow") apply under current copyright regime.
Can a different person copy the style of copyrighted work so closely that it violates the copyright of the original artist? I feel like this is a question that must be answered before we add AI into the discussion.
I think one of the problems is that it is fuzzy. If we can't draw the line with humans, AI is going to increase the scale of the problem. Now, instead of a few artists getting close to the fuzzy area of acceptable or not and being able to be judged partially on intention, AI allows for magnitudes more to get into the fuzzy area. Perhaps we were too lax in allowing the fuzzy to remain fuzzy due to such a low number of cases, but now technology has increased the intrusions into this fuzzy space far faster than the law can keep up.
It’s a good question and I do not believe a style has ever been found to be copyrightable, only actual works.
Styles would seem to be more in the trade dress / design patent side of IP law. Apple’s famous 22-element iPhone design patent, etc.
But it’s hard to imagine an artist reducing their style to clearly defined elements and applying for protection. “A geometric grid or squares and rectangles with white background and black lines extending entirely across the canvas horizontally and vertically. Some grid rectangles are filled in primary colors. The square intersections of two black grid lines may be filled in primary colors. At least some of the grid lines are spaced irregularly”
So your argument boils down to the idea that humans impart a magical, invisible mark on their own artworks, and though no one can measure this mark nor see it, the fact that it exists within your own imagination exempts them from overburdensome rules that you would mercilessly apply to one of the most exciting innovations of not only the 21st century, but even the 20th century before it?
(I emphasize again, that I don't know where I stand on the issue. But for the sake of argument...) Well, I think my argument is that the mark is not invisible or undetectable. Simply by nature of being human, we impart something of ourselves onto things we create. For example, maybe by how you choose to compose the objects within a work, in a way the artist whose style you are mimicking would not have done, or the context in which you present it. Or maybe in how you hold the tools, where it creates an unintentional but different effect in the work. Whereas with an AI model, it will always create the same output for the given inputs.
> Well, I think my argument is that the mark is not invisible or undetectable. Simply by nature of being human, we impart something of ourselves onto things we create.
You're one of those people who think that if you can imagine something and then describe it poetically, that this has any relevance to the real world that the rest of us inhabit.
I think the detail that makes it tricky is that, in theory, an AI could just have a single neuron and simply output verbatim the image it was trained on. In which case we obviously wouldn't say that the image it outputs is a new work and freed from copyright.
So clearly there is some threshold of "transformativeness" which an AI must possess in order to be trained on one image and produce a different one that is considered a new work. Where exactly that threshold lies is difficult to define from a legal perspective.
That sounds a lot like music copyright, which also has a subjective threshold. It’s kind of a mess but it’s more or less worked for a century. Certainly better than any formula for rhythmic, melodic, and harmonic similarity would.
You can argue that artists given an implicit license for people to copy art into their own brains while viewing it. The same way websites give an implicit license for a browser to copy the page to display it. That does not necessarily mean they given an implicit license for someone else to copy the work as part of AI training.
Separate to that is the issue of producing new works that are too similar to existing works you've seen. If you do that as a human then you're liable for copyright infringement.
I think you’re proposing two big expansions to copyright: 1) that it covers what an observer can do, and 2) that everything is illegal unless explicitly made legal.
I'm proposing nothing, as far as I know that is how it is now. A browser viewing a web page is making a copy of the web page which is governed by copyright. Except by making a web page you give people the right to make that copy. Lots of interesting lawsuits around that regarding bots, scrapping, etc.
Copyright has always been about the reproduction of works. Copyright lawsuits about scraping, etc, are about the use of scraped data. That’s why EULAs, which are not copyright, are often used to try to further restrict rights.
None of that covers the idea that observing a copyrighted work is a violation of copyright only made acceptable by an implicit license. As the link you posted notes, that’s not what an implicit license is.
See this kind of thing here is what I think is a problem and has been a problem forever. We have two opposite laws on this,
1. Copying is theft (contents of a CD [software], movies, stuff on a website sometimes, etc)
2. Copying is not theft (VHS copies of broadcasts, cassettes, etc)
Problem is "implied license" (eula-roofie) where we've, without any good reason I think, decided that,
(a) Copying is theft when someone unilaterally declares what you buy to be different from the medium (CD vs what's on it [software, movie, etc])
(a)(2) except when it's a really old medium like a book which you're allowed to quote, and use in derivation, but don't you dare "quote a movie" by copying portions off that DVD or blu-ray because that's different. Shut up is why.
(b) Copying is theft when someone unilaterally declares their putting something up to view, which necessarily requires copying to view (sent over internet), can't be copied for derivative fair use
(b)(2) except when it's over another old medium like public broadcast because... fuck you? I haven't a clue myself how this makes any sense and I suspect it is because "it doesn't".
(c) Copying is theft if you take the copy sent you over the internet and do anything with it somebody unilaterally declares they don't want you to be able to do
(c)(2) except if it's an old medium, like a book, where you can borrow it and quote from it and do all the fair use stuff, or borrow a picture, or a painting, or...
All the fair use and copying and other sensible stuff without this weird "implied license" stuff doesn't exist for formats we've had before to the extent it does now. These formats, in principle, are still just "you get a copy of a thing on some medium", and yet we have two completely different laws based on this pure fiction of an "implied license" that declares you can't copy even portions of one on some format because... because.
I am trying, desperately, to explain there is fundamentally and in principle no difference. The root of all of our problems and why the law doesn't make any sense is the pure fiction that there IS a difference. And I feel like a goddamn madman yelling in the streets trying to explain something that feels so dog gon obvious to me but seems like is obvious to nobody else.
> You can argue that artists given an implicit license for people to copy art into their own brains while viewing it.
You can, but it would be more correct to argue that memories in a human brain are not copies under the law at all, and so need no license (or fair use analysis) in the first place. This has different results than an implicit license argument, especially when there is an explicit denial of license or an explicit license which is also explicitly bounded to exclude any rights not explicitly licensed.
This is the crux of the issue. If courts find that AI draws like a human (at least for the purposes of copyright law), then you can only claim it infringed upon your work if it actually looks like your work was de-facto "traced"/copied by the AI and the output image only has small imperfections that don't qualify for fair use.
Copyright law does not care about how an infringing work was produced. It can be traced, or memorized and recreated, or described to a confederate over the phone, or whatever. It is only the actual work that is judged as infringing or not.
You are not missing anything: you are asking the key question in that process. We don’t have an answer.
As a society, we agree that you, a human, copying art you’ve seen is original (or at least derivative enough that it warrants a new copyright).
We agree that a photocopy machine, or rather a human operating one, is copying art in a way that is not creative enough to warrant a copyright — so the human goes to jail.
As a society, we have not agreed upon how transformative a “work” from a generative AI has to be to count as original, or even if it can. A recent decision claimed that no machine work without human effort would count — ignoring the effort to find good prompts, pick an image and edit it further. An AI is different because no much was “trained” to make art before AIs: some animals were (elephants famously), and who owns their work is tricky — there’s a case of a photo, a selfie actually, taken by a macaque that is really interesting in that regard.
