An IP attorney’s reading of the Stable Diffusion class action lawsuit
katedowninglaw.com
katedowninglaw.com
How can this possibly be a valid good faith argument? Either they're in breach of authors' copyright which extends to every piece of art that they included in the dataset without permission, or they're in the clear and aren't obligated to respond to removal requests.
This reads like damage control to me in an effort to temporarily silence the loudest critics.
These artists complaints are ridiculous, and are being made by people who don’t understand how things work.
If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing.
They either don’t understand how it works or they are just upset that a computer can make art as good as (or better than) they can in a fraction of the time.
All knowledge workers and creatives are going to face this in the future. It’s going to suck, but it would be great if we all could try to understand reality first.
> All knowledge workers and creatives are going to face this in the future. It’s going to suck
This is not a given. It's up to us and the copyright law. Real original work should be compensated appropriately unless you're proposing that we accelerate deployment of universal basic income and completely abolish copyright law.
I have a feeling you might not like the violent outcome if you effectively strip original creators of their copyright, give corporations the right to effectively generate infinite profit off the backs of their work and tell the creators (and other people whose jobs will be automated away) to pound sand when they ask how they're supposed to pay rent from now on.
I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far.
No I don't want that but there's a very clear distinction between people who are truly inspired by each others' work and produce an average human output that maybe covers their bills and AI that can pump out billions of replicates per day to drown out all original work, with entirety of that value captured by some corporation.
> I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far.
Hey, me too. But those human issues should be addressed first, not after automation is allowed to wipe out people's livelihoods. I'm not anti-AI, I'm anti-big tech that seeks to exploit billions of people's original work for their own benefit.
Copyright what? Someone’s brain? You can copyright a specific work or a character, but you actually want to live in a world where someone can copyright the color red with a dark black line, or the G# chord?
Real artists are going to art, and musicians are going to make music. People who do creative work do it to express themselves, their point of view or to say something.
Corporate art exist to sell you soda - I am not sure your argument lands quite like you want it to.
I am glad this is going to court, because, with my understand of how neural nets work, I fail to see how any copyright is being infringed.
This argument is assuming its own conclusion, that such a situation must be bad. But I don't think that's necessarily true.
If somebody can make 1000 different derivatives that the public likes as much as the originals, then it must be that in whatever criteria the public is interested in, these works are just as good. If they were inferior, the public wouldn't accept them as substitutes. The fact that they are (hypothetically) accepted indicates that the public is OK with them.
For my own personal aesthetics, I would like to think that today's popular music, which is written by some combination of algorithm and committee, and produced through tools that correct the performance via autotune, quantization, etc., is inferior to the music that I enjoy. But given that the public seems to like this music (and indeed, they like music generated this way even though it's not even cheaper for them to consume) seems to say that we as a society are getting what they want, and who am I to put a normative judgment on that?
In our future AI-infested world, I'll personally seek out "certified non-AI" creators because part of what inspires me is not just the content, but the creation of it.
Art is not just a destination.
And as I mentioned above, I think the current music industry is already there: the vast majority (by sales) of music entertainment, produced by algorithm and by a committee in order to drive sales. Despite this, in the genres I care about, at least, the volume of high-artistic-merit music (by variety) that is probably at its highest point in decades, if not ever. To be sure, this has meant that fewer artists are able to make a living purely off their music. But this is a return to the norm: the rise of the "star" in the late 20th century has been an aberration.
On the other side of the coin, AI assistance will be (I expect) a huge democratizing force. Recently we've seen computational photography enabling people to take photos of astounding quality with just their phones. And the results of machine learning is allowing artists to make huge improvements in post production as well.
I imagine that the stuff we've been seeing over the past year, with ChatGPT, Stable Diffusion, and such technologies, will be purposed towards (among other things) tools that enable greater productivity for the serious artist, and putting the means in the hands of those who would be otherwise unable to get to table stakes. I've started working on a short story myself, using ChatGPT to help me work through some plot points.
So yeah, we'll get a lot of meritless dreck suitable only for base entertainment. But we'll also see a proliferation of art and of artists, as productivity increases and new entrants are enabled.
Yes, much in the same way that I am glad I live in a future where scribes aren't required to put text on paper: There is a massive amount of efficiency to be gained and enjoyment to reap for everyone who doesn't happen to be employed as an artist.
Current proposals don't have any intention of addressing that, they just silently kick the can down the road. What happens when nearly everything is automated there are no new profitable jobs that people can take on?
The value will flow to society via the cheap books (art). Value will also flow to authors (artists)[0] due to the facilitation of creation/distribution/replication. Where that value will come from is from the scribes (painters, sculptors, etc.)[1] whose contribution is rote.
[0]The intentional distinction here is that the value is in the conception of art, not in the execution of it in media.
[1]Imagine how much more productive Mozart or Beethoven could have been had an AI-powered orchestra existed in their time.
IMO an artist that wants their name out of the dataset is a moron. In the end , people copying an artist style over and over will just send the price of originals through the roof. This is completely obvious.
My creative output per minute has probably increased threefold in the last few months from incorporating these tools into my workflow. What I've been making doesn't look or sound like anything that anyone else is making.
You're going to have a really hard time using Stable Diffusion to make quirky cartoon daily desk calendars in the style of plaintiff Sarah Anderson. You're going to have a much better time if you think more like a Creative Director and have less of an idea ahead of time of what the tool is going to give you... so you can iterate, much like a painter iterates while working.
These tools require the creative agency from the artist who is using them in order to produce things that people find interesting, entertaining, valuable or otherwise meaningful so I really don't see a "corporation goes brrrrrr" doing anything other than flooding the lowest-common denominator content feed pipes on the internet contrasted with the highest quality art using these tools in incredibly transformative ways.
More pointedly, how do I keep my GPL'd code from spewing, license free, out of CodePilot?
> https://huggingface.co/spaces/stabilityai/stable-diffusion/d...
They were prompted with the text "Mona Lisa Smile". Would you not say that they are an extremely close reproduction of the Mona Lisa, with barely any kind of synthesis?
I can hand paint a Mona Lisa like image that are this removed and be fine.
And one could also try a parody angle - make enough of these of famous art and find some angle about mocking or parodying that art, and again it may well pass copyright muster.
A court could simply rule that these images are clearly not the Mona Lisa, and, if taken as a style, could be ruled transformative, just like the above cases.
The fact is these are transformative, with a different style than the original.
Try making a comic book with a character that looks like Mickey Mouse and see how well that goes.
I don't see how more people copying the artist style would not increase the value of originals.
A smart artist here should promote that their style is staying in the dataset. It is as good free publicity as they will ever get.
e.g. Ask 10 random people who wrote the song “Hurt” and 9 out of 10 will probably say Johnny Cash.
sd might not directly include this info in its outputs, but it really is free marketing for some (in fact, i suspect a lot of artists are probably flattered for being included by name in the dataset, but some are riding the waxing wave of outrage for the free media exposure)
Stable Diffusion's U-Net is trained to remove noise from images in latent space, which the variational autoencoder (VAE) converts to and from pixel space. CLIP embeddings are used to improve the denoising step of the U-Net by using the correlations between human language descriptions of the pixel image to reduce latent noise. Neither the U-Net nor the VAR are trained to interpolate or reproduce images from the training set; if that happened the model would be overfitted and loss would be terrible on the validation set. The VAE is trained to produce a latent space that can accurately encode and decode any pixel image, and the U-Net is trained to remove gaussian noise from the latent space.
Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images (minimum of 170M in the 512x512 step alone). That leaves a maximum of ~20 bytes per image that could conceivably be a copy, which is certainly not enough to directly reproduce either the style or contents of any individual image.
There is no artwork included in Stable Diffusion. There is a semantic representation of how images are composed of varied subjects represented in the latent space and what pixel probabilities over those subjects relate to human language phrases during decoding, and finally a method to remove noise from the semantic representation, e.g. starting with a blank or random canvas and interpreting what may be there, iteratively guided by CLIP embeddings. If you give Stable Diffusion an empty CLIP embedding you get a random human-interpretable image obeying the distribution of the learned latent space.
You might as well say that there's no artwork included in a .jpg, just data that can be used to recreate a piece of artwork using a carefully crafted interpreter.
It is not a copyright infringement I go to Disney's website, download a JPEG, convert that JPEG to 1s and 0s, print just a bunch of 1s and 0s and not the image and not ascii art of the image, just like a printing press made up of just [1] and [0] character blocks, and sell that. Yes, the 1s and 0s are mathematically derived from the image but the image of 1s and 0s is not a visual derivative of the Disney image. That is, no one is going to buy a t-shirt of 1s and 0s instead of a Mickey t-shirt. Anyone can go to the Disney website and get those same 1s and 0s.
Again, anyone can go to the Disney website and get those same 1s and 0s, so this is not at all about access. This is about putting things on t-shirts and selling them.
Only because of applied license.
> convert that JPEG to 1s and 0s, print just a bunch of 1s and 0s
Yes, there is, barring context that maked it fair use (which is an exception that applies to things that would otherwise be violations), this is a fixed exact copy in an alternative form jist as much as a .zip or a lossless .png converted from the .jpg would be, and violates exclusive rights in copyright.
> Anyone can go to the Disney website and get those same 1s and 0s.
That’s an argument against damages, not against it being unauthorized copying.
Aaargh. Implied.
Latents are a compressed representation of the source images that are fully recoverable.
If you train a model on a compressed jpg of an image, or on any deterministic transformation of it, you’re still training it on that image.
Any suggestion otherwise is only because someone is trying to put some spin on things.
> Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images…
And yet! Remarkably! It can generate pictures of the Mona Lisa!
Here’s a question for you: if you encode the process of drawing an exact copy of an image, does the pure code that implements that mean you have a copy of the image in it?
Have you encoded pixels as code?
Does that mean there’s no copy of the image?
How about a zip file full of images? It’s just a high entropy binary blob right? Yet… remarkably!!! It can be transformed into images by applying an algorithm.
I don’t know the answer, but this handwavy “it couldn’t possibly encode them it’s too small” is…
Pure. Nonsense.
Of course some part of some images is embedded in the model in some form.
Stop trivialising the issue.
The issue here is: Does an algorithm that generates content infringe copyright?
Does a black box that takes the input “a picture of xxx” and a seed and outputs a copyrighted image infringe?
You know that’s possible. Don’t dodge. Technical details about oh “it couldn’t possibly have…” are pure rubbish.
Sure it could. It could have a full resolution copy of a photo of the image in that black box.
Of all the training data? Probably not. But of some of it? In compressed latent form? Most definitely.