My recommendation would be to have a similar system to what we have for music (and I think plays) where there are two or three “authors”: the person who wrote the score, optionally the person who wrote the lyrics, and the interpret: the singer. It gets really fun when you have orchestras, animals involved in making music, or sound-sampling like in RnB.
We could have original human authors and AI adaptors, each paid based on use. If I like your work, I could commission OpenAI to make “a portrait in the style of alistairSH” based on your DeviantArt work, but of my mother (whom you’ve never met), using photos that I’ve uploaded of her. I can pay OpenAI, who would pay you. Style vs. theme could be separated with current AI technology, although that could get just as complicated as every intellectual property process very fast.
If I build a Rube Goldberg painting machine and somebody else "pulls the trigger" to start the machine, who owns the output? Me or the other person? In the case of today's AI models, I could see arguing that the creator of the AI or the creator of the prompt have some rights to a specific output, but not the AI itself.
The monkey case seems fairly cut-and-dry to me - the monkey didn't steal the camera (as often reported) - the photographer engineered a situation to get the result (he set up the camera, the tripod, and held onto the tripod so the monkey wouldn't steal the whole rig). It's his art. If the monkey actually stole the camera, then it's nobody's art - just a random occurrence.
If an artist orchestrates a group of humans to create art, the art belongs to the artist, not the group (as far as I know - I'd love to hear otherwise).
I guess for me, given today's technology, I don't see an "AI" as any different than a Rube Goldberg machine hurling paint at a wall. A human had to create the AI and a human had to prompt the AI, one or both of them own the art. And in all cases, the art is always based on something the human saw previously. No art exists in a vacuum.
And I agree that a system like music rights might be appropriate. The AI author, the AI prompter, and probably anybody involved in the training data, all have some right to the output. Then the question is what % to each, but that's solvable.
The best equivalent in this case would be the engineers working on Photoshop for Adobe: they write Rube-Goldbergian code to detect shapes, colors, and let artists click once to detour, cast shadows, adjust palette, etc.
Excellent point regarding the famous Monkey selfie, which I think points to the emerging issue of biological vs artificial agents.
It presents a sort of mirrored view of the GenAI issue: The human photographer does not get credit because the monkey is deemed to be the "agent of intent" rather than him, presumably because it's a living thing. If a photograph was triggered instead by a human tossing a rock onto the shutter, no one would claim that the rock now holds the copyright!
Interesting thought experiment to replace the monkey with other "organic agents" and some method for them to trigger a photo: A squirrel? A lizard? A cockroach? A slime mold? At what point is the organism so "simple" that you say it was just part of the system/mechanism?
Then on the artificial side, to replace the monkey with different mechanisms (as was your point): A photocopier seems quite different to a GenAI model. At what point does the human who set up the situation stop being the "agent"?
Yeah, I am equally confused and I fear the law is going to royally bungle this one and we'll be stuck with something really stupid for some arbitrary number of decades or generations.
It just seems obvious to me that "human made thing to transform some set of things into a composite+transformation of those things" is fair use if fair use is to make any sense at all. Since analogously, like you said, in the same way you are taking your experience including experience of copyrighted stuff as a set of things you are drawing upon producing something of your own. All art is derivative.
It makes zero sense to me at all that it's suddenly a problem that art is derived if, instead of artist directly painting something, artist sets up some device that paints something. In both cases "human takes inputs and produces something unique" (not a copy, not a scanner, in case that was not clear from context).
Unless I'm missing something too, but based on what I see so far I just feel like I'm from mars and don't belong here.
> It makes zero sense to me at all that it's suddenly a problem that art is derived if, instead of artist directly painting something, artist sets up some device that paints something.
I think a big difference in how this tech makes people feel is the amount of effort required. It's a problem we see with technology in other areas, too: it's not (much of) a problem if a cop sits outside a suspect's house for a few hours to monitor their behavior. Most people think it would be a problem if the cops pointed a camera at the house 24/7 for months. Pretty much everyone agrees canvassing society with cameras and creating a public space surveillance panopticon is bad. Where's the line?
Similarly, if some artists spend years learning the craft and mimicking a couple of styles with a relatively small output, it's not a huge problem. The scope is small. But if everyone on earth can do it to anyone on earth at massive scale just by typing "in the style of ..." to an AI, does that start to be a problem?
The line is pretty simple for me, because it's the same line that applies to people. It doesn't matter what the training set is so long as it is not reproducing either the same thing or too similar to the thing it's trained from. Somebody else linked that you can't generally copyright style for instance, and if I recall you can't copyright algorithms or things of general knowledge either. So you can copyright "a specific instance of a thing", but not "the general idea of a thing". Here, near as I can tell, these algorithms generate some general idea of the things it's trained on and produce something specific different from the specific things. What should matter is only if it is different enough, same as it matters for people.
We've got much bigger problems though but if we, we as people generally not specifically you, can't even agree on something I see as so fundamentally basic bringing up things I think are problems would start war
> so long as it is not [...] too similar to the thing it's trained from
Aye, there's the rub ;) How do you define "too similar"? If I invert the colors of Munch's The Scream, is that an original work? What if I pass it through a computer program I wrote to swirl it around in a spiral? I think those are clearly derivative works--you put the image in as input, put it through a mechanical transformation process, and get the same output for the given input. That description also applies to AI.
> these algorithms generate some general idea of the things it's trained on
Well, be careful here. AIs don't generate ideas. They take inputs, do some mechanical work on them, and output something derived from the inputs. There are no "ideas" involved here, it's a(n extremely complicated) mechanical transformation.
(To repeat myself, I don't know where I stand on the issue. I think there's good arguments on both sides.)
>Aye, there's the rub ;) How do you define "too similar"?
... Same ways I do for humans. That's why I wrote "the same as for people" that line was supposed to contextualize everything else I wrote and somehow it hasn't.
edited to fix a screwup. because somehow I swear I hit "copy" but it didn't copy
When things seem so obvious and simple, it can be a good mental exercise to try to put yourself in the shoes of the "other side" for whom it also seems obvious and simple.
While never my main thing, I've worked as a professional artist and know people for whom art is their livelihood. The holy grail in that world is to create a unique style that will command a premium: an art director decides your style is perfect for their new campaign, a building designer decides your style fits the lobby of the new building, etc. This style is the result of years and years of refinement, false starts, watching trends, etc. And this style is why they get paid, and how they feed their family.
When you tell such an artist that now any schmuck can create art in their style just by writing "In the style of..." you should understand that —especially to a non-technical person—the end result is just a slight deviation of from a copy machine. To them it seems "pretty simple" and "obvious" that this is just a fancy way of stealing.
(And I'm not taking sides here, just saying that step one is to realize that the issue is not simple, and is thus super interesting.)