"Image diffusion models such as DALL-E 2, Imagen, and Stable Diffusion have attracted significant attention due to their ability to generate high-quality synthetic images. In this work, we show that diffusion models memorize individual images from their training data and emit them at generation time. With a generate-and-filter pipeline, we extract over a thousand training examples from state-of-the-art models, ranging from photographs of individual people to trademarked company logos. We also train hundreds of diffusion models in various settings to analyze how different modeling and data decisions affect privacy. Overall, our results show that diffusion models are much less private than prior generative models such as GANs, and that mitigating these vulnerabilities may require new advances in privacy-preserving training. "
I think it is to avoid any common law wrongs related to the publicity rights of the defendants. It seems like something that a legal team would flag as an unnecessary risk for the product. Removing their names and images from the training data doesn't impact the usefulness of the model while at the same time creating a much smaller surface area for collecting subpoenas.
Notice that it's a legal posture that implicitly condemns Copilot, which ignores explicitly formulated opt-outs in the form of licensing.
They could just as well not do anything and continue on - it's likely this case will be in defendents favor. Same as how Google can crawl the net, cache data, transform it, etc.
This didn't make any sense to me. Without the curated training data (images) how are they making the models?
No matter what, putting images into your machine then selling the output generated with them and not compensating the original creators is going to be seen as problematic. Machines aren't people.
> It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for
I bet that you are massively underestimating the cost of that.
There's no reason why that is the significant detail. Why does it matter? If you can look at millions of images over your lifetime and faithfully reproduce famous works of art by hand, aren't you just as wrong?
Setting aside the question of "is the model a derivative work", running the program cannot create a work that is copyrighted. Only humans (and not monkeys) can hold a copyright.
And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?"
If think you're going to need more clarity on what you mean by that. Programs are used to create copyrighted works all the time. And machines can and do create copies of other people's copyrighted works.
> And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?"
My point is you can take the machine or model out of the question entirely. If you learn stuff and then produce something new with what you learned, is that a derivative work? That's already a complex question but it has nothing to do with how you learned it. It depends entirely on the output and has little to do with the input.
> The US Copyright Office has rejected a request to let an AI copyright a work of art. Last week, a three-person board reviewed a 2019 ruling against Steven Thaler, who tried to copyright a picture on behalf of an algorithm he dubbed Creativity Machine. The board found that Thaler’s AI-created image didn’t include an element of “human authorship” — a necessary standard, it said, for protection.
https://www.wipo.int/wipo_magazine/en/2017/05/article_0003.h...
> Creating works using artificial intelligence could have very important implications for copyright law. Traditionally, the ownership of copyright in computer-generated works was not in question because the program was merely a tool that supported the creative process, very much like a pen and paper. Creative works qualify for copyright protection if they are original, with most definitions of originality requiring a human author. Most jurisdictions, including Spain and Germany, state that only works created by a human can be protected by copyright.
https://www.copyright.gov/comp3/chap300/ch300-copyrightable-...
> 306 The Human Authorship Requirement
> The U.S. Copyright Office will register an original work of authorship, provided that the work was created by a human being.
> The copyright law only protects “the fruits of intellectual labor” that “are founded in the creative powers of the mind.” Trade-Mark Cases, 100 U.S. 82, 94 (1879). Because copyright law is limited to “original intellectual conceptions of the author,” the Office will refuse to register a claim if it determines that a human being did not create the work. Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 58 (1884). For representative examples of works that do not satisfy this requirement, see Section 313.2 below.
> 313.3 Works That Lack Human Authorship
> As discussed in Section 306, the Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a) (emphasis added). To qualify as a work of “authorship” a work must be created by a human being. See Burrow-Giles Lithographic Co., 111 U.S. at 58. Works that do not satisfy this requirement are not copyrightable.
> The U.S. Copyright Office will not register works produced by nature, animals, or plants. Likewise, the Office cannot register a work purportedly created by divine or supernatural beings, although the Office may register a work where the application or the deposit copy(ies) state that the work was inspired by a divine spirit.
> ...
> Similarly, the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author. The crucial question is “whether the ‘work’ is basically one of human authorship, with the computer [or other device] merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.”
----
A machine cannot create a copyrighted work. The human who uses it can - and it is the human who uses the ML model to create a derivative work - not the ML model itself.
Thus any derivative work infringement from using Stable Diffusion is from the human creating the prompt doing it. This doesn't attempt to answer the "is the model itself a derivative work".
Unoperated machines are not spontaneously creating art of their own motivation.
The model behind Stable Diffusion (or Dall-E) is not infringing itself. It has the latent images it, but there is nothing that you can point to that says "that number there is infringing."
It takes a human action to pull together the parts of the model and select what is right and craft the prompt. Asking for "a picture of Mickey Mouse in the style of Pixar" is all sorts of infringing - by the person asking for that.
By itself, that image isn't a problem any more than the doodle in my notebook is. The question is what the human does with the image. If they publish it or try to sell it - yes, that's a problem. If they hang it in their kids room, yes, its infringing but unless you invite a Disney lawyer over for a play date no one is going to care (this also applies to if you drew it yourself rather than instructing a program to generate it).
My (I am not a lawyer) take is that the model isn't infringing. Use of the model may be infringing, and issues of infringements should be taken up with the human who created and published it them just as if it was done by a painter drawing heavy inspiration from a style or a photographer with a photograph of the Eiffel Tower at night.
What about a company where you submit images and it tells you which faces are in them?
The reason we have copyright for a limited time is to promote the arts, not to give people a monopoly on a specific style so they can milk it. (although that is what seems to be happening)
If the image is freely viewable (say you can browse to it), and you just look at it, are you violating any rights?
It seems that violation would only come if you would use the model to produce images that are derivative of that original image, the same way a counterfeiter would make a copy of it. Have the skill to copy is not the same as actually copying.
The fundamental issue with this line of argument is that it equates the process of human vision and the consequences of that with that of a computer program ingesting that image and the consequences of that.
This anthropomorphization seems like a form of deep fallacy when considering the nature and impact of AI software. In the case of "seeing" an image, the two processes could not be any more unlike each other, both in content and context.
Without consciousness, it’s just a biologically inspired computer program. With consciousness, I suspect an AI modeled to understand ethics would refuse to provide certain outputs of its own accord.
And the analogy quickly breaks down the moment you continue to compare these processes and their context.
Whether or not it is superficially similar, the barrier to entry and the upper ceiling for infringement have both drastically changed overnight.
AI is not an independent entity that has entered the game, it is (currently) a power to be wielded by anyone regardless of their background. It can only be used as ethically as the person sitting at the keyboard, who most likely does not have a sufficient understanding of the underlying systems to make an informed decision (I suspect that if using the AI software involved the end-user feeding images into the model as a prerequisite step, they might have better intuitions about how to understand the implications of the images they generate from the resulting model).
> so nothing really changes for the "ethically sensitive" use-cases.
I think the thing that changes is the whole playing field. When overnight, anyone with a recent iPhone can generate highly sophisticated art/images with no artistic practice/training, it seems hard to argue that nothing has changed.
Before AI, even with the constraints of human capability, the art world was full of stories of stealing and bad behavior. Some blatant, some ethically questionable but thought provoking, etc. For all of their promise, the tools at hand have the ability to grow that kind of misuse at unprecedented scale.
What it even means to exist in an "ethically sensitive" framework likely needs to change. Or at the very least, current thinking needs to be examined to determine if it still makes sense in light of these new tools.
Considering the definition of that word, may I ask what you're trying to say?
> What's your definition of consciousness?
I like Thomas Nagel's:
"A creature is conscious if there is “something that it is like” to be this creature; an event is consciously perceived if there is “something that it is like” to perceive it. Whatever else consciousness may or may not be in physical terms, the difference between it and unconsciousness is first and foremost a matter of subjective experience. Either the lights are on, or they are not."
It is because of this subjectivity that I find it problematic to give weight to arguments that equate human consciousness with machine consciousness. Even if we achieve AGI tomorrow, and even if we know with certainty that it is conscious, it does not automatically follow that we would apply the same frameworks to a newly conscious entity on the basis of consciousness alone.
Consciousness and the implications of that consciousness can vary drastically, e.g. no one wants to be in the same room when the sleeping grizzly bear wakes up.
> How do you know that a (sufficiently complex) biologically inspired computer program doesn't have it?
I think we will eventually have to take this question seriously, but current systems do not seem to approach the levels of complexity required. But taking this seriously is not at odds with the belief that current AI programs are nowhere close to the level of complexity we associate with conscious creatures.
> What's special about meat?
I think this is the question that many scientists and researchers would love to answer.
There are some lines of thinking that consciousness is an emergent property of a sufficiently complex biological system with a sufficiently complex nexus of computation to make sense of those systems. In this line of thinking, the experiential aspect of consciousness - e.g. "what it's like to feel pain" - is just as critical to the overall experience as the raw computation capabilities in the brain.
Maybe meat isn't special at all, and consciousness springs from some other source or confluence. Even if it does, we then need to have a conversation about whether consciousness is the great equalizer, or if the "kind" of consciousness also plays a role.
Going back to that grizzly bear, no one wants to be there when it wakes up, but neither do we hold the bear to human standards of value. If the bear kills someone, we don't ascribe to it titles like "murderer".
But again, even if biology is not a key component, I still don't believe arguments about consciousness can be used as a basis for the ethics of the current generation of tools, which are far too primitive relatively speaking.
Computational neural networks are not models of biological brains, nor are they even attempting to be.
The basic functioning of a computational "neuron" in a neural network is at most reflective of an extreme distillation of the most fundamental concept of how a biological neuron works. And it really is just their functioning - ie executing.
The most important parts of making a computational neural network actually give meaningful output - training - doesn't even rise to the level of being vaguely inspired by the deconstruction of the concepts behind biological functions.
So, no. They aren't models of biological brains any more than boids are models of actual birds.
As for the goals of reasonably anthropomorphizing them... you're talking pretty much full on artificial general intelligence there. I don't believe anybody is reasonably suggesting modern deep learning is even a particularly viable route there, never mind something that's an active goal.
I'm not putting any weight here on what is good or bad for society, but relying on that humans somehow work in a completely different way from where AI is and is going is not going to help.
I do think it will take longer for the AIs to know all about human contexts though, so the pairing of human AD + bulk-gen AI seems to me to be an obvious near-term tag team that's hard to beat.
If we could develop literal eyeballs that could look at these images and translate the information the way humans do, the resulting capability is still no more human-like (in the sense that it should be afforded some human-like status) than any other program IMO.
If we achieved AGI tomorrow, we'd still need to have a conversation about what it is allowed to "see", because our current notions about humans seeing things are all based on the constraints of human capability. Most people understand that a surveillance camera seeing something and a human seeing something have very different implications.