>When things seem so obvious and simple, it can be a good mental exercise to try to put yourself in the shoes of the "other side" for whom it also seems obvious and simple.
I would if I could but as I keep tryin to explain I can't because it requires I believe something that's a contradiction.
Just because it's a contradiction that makes ya more money don't mean it isn't a contradiction. Also I am that non-technical person.
You're not missing anything. When humans do this it's fair use.
The question is... Is fair use an exclusively human right? For example, we have many rights that do not extend to our possessions. Animals have no protected right to life or liberty. The human rights enjoyed by their owner do not flow to them.
Similarly, just because a human has a right to fair use, do AIs?
Put another way, if I create a Rube Goldberg painting machine, what copyright law applies? Why is that different than if I create/train/prompt an AI to do the work? One is physical, one is digital, but both require a human to create and initiate. And in both cases, every decision that human makes is made based on their life's experiences. No art exists in a vacuum. All art is derivative.
Nah, Altman doesn't really want you think machines are like humans and have rights either, because if the law thought that, it would also think Altman was aspiring to be the greatest slaver in all of history.
The difference in training an AI compared to training a human is that an AI is not a human. Law makes differences between humans and non-humans all the time. AI drivers vs human drivers is a prime example.
If the question is in how biologically we learn compared to and AI, the answer is that humans don't learn by calculating math in order to determine how pixels relate to each other. Humans uses chemical signals between neurons and neuron activiation among other things in order to approximate the concept of what we consider to be learning.
I find a similar and equal important question when asking whats the difference between an AI learning and a human learning, is to ask whats the difference when a computer remembers information by storing it on a file compared to a human remembering information by storing it in long term memory.
Copyright is new age nonsense that denies the chain of influence and inspiration that exists between all people. If you can create something sell it. If it can be easily recreated (or stored digitally) then it isn't yours.
That doesn't hold though, especially the part about stored digitally as then every movie and TV show that is streamed would no longer be copyrighted.
As for easily recreated, it is not hard for anyone to make a wooden block and then use it to make prints but if you made a wooden block print of Campbells soup cans you would probably face a copyright claim from Andy Warhol depending on the "look" of the print.
If we go that route, then doesn't that remove almost all financial incentive to produce new content that could be digitally stored / copied / recreated?
Because as soon as you create it and try to sell it for $1, someone else will recreate it instantly and put it up for $0.50, and so on until the value of all non-physical works is effectively $0 the moment after creation.
Feels like that would result in way less human art being made.
If you create a painting someone else reproduces instantly with ease then it would imply the quality was not high enough to command the price. If incentive falls then supply would fall and either find a new equilibrium price or demand would remain and prices would rise.
Seems more like quality would matter. Hard copies would matter; perhaps the disc you issue your media on is rated to last 100 years, whereas someone's digital copy can be deleted with a keystroke.
Copyright is bullshit in a market of perfect reproducibility.
The copies are non-scarce, but the production effort is indeed scarce. That's why alternatives to copyright tend to focus more on funding projects before they are released, such as patronage, commissions and public funding [1]. There's also the comically named "hostage funding" where you have a finished project, but you only release it after a certain funding goal is met.
It's worth noting that copyright is rather destructive since it impairs people from building and innovating on previous works [2], so any benefits need to be weighed against that.
This is really good news, I support artists but these short sighted individuals if successful in these claims will just build a world that harms artists as these tools wont cease to exist they'll just be put behind the paywalls of some of the richest companies in the world which then will be able to change what they want.
Like look how expensive AutoCAD is, why wouldn't Adobe charge that sort of money or more to access image gen if them and 2 other companies max can possibly provide it.
Sorry but at least with open models we all can benefit, you download it once and its FREE FOREVER you only pay for power used. It's petty and short sighted to try and take that away under the guise of "ethics".
"Judge Orrick writes that it is “unclear” as to whether Stable Diffusion holds “compressed copies” of the images and points to the defense’s argument that the training dataset, which contains five billion images, can “not possibly be compressed into an active program.”"
Five billion images in compressed lossy format can easily fit into an active program. The only variable is quality. Five billion images in the original format and quality is a different matter, but then that requirement for copyrightable would be drastically insufficient.
Assuming stable diffusion is ten gigabytes, that's 2 bytes per image. 2 bytes would be too small for a short textual description of the image, let alone a scaled copy of the image.
Assuming they don't need to use all Five billion images, the bytes per image will be larger. They can also use composites and filters to remove duplicates and images similar enough, or generative series.
One do not need 2 bytes per images if one million of images are blank, even if they are uniquely of different sizes. The amount of bytes to represent one million blank images in code is very few. One only need to store the unique aspects of the images, and if one compresses the images then the range of uniqueness drops based on the quality required.
Very unique images will naturally have larger footprint, based on the quality that one want to retain.
146 comments
[ 4.1 ms ] story [ 233 ms ] threadLots of data they trained on is available for purchase (e.g. artists often sell prints or reproduction rights, the books in books3 are widely available, etc). It's my understanding that companies like Stability and OpenAI did not attempt to determine if the data they trained on was available for purchase and then buy a legally purchased copy for training. That might cause them to run afoul of fair use doctrine in the US (not sure of other jurisdictions).
See these excerpts describing fair use for copying library materials [1] (many of these collections are being released by groups referring to themselves as libraries):
> Copying a complete work from the library collection is prohibited unless the work is not available at a “fair price.” This is generally the case when the work is out of print and used copies are not available at a reasonable price. If a work, located within the library’s collection, is available at a reasonable price, the library may reproduce one article or other contribution to a copyrighted collection or periodical issue, or a small part of any other copyrighted work, for example, a chapter from a book. This right to copy does not apply if the library is aware that the copying of a work (available at a fair price) is systematic. For example, if 30 different members of one class are requesting a copy of the same article, the library has reason to believe that the instructor is trying to avoid seeking permission for 30 copies.
> The copying, whether performed by the library or whether unsupervised by the library patron, cannot be for a commercial advantage. This means that the library (or a copying service hired by the library) cannot profit from the copying. In addition, the copying for the patron must be done for purposes of private study, scholarship, or research.
[1]: https://fairuse.stanford.edu/overview/academic-and-education...
That quote from Stanford's library is not discussing fair use doctrine in general, but rather is stating what is permitted in those specific circumstances. There are plenty of instances of fair use where the underlying work used was available at a fair price. That's the whole point of fair use law: some use of a work that is facially infringement escapes liability because the particular use is considered fair.
Tail end of the article:
>However, the judge allowed Andersen’s complaint that her 16 copyrighted works were used without her authorization to move forward.
I'm not loving this headline.
(edit: stamped comment w/ the current headline to be safe)
Again, I don't remember what the exact description was, but I remember it jumping out as obviously wrong (a layman's interpretation of a neural network). Feel free to go over that and it should be pretty obvious how the claims could be amended. Barring that, just wait for the filing because I'm sure the plaintiff lawyers will be jumping all over this.
I get the emotional appeal to companies are using my stuff without paying me or even crediting me. But, as I understand it, works are normally considered derivative of specific works—not some large corpus.
The argument as it went went like this. Is the theory that no infringement has occurred believable, or can it be said as a forgone conclusion that over the course of the operation that some specific copyrighted work has been infringed on at some point in time and that the operator knowingly were aware that such cases was likely to have happened.
It is true that many lawyers and legal experts thought that the pirate founders were immune to such claims since usually one need to produce specific cases of specific copyrighted works in order to find someone guilty of assisting. The case however illustrated that such requirements are not always needed, and the Swedish supreme court did not feel it necessary to analyze it further.
Though in this case, I assume that a great many lawyers would argue that it's hardly a foregone conclusion in this case.
AI developers has also started to add filters and other techniques to remove outputs that are too similar to existing works, which is both a good thing and a bad thing. It shows that the tools do output such works (as otherwise they wouldn't need to filter it out), but also show that they are working to minimize it. Courts would have to look at it and decide if such efforts are an admittance of the issue or if the mitigations are enough.
In this specific case, I doubt such discussion will occur at this point since the main point that the judge brought up is that the copyrighted works need to be registered first at the copyright office before it can go any further (except for 12 works which already are registered).
With the "edge computing in AI" push (offline processing on phones and the like), we might see soon enough.
Judge is also letting the 'unauthorized use' charges move ahead.
To me, this feels like the right way to make sure the questions at the heart of the trial are satisfied by the results of the trial. SOP, for sure, but it definitely could have gone the other way, with the judge just throwing it out altogether, so I'm glad this is moving in this direction, instead.
> The three artists — Sarah Andersen, Kelly McKernan, and Karla Ortiz — immediately ran into problems as two of them — McKernan and Ortiz — did not register their works with the U.S. Copyright Office.
Two of them got bounced, Anderson remains.
> Judge Orrick writes that it is “unclear” as to whether Stable Diffusion holds “compressed copies” of the images and points to the defense’s argument that the training dataset, which contains five billion images, can “not possibly be compressed into an active program.”
Once again, regulation that's needed is complicated by the fact that people in power do not understand how this shit works.
However:
> The judge has offered the plaintiffs an opportunity to amend and clarify their theory as to how Stable Diffusion operates its training data.
And:
> However, the judge allowed Andersen’s complaint that her 16 copyrighted works were used without her authorization to move forward.
So the case is still developing.
However, this is frustrating:
> The problem for the artists is that the training data for these programs is a black box. Outside of LAION, very little is known about what exactly went into training AI image generators but it is widely assumed that the companies did an almighty scrape of images on the internet which included taking copyrighted and copyrightable pictures.
I'm not sure if the publication stated "widely assumed" here because of legal reasons or because the author doesn't understand this either, but the entire crux of Anderson's case here is that if you include the proper prompts, you get comics out of these generators that strongly resemble hers, in both style and content, which are copywritten. That an generator can do that basically requires that a good amount, and I mean a LOT of her comics were pushed into the training data, and not only that, but were tagged explicitly as her work. Given how much of LAION is scraped from social media, this is not surprising.
In order for the generator to know what Sarah Anderson's style looks like, her comics must be in the data set. Period, paragraph.
My read on the case is that it happened, John made a big public outrage post, and then gradually after news interest died down, more facts were revealed that made it much more obvious that oculus was in the wrong but nobody cared anymore.
It's also not relevant to making an ai model. That is at best relevant to producing a similar output using an ai model
This an entirely new schema and the endless slap fights between people trying to say "well the computer is just learning like a person does" is just, bafflingly, wildly wrong headed. No it's not. A person learns so they can accomplish a goal, and an ML algorithm by it's nature wants nothing, it is incapable of want, versus the people who are saying "well ML lets people copy other people" which is also wrong headed, because it isn't copies, it's remixes and re-imaginings of similar material are just not productive. On the one side people trivialize the act of learning how to art being just looking at shit and reproducing it, which is just wrong, completely wrong, and on the other side people think it lets the ML model copy artists, which it dooooooes, kind of? But that test also fails because they aren't strict copies and them being copies isn't the goddamn problem. The problem is the style, the vibe, the unique look of a given creative is trivialized into something you can get from stable diffusion after a few tries with the right prompt. This isn't a problem because Sarah Anderson lost a sale or whatever the fuck. It's a problem because it completely undermines the notion of the output of paid creatives having value, which is already a shaky proposition for most people and why tons and tons of prominent creatives cannot earn a living via their creations. This is going to make that situation even worse and no matter how many times this is explained to people, they just run back to this "well I wasn't going to buy things from them anyway" and the entire thing makes me want to fucking scream.
All that to say: AI is not AI, it's ML. ML is not learning like a person does, it's fundamentally different, stop equivocating. If you're so determined to cheer on as silicon valley knocks already struggling creative professions down a hill with systems that COULD NOT POSSIBLY EXIST without their pre-existing output to exploit, without permission, fundamentally unethically, if you just don't give a shit about that and are going to gleefully cheer that process on, then all I really ask is you be honest about that. Just say "I know this is fundamentally corrosive, unethical, and undermines an entire segment of the economy, but now I can get as many over-tuned generic shitty pieces of art as I want, and that's just more important to me." Just say it, and stand by it.
I don't find the rest of this readable tbh
I have ZERO issue with image generators on their own. They're quite cool technology. The only thing I and tons and tons of other people have requested is they not use images without permission to train them, that's literally it. That's my sole, single, solitary objection and yet every AI interested person, advocate, and company rep will do intellectual back-flips to explain why this is unreasonable.
It's not unreasonable to ask permission before you use something. It just isn't.
It's also not unreasonable to look at and learn from things that people have put on public display.
But if I "copy" the experiential knowledge of your art into my brain by viewing it, I'm not violating your copyright. My brain doesn't contain a copy of the art, it's just been influenced by viewing it, and I might be more capable of producing art that mimics your style.
What these models are doing feels, to me, vastly more like the second case.
In this way, ML training is just not the same as a person being inspired by or even being asked to recreate another person's creative work. Over the course of making something, an artists' "voice" would be best characterized I feel as the tiny choices they make along the way that all point to and reinforce a larger point or purpose to the piece. This "voice" shows up in all creative output, not just spoken word. That is what I feel people are feeling is lacking in generated art: because a machine-learning model does not have a voice, it does not have intent, it has a mandate from a third party from which it tries to draw from, and instead of making numerous, tiny but contributory choices, it instead decides on a weighted average of all the choices made in the art that the model was trained upon, which is simply not the same thing.
It makes all generated art have this very sterile, soulless feeling to it because these tiny choices that would otherwise be made by a person trying to illicit an effect are instead just the machine sort of shrugging and being like "well in most things I've seen where a woman is sitting this way, her hand is tilted this way" but it doesn't know why the hand is tilted or what that means for the subject, which means the hand-tilt might be applied to subjects for whom it makes absolutely no sense at all to tilt the hand.