In the short term, it's a conflation that I'd argue makes us see less clearly about what these systems are/are not, and leads to some questionable conclusions.
In the long term, it's a whole other ball of wax that will still require either new regulations or new ways of thinking.
If you place a human and a computer in front of a painting. A human seeing the painting is a consequence of biology. A computer seeing the painting is a consequence of design.
There's always a distinction between happenstance and premeditation.
You said a lot of words, but I believe your argument comes down to “computers are super powered compared to humans doing the same thing”? Is that accurate? Because magnitude of ability, to me, makes no difference at all. It’s perfectly acceptable for a human to study the artwork of a specific person and then create their own works based on that style. Why wouldn’t it be the same for an automated process?
No one knows why we are conscious. We have sliced the brain up a thousand ways and we will slice it up a million more and will never find consciousness because it is an emergent property of healthy brain, just like light is an emergent property of a working light bulb. No matter how you disassemble a light bulb, you will never find the light, though I grant you'll eventually figure out how light is produced, the assumption that a light bulb contains light is wrong headed. It's just a metaphor.
There is no worse slander than the truth: Strong AI can not be achieved, not with digital computers and programming and machine learning, and most likely by no other method either. Please, please grow up, and set aside your childish beliefs, because we need you now more than ever, here, in the real world.
Regardless of the details of how a brain is a conscious, it can be reduced to its constituent pieces or nuts and bolts, so to speak. Everything from the electrochemical potentials within neurons to the encoded chemical information in the form of DNA and RNA that spontaneously replicates and orchestrates a maddening array of complexities, we can partially explain. Even if our explanation is basically parts in a bucket, that's enough to paint a future where humanity understands enough of those processes to replicate consciousness without actually understanding why it works. Perhaps we don't need to understand the emergent properties, but merely discover them like the standard model in physics. We equally can't explain /why/ the fundamental physics constants are the way they are, but we can use them to do extraordinary things.
> I believe your argument comes down to “computers are super powered compared to humans doing the same thing”? Is that accurate?
No, that doesn't really touch it. The speed/power disparity between humans/computers at certain tasks are certainly a factor to consider, but the more fundamental point I was trying to make is much simpler: "computers and humans are fundamentally different, so let's stop building arguments on the mis-belief that they are the same".
> Because magnitude of ability, to me, makes no difference at all.
What is your position on autonomous AI weapons? Does that position change when there's a human in the loop? If such weapons were suddenly available to everyone, would that be functionally no different than allowing people to own firearms or baseball bats?
> It’s perfectly acceptable for a human to study the artwork of a specific person and then create their own works based on that style. Why wouldn’t it be the same for an automated process?
I'd turn that question around: why would it be the same for an automated process?
It is perfectly acceptable for a human to shoot an intruder entering their home in most states if they believe their life is in danger. An AI-controlled gun would be far more effective (I wouldn't even have to wake up!), but is clearly in a different category.
Is a human sitting on a neighborhood bench in view of your house the same thing as a surveillance camera on a nearby telephone pole? I think the answers to this question are useful when looking at the emerging issues of AI, at least to orient our basic instincts about what feels ok vs. what doesn't.
The AI software has only "learned" in the sense that it has operated on the input data such that it can now provide outputs that are of convincingly high quality to make it appear to "know" what it is doing.
Whatever the similarities, such learning lacks the vast majority of the context and contents of what a humans learns by viewing the same image, such that the word "learn" means something fundamentally different in each situation.
It's perfectly acceptable for a human being to drive a a car, but driving one drunk is completely unacceptable. Conversely, there is no rule against creating or consuming art while intoxicated.
So to answer your question, because it is not a matter of life and death. Take your argument and apply it to mass produced goods that were once the realm of only skilled craftsman.
> Is a human sitting on a neighborhood bench in view of your house the same thing as a surveillance camera on a nearby telephone pole?
If a person never leaves and keeps notes, yes, it is exactly the same. I'd call the police for stalking. The issue here is privacy, which is tangential to AI reproducing the styles of known individuals.
> The AI software has only "learned" in the sense that it has operated on the input data such that it can now provide outputs that are of convincingly high quality to make it appear to "know" what it is doing.
Completely disagree with you about the nature of learning here. If a person produces art in the style of an individual, they have no idea the internal machinations of the original artist, they just "appear to 'know' what they are doing".
You've lost me here. Are you saying that the most important factor when judging whether or not something is appropriate is based on whether or not the activity is dangerous enough to be fatal?
There are plenty of laws and cultural/ethical norms that restrict behavior for many other reasons.
> If a person never leaves and keeps notes, yes, it is exactly the same. I'd call the police for stalking.
You're arguing that a person taking notes with a pen and paper is the same as a video camera recording the same scene?
> The issue here is privacy, which is tangential to AI reproducing the styles of known individuals.
The point is that two forms of "seeing", one mechanical, and one biological, have two very different implications. If you don't believe that, ask the hypothetical person with a notebook to provide you with a 4K rendering of the scene over the last 30 days.
The AI reproducing art is just a single use case. The point of concern has little to do with how innocuous it is to produce images, but whether or not it is acceptable to use arguments about humans when judging what is or is not acceptable in an AI program.
> Completely disagree with you about the nature of learning here. If a person produces art in the style of an individual, they have no idea the internal machinations of the original artist, they just "appear to 'know' what they are doing".