I am capable of doing this.
> If you show someone a painting and ask them to recreate it, even setting aside the skill gap, you do not get the same painting back. You get a different painting, with some overlap, with the "focus" of it being what the person paid the most attention to during their viewing.
same for ai models. in fact, more true for ai models. You'll have a much harder time recreating an image from popular models than from human memory.
> It makes all generated art have this very sterile, soulless feeling to it because these tiny choices that would otherwise be made by a person trying to illicit an effect are instead just the machine sort of shrugging and being like "well in most things I've seen where a woman is sitting this way, her hand is tilted this way" but it doesn't know why the hand is tilted or what that means for the subject, which means the hand-tilt might be applied to subjects for whom it makes absolutely no sense at all to tilt the hand.
this will surely age well
People are already experimenting and/or deploying systems that run an LLM before doing text2image [1]. This helps a lot with 'understanding'.
The next low-hanging fruit would then be to go back and improve the labeling used for training the image-generation models in the first place (using new multi-modal LLMs).
[1] eg. the current GPT+ standard model will do Prompt -> GPT-4 -> DALL-E
AI companies argue (long story short) that their usage is closest to taking statistical data, and thus doesn't trigger the need for a license.
I do realize that some people find this to be a somewhat unfair outcome as is, but
A) It is the current law as is. (caveat: we still have to see if the courts actually confirm this particular interpretation)
B) I'm not actually sure we should let those people have their way in the first place.
Of course, ignorance of the law is no excuse.
I feel like some creative folks never actually bothered to learn the laws they were working under, or (more charitably) this turned out to be a corner case, and they were caught by surprise.
It can be unreasonable if it makes it harder to society in general to reuse and innovate on previous works. Intellectual property in general can be quite destructive. Heck, take a look at Permission Culture, which is actually a negative thing:
Source: https://en.wikipedia.org/wiki/Permission_cultureThe "style" certainly isn't copywritten because you can't copyright a style. Someone creating doodles in her style is absolutely allowed.
She doesn’t own black and white scribbles and striped clothes.
Especially when really she’s just deriving from Tim Burton and other similar styles.
> However, the judge allowed Andersen’s complaint that her 16 copyrighted works were used without her authorization to move forward.
So there is that and Stability is already having another lawsuit with Getty with obvious watermarked images as well.
This is where I think the judge misspoke. "Copyrighted" does not mean registered. Copyrights attach at creation and does not require anything of the creator. What the judge is referencing is the fact that registered copyrighted works are given additional protections, notably the award of attorney fees in disputes. But the fact that a work was not registered, what the judge incorrectly describes as "copyrighted", does not mean it is outside copyright. All the same rules apply governing how the material may be used or not used. You just don't get as money when you sue over a non-registered work.
I would hazard that the vast majority of copyrighted works are not registered. Every holiday photo taken is copyrighted. Every word typed on HN is copyrighted. Virtually none are registered. That doesn't mean such things are public domain, free for use by whomever/whatever runs across them online.
Aren’t there visual art in the public domain?
INAL but the judge’s ruling seems technically correct, so what am I missing?
Yes, photos taken by people working for the government for example, and anything shared with Wikipedia should have a license for free use (typically CC0) [0] That’s an important, valuable part of all images, but a very small segment overall.
The vast majority of images (holiday photos, posts on social media, Deviant Art, etc.) are not in the public domain, even if they are online.
[0] https://en.wikipedia.org/wiki/Wikipedia:Image_use_policy#:~:....
This is a huge leap that would change copyright law dramatically, even for humans.
Youre asserting that copyright doesn’t just protect from reproduction, but from use as an example to learn from.
I think this would break copyright altogether. What musician or artist can say a work is 100% original and does not incorporate any details or concepts from work they’ve seen / listened to / read?
I know people hold that position, but it’s important to be transparent that applying copyright to training/learning (in addition to output) would be an overhaul of copyrigh, not a clear and obvious application of the law as it stands.
Which, I don’t think that’s an inherently absurd statement. I mean, at some point humans physically exist, so a sufficiently complex artificial being could exactly replicate our physically processes, and it would be hard (imo) to argue that such a thing isn’t a person.
But I don’t see the non-big-leap option. Either we admit that machines can actually perform creative works which are their own ideas, or we call the thing they do a derivative work, and we admit that we can’t just eyeball the difference between an original work and a derivative work anymore. That there’s something intrinsically special about having a human in the loop.
Why do you think humans can do that, and that what we call “original works” of humans are not just mechanical transformations of the human’s inputs.
How, in a universe of deterministic physical laws, is it possible for a human to do anything else?
This presumes that an input having influence on an output means that the output qualifies as a derivative work of that input. A work being in some way derived from another work does not mean that it is a "derivative work" for the purposes of copyright law. Many artists want to make this presumption, but IMO it's extremely dubious as a general claim. It is likely that when this all shakes out, specific works will be able to be established to be derivative works of specific other works, but simply being an input will not cause that. (caveat: IANAL, but I do have a law degree and IP experience)
That would be like requiring human artists to enumerate exactly every single influence on each piece of art they make. Did they see a painting by so and so at six years old that had a 0.1% influence on them? Better include it too.
This is just an attempt at a landgrab by idiot copyright maximalists.
But I do suspect that there are two overlapping schools among the most vocal supporters of this technology.
- Those who think that generative AI is so wonderful because they have a vested interest and/or have bought into the most breathless hype
- Those who more or less think copyright shouldn't exist in any case
I still believe a little bit in Jefferson's vision. It's just that that has so little to do with copyright that we're probably not even talking about the same thing. He envisioned copyright as being an incentive to create and do new things, to learn new things.
And here, as it would be used, it would shut down new technologies and quash new ideas. Clearly, such people are doing copyright wrong. I am not religious, I do not believe in "sin", but if that word did have any real use, it might be strong enough to describe those who would misuse copyright.
In any event, Jefferson was granting a very limited privilege and not describing a fundamental human right. This privilege was granted by the people, and the people do have the right to revoke that privilege if it continues to be misused. This, I think, is all the warning the misusers deserve. A comment that will be read by hardly anyone, that they will surely never see, and which will give them no headsup at all.
In the way in which you literally must make a copy of the dataset to use it in training, and that literal copy if not excused by fair use analysis including consideration of the downstream purpose, is a violation of copyright.
AI companies argue that insofar as they are copying, they are only making ephemeral technical copies.
US fair use analysis includes purpose, so copying which would be fair use for one purpose can cease to be when the purpose changes.
Since this is a case under US law, and the US is not currently a member of the EU, EU law is somewhat tangential.
- In the context of specifically training an AI-model under EU law: -
Article 5(1) of the Directive 2001/29/EC [1] is argued to apply. Of course, if sued, AI companies still need to show that their use is otherwise lawful (which should be fairly easy, since they actually do not retain anything remotely resembling a copy at all)
- In the context of specifically training an AI-model under US law: -
The purpose of the copy can be argued to be non-exploitative, incidental, for the purpose of enabling technology, and temporary.
By contrast: Google Books even argued that permanently retaining entire copies of books wholesale was fair use [2], provided they didn't provide copies of those books to 3rd parties.
OpenAI argues that there's no way they're doing anything even remotely close to that. [3]
Disclaimer: IANAL, YMMV.
[1] https://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=CELEX...
[2] https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....
[3] https://www.uspto.gov/sites/default/files/documents/OpenAI_R...
See also:
[4] https://news.ycombinator.com/item?id=37879938 Previous thread where I looked up a bunch of sources.
Every day, I commit a huge number of non-criminal actions. So as long as I have some non-crimes, and I guess it doesn’t even need to be the majority under this standard, I’m fine to commit a bunch of crimes?
> The judge has offered the plaintiffs an opportunity to amend and clarify their theory as to how Stable Diffusion operates its training data.
Since they are allowed to amend part of their claims, it seems more like the judge is just asking them to be a little more specific about which copyright works were included.
The issue here is that the artists are making broad claims that DeviantArt infringed all of their works (by letting them be used as training data), even though they didn't register all of their works. Only some of the artists that are suing registered their works, and they may not have registered all of the works they are suing over. That's what the first sentence in the next paragraph is about: "[E]ven if plaintiffs narrow their allegations to limit them to Output Images that draw upon Training Images based upon copyrighted images..." The judge didn't misspeak here.
> registered before one can sue for infringement.
It's a little unclear quite what it means for a work to be protected if you don't actually have any recourse when infringement occurs.
> Title 17 U. S. C. §411(a) states that “no civil action for infringement of the copyright in any United States work shall be instituted until ... registration of the copyright claim has been made in accordance with this title.”
> ... however, a copyright owner can recover for infringement that occurred both before and after registration
https://www.supremecourt.gov/opinions/18pdf/17-571_e29f.pdf
Far more then that, it's an absolute show-stopper for any potential lawsuits, as it's really unlikely any potential litigant will be able to prove that even a large majority of the images have been registered, and thus eligible for a lawsuit.
Well, with the registration issue. They still have to prove us of an exclusive right and overcome any argument that the other side puts up that the use involved fell under fair use, but they'll have dealt with the threshold issue of “are we allowed to sue over the works we are alleging were used".
Based on what legal reasoning?
You need to register prior to bringing a claim. That isn't a big deal and can be done easily after any alleged infringement. But to get the attorney fees you need to have registered before the infringement occurred, something you cannot fix after you decision to sue.
Then to split hairs yet again, was the work registered at the time it was copied by the AI, or are you suing over infringement that occurred after a late registration? Are you claiming it was registered when the AI was trained on it, or are you claiming registration during the work's subsequent use by the AI after the training process? Maybe you can get attorney fees despite late registration if you can prove ongoing infringement after the pre-registration training process.
How is training an AI any different than "training" a human? In the human case, isn't any court case really based on the output (my art looks like previous work), not the input (I saw some art somewhere)?
I feel like I'm missing something (obv IANAL).
If a line appears on the page, it’s because that’s your intent. For an AI, when a line appears, that is always someone else’s intent.
AI can only copy.
I mean. Do we really need to go over all the reason why an art scanner that copies art and maybe paints it differently is fundamentally not the same as a human learning to create art?
That's just plain wrong. AI can generate things that are not in the training set, hence it can not "only copy" (on the other hand, I could understand trying to argue that AI can only create pastiche, or that it "uses" training examples in a fundamentally different way than we do).
Fundamentally, AI image creators stitch together images, then try to apply an art style.
You might call it “unique” because it stitches things together wrong (hello 10 fingers on one hand), but it is still fundamentally just copying, pasting, and then deforming lines together.
People claiming that AIs understand what they’re doing and that what AIs do is not fundamentally copying and pasting are the ones who have the burden of proof here.
That's called "photobashing", very common to see used by concept artists. Also, it's fair use.
So now you've got a bigger problem in that almost all concept artists are engaged in theft I suppose? Since they are "only" copying, pasting, and then deforming things together?
It is also massively frowned upon in art sharing circles.
[[citation needed]]
>a somewhat ambiguous grey area at the moment
That's a funny way of admitting I'm right.
>It is also massively frowned upon in art sharing circles.
Who gives a shit? What part of "commonly used in concept art", and among professionals in drafting and conceptual stages in general, do you deliberately seek to not understand?
You mentally unwell or are you just an internet troll?
I am not the one arguing that a computer has thoughts, feelings and intent as you are here. It is you that is deliberately attempting to misunderstand through ignorance, lies, and insult.
https://www.owe.com/is-fan-art-legal-fair-use-what-about-mas...
Your “carte Blanche fair use” claim is nonsense. Copyright owners can, in fact, today stop derivative works that they know about, including photobashing.
By your own admission, AI is barely doing photobashing, so is very obviously not creating original works and is subject to copyright claims.
It gets tiring listening to HN AI enthusiasts pretending that AI is an actual comparison to human intelligence when it’s actually demonstrably a specialized search engine that does little more than copy and paste. If this were not the case, then:
-An AI would be able to feed itself without becoming shit tier nonsense
-An ai would be able to create materials not discernible from its training material
Neither case is true for any AI, including diffusion models.
Since you seem to not be aware of this: Fair use is an affirmative defense. By definition nothing I have written is therefore "carte blanche" as no defense of fair use can be ever said to be "carte blanche". This is pure sophistry on your part and transparently so.
>Here’s an article specifically on the legality of photo bashing
And mashups, photobashing, and many things likewise, remain legal and with fair use defenses in the courts as precedent. You continue to reach to any dishonesty or deflection including projection to avoid dealing with any of the points anyone has brought up.
That it doesn’t copy it’s very easy to prove: ask for something that is not in the training set. Does it generate it? Then it’s not a copy.
“Copy-pasting and deforming lines” is such a generic description (and to a certain extent a misrepresentation of what is actually happening) that I don’t see how it’s useful, or relevant. Even if we supposed this was true, you would be claiming that many of Duchamp’s works or potentially Arcimboldo’s or mosaics or collages are “just copies”.
>humans copy techniques
And as I already addressed, when humans are creating, there is real intent. When an ai does it, it’s a copy and paste with no intent.
What is your actual intention here? Where does the "intent" in humans come from that AI doesn't, even when commanded by a human to perform something?
If you ask an AI to train itself from its own images, it’ll always actually become worse using its own evaluation.
If you ask a human to train themselves, the human will notable become better of their evaluation.
AIs are fundamentally limited to their training set. Humans are not.