Frankly, this is nonsense. We may not understand all of the underlying processes involved in learning, but we certainly know a lot more than nothing. Even if we knew literally nothing at all about the human brain, there is no standing to conclude that this lack of knowledge must imply that humans use some internal denoising algorithm when imagining what they will draw next.
We know enough to know that human processing of information is subjective, contextual, cultural, emotional, and there are a myriad of factors involved.
We know enough to know that what software like Stable Diffusion is doing looks very little like the human process for achieving a similar outcome, even if there are biologically inspired components inside.
> ask the hypothetical person with a notebook to provide you with a 4K rendering of the scene over the last 30 days.
You just seemlessly transitioned from a machine learning model to literal recording. These aren’t at all the same. In the context of your example, the person on the bench could have easily been wearing a body cam or recording with their cell phone, in certainly something they are capable of doing, so why would I treat it any different? The camera you mentioned could also be a CCTV feed with no DVR, in which case it couldn’t reproduce anything. The AI/person would be what would allow instant pattern recollection, like, “the person usually leaves in a hurry, but not on the weekends with a few exceptions” or “they usually turn lights on around X in the morning, and Y minutes after sunset”
> norms that restrict behavior for other reasons
What’s a concrete example of a tool that has been banned for other reasons. ML models like SD are tools, we usually let people use tools freely unless they can cause great bodily harm, and often even if they can.
Also I wonder where you get the view that future ML systems will not require large amounts of learning? I don't see any development in current systems that would allow that, or do you mean you have a network trained on large amounts of data which can then adjust to a style from a single image? If that's the case we are still at the same question, how was the original model trained.
Not only do I think the two processes are essentially the same, but I can't think of any laws in my jurisdiction (the UK) which actually distinguish between them.
E.g. we are allowed to make copies of digital media for personal use.
Essentially this is what the brain does when you do oneshot learning of traffic signs or characters when learning a new alphabet etc. (yeah sometimes it's not that easy but still it's "theoretically" possible :). The rest of the recognition pipeline is so general that styles and objects etc are just a small icing on the cake to learn on top, you don't need to retrain all the areas of the brain when adding a roadsign to your driving skill set.
But my point was that you could train the rest of the network on more general public data and not greg rutkowski. Hooray. Then someone shows it a single greg image and you're back to square 0.
I agree that this is a dangerous fallacy. Something that legislatures and culture have agreed is fine for a human to do - limited by human scaling, memory, and skill - may not be fine for a computer to do.
This doesn't mean anything. If an unsecured SSH server is connected to the internet that lets anyone who connects to it in and gives them a root shell, it is still illegal to 'hack' that machine. The law cares about intent, not technicalities.
edit: Since HN decided to break with "You're posting too fast. Please slow down. Thanks." again, banning me from replying: This is obviously just an example intended to show that the law cares more about intent than technical measures.
@dang Calm down dude.
If I read Harry Potter, then turn around a write a book about a wizard with a z-shaped scar? Who works at a school for wizards? With a pet owl? Who is an orphan? At some point I have started to violate intellectual property rules. (Ignoring all the Harry Potter material that was itself lifted from prior public domain art.)
AI systems aren't just reading, they are generating material based on the stuff they have read. They and the people controlling them have to abide the copyright rules just like any other "author".
See this link, lots of recent works in there.
https://en.wikipedia.org/wiki/Harry_Potter_influences_and_an...
The flaw in this argument is the word "artist". If you remove all the pictures from the data source, the AI isnt capable of generating anything. Because it's not an artist.
Second, your Outsider Art is something completely different, funnily enough a strawman. Surely you're not claiming that the creators of outside art have literally never had any sensory input in their lives? Or do you really think that one painter in two timelines, in one blind and in the other not, would paint exactly the same stuff?
Only a bodyless, artificial brain cant draw anything, blind.
No reason you can't apply that framework to AI.
Where AI might get into more trouble is that you might be able show literal copying in a way that it's impossible to do in a person mind. Like saving chunks of a work into its model.
Now if you are the company selling this product, how many people are feeling wronged and will sue - that's the class action part?
If you use the product to generate an image that is very similar to someone's art and they feel wronged and sue, would you still use it commercially?
This isn't the kind of question that the lawyers of the defendants are going to ask the court.
They'll more likely ask if it isn't clearly fair use similar to Sony v Universal and Authors Guild v Google and then present evidence of significant non-infringing commercial use.
> It seems that violation would only come if you would use the model to produce images that are derivative of that original image, the same way a counterfeiter would make a copy of it. Have the skill to copy is not the same as actually copying.
Yeah, that's basically how the courts see it these days although for a different reason. They don't ask questions about skills or work or anything like that. They ask questions like, "is this supposed infringing work a replacement in the market for the plaintiff's work?".
The deeper questions about what the hell anyone meant by the words in the Constitution about Copyright wait for the highest courts to get involved, which is where we got this nice division between tools and what the tools are used for which allows for innovative fair use of copying other people's protected works with tools like VCRs, online book search and large language models.
Those were not cases about 'generators' but about 'aggregators', a completely different class of application.
Those cases hinged on republishing works or significant parts of works as themselves, they weren't trying to pass them off as new, original works in their own right.
And this is exactly what this court case is about, whether or not Stable Diffusion ultimately is just another - complex - form of mechanical transformation or whether it creates original work.
In my opinion the only things all of these companies get wrong is that they (1) failed to obtain consent from the suppliers of the inputs to their models and (2) that they themselves have not contributed even a little bit of the input data.