The fact that you can tell an AI to draw a specific thing, and it’ll steal content from other people doesn’t result in the AI having uniquely created an image. It has searched and blended strings together. That’s it.
Diffusion models are exposed to a blank canvas and then instructed to produce something. At no point are they looking into a database. They simply produce an image that will score highly in their aesthetic estimations.
Train an AI just on a few techniques. Then a couple fundamental stock photos with nothing but the subject. Then have it generate things outside of the realm of the stock photos recombined in a general style. Let us see your completely unique and unpredictable results.
Exactly dude.
The human did it.
I don't know where I fall on the issue myself, I think there's many good arguments on both side, both in pure IP terms and economic ethics terms.
A good counter to that is that the training data that was chosen as an input, and the prompt, are sufficiently original to count as their own work. Kind of like how an artist might use a copyrighted stamp to create their own work (Campbell soup cans).
And a good counter to that is that a translation of a book into another language is still considered a derivation of the original work, not its own work, despite requiring a whole lot of creative human input on the part of the translator.
I don't know!
At the end of the day, it's all about legislation and case law--and judges making calls on the borderline cases, which creates new case law. It even plays into whether models themselves are copyrightable which, in turn, depends partly upon whether certain doctrines (like "sweat of the brow") apply under current copyright regime.
Styles would seem to be more in the trade dress / design patent side of IP law. Apple’s famous 22-element iPhone design patent, etc.
But it’s hard to imagine an artist reducing their style to clearly defined elements and applying for protection. “A geometric grid or squares and rectangles with white background and black lines extending entirely across the canvas horizontally and vertically. Some grid rectangles are filled in primary colors. The square intersections of two black grid lines may be filled in primary colors. At least some of the grid lines are spaced irregularly”
You're one of those people who think that if you can imagine something and then describe it poetically, that this has any relevance to the real world that the rest of us inhabit.
So clearly there is some threshold of "transformativeness" which an AI must possess in order to be trained on one image and produce a different one that is considered a new work. Where exactly that threshold lies is difficult to define from a legal perspective.
Separate to that is the issue of producing new works that are too similar to existing works you've seen. If you do that as a human then you're liable for copyright infringement.
I don’t think I want to live in that world.
https://en.wikipedia.org/wiki/Implied_license
None of that covers the idea that observing a copyrighted work is a violation of copyright only made acceptable by an implicit license. As the link you posted notes, that’s not what an implicit license is.
1. Copying is theft (contents of a CD [software], movies, stuff on a website sometimes, etc)
2. Copying is not theft (VHS copies of broadcasts, cassettes, etc)
Problem is "implied license" (eula-roofie) where we've, without any good reason I think, decided that,
(a) Copying is theft when someone unilaterally declares what you buy to be different from the medium (CD vs what's on it [software, movie, etc])
(a)(2) except when it's a really old medium like a book which you're allowed to quote, and use in derivation, but don't you dare "quote a movie" by copying portions off that DVD or blu-ray because that's different. Shut up is why.
(b) Copying is theft when someone unilaterally declares their putting something up to view, which necessarily requires copying to view (sent over internet), can't be copied for derivative fair use
(b)(2) except when it's over another old medium like public broadcast because... fuck you? I haven't a clue myself how this makes any sense and I suspect it is because "it doesn't".
(c) Copying is theft if you take the copy sent you over the internet and do anything with it somebody unilaterally declares they don't want you to be able to do
(c)(2) except if it's an old medium, like a book, where you can borrow it and quote from it and do all the fair use stuff, or borrow a picture, or a painting, or...
All the fair use and copying and other sensible stuff without this weird "implied license" stuff doesn't exist for formats we've had before to the extent it does now. These formats, in principle, are still just "you get a copy of a thing on some medium", and yet we have two completely different laws based on this pure fiction of an "implied license" that declares you can't copy even portions of one on some format because... because.
I am trying, desperately, to explain there is fundamentally and in principle no difference. The root of all of our problems and why the law doesn't make any sense is the pure fiction that there IS a difference. And I feel like a goddamn madman yelling in the streets trying to explain something that feels so dog gon obvious to me but seems like is obvious to nobody else.
You can, but it would be more correct to argue that memories in a human brain are not copies under the law at all, and so need no license (or fair use analysis) in the first place. This has different results than an implicit license argument, especially when there is an explicit denial of license or an explicit license which is also explicitly bounded to exclude any rights not explicitly licensed.
As a society, we agree that you, a human, copying art you’ve seen is original (or at least derivative enough that it warrants a new copyright).
We agree that a photocopy machine, or rather a human operating one, is copying art in a way that is not creative enough to warrant a copyright — so the human goes to jail.
As a society, we have not agreed upon how transformative a “work” from a generative AI has to be to count as original, or even if it can. A recent decision claimed that no machine work without human effort would count — ignoring the effort to find good prompts, pick an image and edit it further. An AI is different because no much was “trained” to make art before AIs: some animals were (elephants famously), and who owns their work is tricky — there’s a case of a photo, a selfie actually, taken by a macaque that is really interesting in that regard.
https://en.wikipedia.org/wiki/Monkey_selfie_copyright_disput...
My recommendation would be to have a similar system to what we have for music (and I think plays) where there are two or three “authors”: the person who wrote the score, optionally the person who wrote the lyrics, and the interpret: the singer. It gets really fun when you have orchestras, animals involved in making music, or sound-sampling like in RnB.
We could have original human authors and AI adaptors, each paid based on use. If I like your work, I could commission OpenAI to make “a portrait in the style of alistairSH” based on your DeviantArt work, but of my mother (whom you’ve never met), using photos that I’ve uploaded of her. I can pay OpenAI, who would pay you. Style vs. theme could be separated with current AI technology, although that could get just as complicated as every intellectual property process very fast.
The monkey case seems fairly cut-and-dry to me - the monkey didn't steal the camera (as often reported) - the photographer engineered a situation to get the result (he set up the camera, the tripod, and held onto the tripod so the monkey wouldn't steal the whole rig). It's his art. If the monkey actually stole the camera, then it's nobody's art - just a random occurrence.
If an artist orchestrates a group of humans to create art, the art belongs to the artist, not the group (as far as I know - I'd love to hear otherwise).
I guess for me, given today's technology, I don't see an "AI" as any different than a Rube Goldberg machine hurling paint at a wall. A human had to create the AI and a human had to prompt the AI, one or both of them own the art. And in all cases, the art is always based on something the human saw previously. No art exists in a vacuum.
And I agree that a system like music rights might be appropriate. The AI author, the AI prompter, and probably anybody involved in the training data, all have some right to the output. Then the question is what % to each, but that's solvable.
It presents a sort of mirrored view of the GenAI issue: The human photographer does not get credit because the monkey is deemed to be the "agent of intent" rather than him, presumably because it's a living thing. If a photograph was triggered instead by a human tossing a rock onto the shutter, no one would claim that the rock now holds the copyright!