The ultimate laugh test is whether or not these companies themselves slap (C) signs on everything and are willing to litigate when they believe their rights are the ones that are infringed upon. I hope for an outcome where opt-in will become the norm, that would seem to be a reasonable middle ground.
Finally, note that copyright is not a local concept but a global one - and has been for a long time - and that anything that happens on that front will have to ratified in a different forum than some US court.
No, it's not at all. This court case is about:
Plaintiffs Sarah Andersen, Kelly McKernan, and Karla Ortiz (“Plaintiffs”), on behalf of themselves and all others similarly situated, bring this Class Action Complaint (the “Complaint”) against Defendants Stability AI Ltd. and Stability AI, Inc. (collectively “Stability”); Midjourney, Inc. (“Midjourney”); and DeviantArt, Inc. (“DeviantArt”) (all collectively “Defendants”) for:
1.) direct and vicarious copyright infringement under 17 U.S.C. § 501;
2.) violation of the Digital Millennium Copyright Act, 17 U.S.C. §§ 1201–1205 (the “DMCA”);
3.) violation of Plaintiffs’ statutory and common law rights of publicity, Cal. Civ. Code section 3344;
4.) violation of Unfair Competition law, Cal. Bus. & Prof. Code §§ 17200, et seq.;
5.) and declaratory relief.
So for each of those complaints the defense needs to establish that their actions fit a different narrative, one that is legally coherent and against the claims for damages.
So for copyright infringement they are going to go for a fair use defense. I'm sure they won't only reference VCRs and Google Books! I'm certain they won't talk about "aggregators" and "generators" because this is not a Supreme Court opinion. They're going to use the established legal doctrines. I'm sure that their lawyers have plenty of other relevant case law at their disposal.
As for DMCA and rights of publicity, this seems to be what motivated Stability AI to adhere to "takedown requests" as they probably had some lawyer whispering in their ear that they probably don't want to spend the time and money testing this in court if it doesn't really impact the marketability of their tool.
I haven't ready anything about the Unfair Competition law in California.
Judges are far from stupid and a fair use defense requires that you primarily acknowledge that you are in fact infringing but that you feel that because it is fair use you should be allowed to continue to do so. This is a pretty risky strategy, especially when you are a party that is in the business of hosting other people's creative content.
We'll see how it all pans out, personally I think their position would be much, much stronger if they had bothered to obtain consent, even an opt-out email that if not responded to within say 3 months would count as consent (and no: obviously that's not the same but we're comparing the relative size of fig-leaves here).
As it stands I don't see how their 'fair use defense' will hold together under scrutiny without opening a much bigger can of worms.
[1] They had to admit infringement because of their first line of defense, then the fall-back required them to point at the site where they presumably took the code from which they could not. Don't interrupt your opponent when they are making mistakes.
This doctrine gives the lower courts a clear test that doesn't require diving deep into arguments about what it means for "computers to learn ideas" or for "language models to author".
The problem with an opt-in model is that it puts enough of a burden on a language model that it discourages non-infringing commercial creation. The opt-out-on-request model, which would need some form of legislation and probably informed by both the DMCA and right-to-be-forgotten in Euro area, would be preferable as it would have much less of a burden on model creators.
You seem to be making a very convoluted argument that eventually boils down to 'because it is useful it must be right', aka an argument from utility. But copyright law has time and again been proven to be highly resilient against such arguments. You either have rights or you don't and in a moment of clairvoyance the people that came up with the current incarnation decided that it is such an important thing that it gets bestowed upon creation. No registration required (though it can help). Just making something and boom you have a bunch of rights which you can only contract out of.
I don't think any utilitarian argument that results in the creation of new works based on the works of others will make those rights go away.
I've read your other comments and I see that this tool is useful to you but don't be persuaded so easily by the utility: If I stole your work and passed it off as my own it might be very useful to society, especially if I re-licensed it under more permissive terms or even placed in the public domain. But I would be clearly infringing on your rights. You may in fact not be in a position to claim these works as your creation.
The fact that 'my' work has a few hundred or even a few thousand such inputs rather than just one does not change the principle: I did not create the work, and that is the bit that really matters, unless you are creating a work it doesn't matter if you have the equivalent of a bitcoin tumbler for art at your disposal to pretend that you have created a work. You did not. The fact that you used a tool that obfuscates attribution and overrules the licensing terms of the original copyright holders does not mean that you can claim your hands are clean: you know exactly what is going on behind the scenes.
Personally I won't go within a mile of these tools to create work that I put my name under .
So I’ve used Stable Diffusion and I’m “literally” stealing from every artist. Prove it. Get a warrant to search my premises for signs of illegal language model use. How do you get a warrant with an image that has no visual evidence of being a copy?
I can prove to you that my writing and my code are mine because you can stand behind me and look over my shoulder to see that I am creating it, one bit (or at most 8) at the time. Visual evidence of it being a copy is not required: what's required is a track of creation aka provenance. This is a very well defined area in copyright. So the proof would be trivial: you recreate your work again, without access to Stable Diffusion or equivalent, while being monitored and if you can not then that would count as a bust in my book. If you're a good enough artist that you can do that and you merely use SD as a way to get some inspiration then that's another matter. But if the output of SD is in the workflow in such a way that its output is directly your input then that would be troublesome (to me). If it is copied then that still would be troublesome (to me). If you feel ok with that it is entirely up to you, but I have my views on that, which I'm perfectly at liberty to share.