Interesting thought experiment to replace the monkey with other "organic agents" and some method for them to trigger a photo: A squirrel? A lizard? A cockroach? A slime mold? At what point is the organism so "simple" that you say it was just part of the system/mechanism?
Then on the artificial side, to replace the monkey with different mechanisms (as was your point): A photocopier seems quite different to a GenAI model. At what point does the human who set up the situation stop being the "agent"?
Relevant concept of the Intentional Stance: https://en.wikipedia.org/wiki/Intentional_stance
EDIT: Or at the far end of the spectrum, what about "art" that is made from DNA activity?! Does the copyright belong to the DNA? https://www.labiotech.eu/trends-news/gene-craft-laurie-ramse...
It just seems obvious to me that "human made thing to transform some set of things into a composite+transformation of those things" is fair use if fair use is to make any sense at all. Since analogously, like you said, in the same way you are taking your experience including experience of copyrighted stuff as a set of things you are drawing upon producing something of your own. All art is derivative.
It makes zero sense to me at all that it's suddenly a problem that art is derived if, instead of artist directly painting something, artist sets up some device that paints something. In both cases "human takes inputs and produces something unique" (not a copy, not a scanner, in case that was not clear from context).
Unless I'm missing something too, but based on what I see so far I just feel like I'm from mars and don't belong here.
I think a big difference in how this tech makes people feel is the amount of effort required. It's a problem we see with technology in other areas, too: it's not (much of) a problem if a cop sits outside a suspect's house for a few hours to monitor their behavior. Most people think it would be a problem if the cops pointed a camera at the house 24/7 for months. Pretty much everyone agrees canvassing society with cameras and creating a public space surveillance panopticon is bad. Where's the line?
Similarly, if some artists spend years learning the craft and mimicking a couple of styles with a relatively small output, it's not a huge problem. The scope is small. But if everyone on earth can do it to anyone on earth at massive scale just by typing "in the style of ..." to an AI, does that start to be a problem?
We've got much bigger problems though but if we, we as people generally not specifically you, can't even agree on something I see as so fundamentally basic bringing up things I think are problems would start war
Aye, there's the rub ;) How do you define "too similar"? If I invert the colors of Munch's The Scream, is that an original work? What if I pass it through a computer program I wrote to swirl it around in a spiral? I think those are clearly derivative works--you put the image in as input, put it through a mechanical transformation process, and get the same output for the given input. That description also applies to AI.
> these algorithms generate some general idea of the things it's trained on
Well, be careful here. AIs don't generate ideas. They take inputs, do some mechanical work on them, and output something derived from the inputs. There are no "ideas" involved here, it's a(n extremely complicated) mechanical transformation.
(To repeat myself, I don't know where I stand on the issue. I think there's good arguments on both sides.)
... Same ways I do for humans. That's why I wrote "the same as for people" that line was supposed to contextualize everything else I wrote and somehow it hasn't.
edited to fix a screwup. because somehow I swear I hit "copy" but it didn't copy
> The line is pretty simple for me...
When things seem so obvious and simple, it can be a good mental exercise to try to put yourself in the shoes of the "other side" for whom it also seems obvious and simple.
While never my main thing, I've worked as a professional artist and know people for whom art is their livelihood. The holy grail in that world is to create a unique style that will command a premium: an art director decides your style is perfect for their new campaign, a building designer decides your style fits the lobby of the new building, etc. This style is the result of years and years of refinement, false starts, watching trends, etc. And this style is why they get paid, and how they feed their family.
When you tell such an artist that now any schmuck can create art in their style just by writing "In the style of..." you should understand that —especially to a non-technical person—the end result is just a slight deviation of from a copy machine. To them it seems "pretty simple" and "obvious" that this is just a fancy way of stealing.
(And I'm not taking sides here, just saying that step one is to realize that the issue is not simple, and is thus super interesting.)
I would if I could but as I keep tryin to explain I can't because it requires I believe something that's a contradiction.
Just because it's a contradiction that makes ya more money don't mean it isn't a contradiction. Also I am that non-technical person.
The question is... Is fair use an exclusively human right? For example, we have many rights that do not extend to our possessions. Animals have no protected right to life or liberty. The human rights enjoyed by their owner do not flow to them.
Similarly, just because a human has a right to fair use, do AIs?
Put another way, if I create a Rube Goldberg painting machine, what copyright law applies? Why is that different than if I create/train/prompt an AI to do the work? One is physical, one is digital, but both require a human to create and initiate. And in both cases, every decision that human makes is made based on their life's experiences. No art exists in a vacuum. All art is derivative.
A machine has no rights, and it doesnt learn as humans do.
I know the cult leader, altman, wants you to think otherwise, but no, stealing people’s work to make ai viable is not OK.
Once he manages to replace y'all with ai, what do you think you will become?
If the question is in how biologically we learn compared to and AI, the answer is that humans don't learn by calculating math in order to determine how pixels relate to each other. Humans uses chemical signals between neurons and neuron activiation among other things in order to approximate the concept of what we consider to be learning.
I find a similar and equal important question when asking whats the difference between an AI learning and a human learning, is to ask whats the difference when a computer remembers information by storing it on a file compared to a human remembering information by storing it in long term memory.
Bits are not a bug.
As for easily recreated, it is not hard for anyone to make a wooden block and then use it to make prints but if you made a wooden block print of Campbells soup cans you would probably face a copyright claim from Andy Warhol depending on the "look" of the print.
Because as soon as you create it and try to sell it for $1, someone else will recreate it instantly and put it up for $0.50, and so on until the value of all non-physical works is effectively $0 the moment after creation.
Feels like that would result in way less human art being made.
Seems more like quality would matter. Hard copies would matter; perhaps the disc you issue your media on is rated to last 100 years, whereas someone's digital copy can be deleted with a keystroke.
Copyright is bullshit in a market of perfect reproducibility.
It's worth noting that copyright is rather destructive since it impairs people from building and innovating on previous works [2], so any benefits need to be weighed against that.
[1] https://youtu.be/mnnYCJNhw7w?si=UzmBasgylpxomUAt
[2] https://www.gnu.org/philosophy/shouldbefree.en.html
Like look how expensive AutoCAD is, why wouldn't Adobe charge that sort of money or more to access image gen if them and 2 other companies max can possibly provide it.
Sorry but at least with open models we all can benefit, you download it once and its FREE FOREVER you only pay for power used. It's petty and short sighted to try and take that away under the guise of "ethics".
Five billion images in compressed lossy format can easily fit into an active program. The only variable is quality. Five billion images in the original format and quality is a different matter, but then that requirement for copyrightable would be drastically insufficient.
One do not need 2 bytes per images if one million of images are blank, even if they are uniquely of different sizes. The amount of bytes to represent one million blank images in code is very few. One only need to store the unique aspects of the images, and if one compresses the images then the range of uniqueness drops based on the quality required.
Very unique images will naturally have larger footprint, based on the quality that one want to retain.