And of course once you've made it work likely it is trivial to make it again so probably the above test conditions would need some sharpening but you get the principle.
Without provenance you are still creating art, but you are not creating original art. And if you were a good enough artist before SD then you should still be after, even if you're not using it and everybody else is, at least your work will be and will continue to be original.
I'm trying to imagine a world in which Rembrandt van Rijn admitted to using SD to create the Nightwatch, projecting it on a big canvas and then erasing the prompt. I figure it wouldn't give me quite the same feeling that it does today. Of course artists will use the tools that are at their disposal, but this tool essentially is a pocket sized library of all the other art that could be vacuumed up into the model and that just doesn't sit right with me. Which parts are yours and which parts are SD? If you claim an outsider can't un-entangle them to the point where cause and effect are separated do you really feel that that is the bar that should be met? If so what if I trained a model on your art exclusively and then used that to produce prompt driven 'art' to compete with you, would you think that's fair? And if not, why would the number of inputs be a factor? It's the principles that matter.
FWIW my dad made a living as a painter for quite a few years and even though I can't paint worth anything I know enough about brush technique to get oil to stick to canvas. Given this tool I could produce 'art' that I probably could pass off as original and it would likely not be detected by your standard.
But without the tool I would be absolutely hopeless so that's a pretty good argument. The more of an artist someone is the more brittle that argument becomes, there may well be a border beyond which what someone creates using such a tool does qualify as original work but again, to me, such a work would feel tainted.
Do you disclose the use of SD to the customers that buy these works?
Also, Stable Diffusion is already on my laptop so I don’t need an Internet connection, but I digress.
You’re right, I can’t make the same art with Stable Diffusion as without. The same holds for the art I can make with DAW software like Logic Pro and without. Or my guitar and without. Basically you’re leaving me with nothing but my voice, and personally, I’m in luck because I do sing and have already put effort into getting better.
I can’t see how an image that has never existed before cannot be considered original… it might not have taken much effort, but that’s something else entirely! How is any part of photography even remotely considered original in your theory of art?
The burden is on the accuser but the accuser has been 'shredded' and that makes them all but anonymous. So if you dilute inputs enough then you can ignore the rights? There is a close equivalent to this in music sampling. The rule there is that if you want to remix someone else's production that you will have to ask for their permission, which, crucially, they are not required to give.
https://www.romanolaw.com/2022/10/14/music-sampling-rights-w...
This seems to me to be a reasonable middle ground. Once you start breaking things down further it gets harder and harder to nail down where the limit lies. In music you probably won't find any useful samples shorter than 1/32th of a beat and even that would be pushing the limits. But the rules don't say anything about how long those samples are: even the tiniest sliver would have to be accounted for: you would need permission. And yes the burden is on the accuser. But it would make sense from an ethical point of view and as an artist to just play by the rules, rather than to see what you can get away with and so even for short samples that might never be detected by that particular accuser as a rule permission is obtained.
> You’re right, I can’t make the same art with Stable Diffusion as without. The same holds for the art I can make with DAW software like Logic Pro and without. Or my guitar and without. Basically you’re leaving me with nothing but my voice, and personally, I’m in luck because I do sing and have already put effort into getting better.
That's disingeneous and you are likely well aware of it, but just to spell it out: Logic Pro 'Sample Packs' all have clear attribution and use rights attached, and your guitar is (unless it is some advanced model that I'm not currently aware of) unable to make you play like Al Di Meola without putting in the endless hours of practice. Neither of these contain a combined mountain of copyrighted artists works that you can then re-use at your discretion without attribution and if and when they do those are very carefully sourced and marked as such. So the comparison with SD doesn't hold.
> I can’t see how an image that has never existed before cannot be considered original…
Ok, so I have this little piece of software here, it allows me to extract the notes, durations and volumes of individual notes of a piano piece and then I can re-play those notes using any instrument from an enormous bank of synthesized sounds. Do you feel that this should count as original work because 'it has never existed before' or do you feel that this is a mechanical transformation and that the original creator should be able to sue me for creating a derived work? Does the mechanism of the derivation count?
> it might not have taken much effort, but that’s something else entirely
Agreed, Marc Rebillet comes to mind (https://www.youtube.com/watch?v=3vBwRfQbXkg) , but he has 3 decades of experience which allows him to create in a few minutes what would take others weeks (or they might never be able to do anything).
> How is any part of photography even remotely considered original in your theory of art?
Photography is original art when it goes beyond 'just reproduction'. My brother, who is a pretty good photographer has two 'modes' in the first one he is simply registering an event, a wedding or something else that needs to be documented. He most likely would not consider this to be even close to his best work and certainly not in the level of artistry involved. In his other mode he can spend a crazy amount of time arranging the shot, including getting up at odd hours of the night, creating special vantage points, the right time of the year to get an object to catch the light just so. In that mode he is an artist. And I'm sure he has made 1000's (or probably more likely 10's of 1000's) of shots that are somewhere in between.
Yes! Always and only! If you are a songwriter then your inputs are a certain ordering of words and sounds. If you dilute those inputs to the point where it indistinguishable you have created a new song!
No, that's not the right criterium. You can create a new song but it may (not will) be classified as a derived work.
Note that in audio it is fairly easy to prove that a part was sampled and unless your samples are pure noise you would most likely not get away with this. Plenty of artists are more than happy to give permission but they are also likely more than happy to sue if you don't ask for it.
And some don't care. But that should not be your assumption.
Using the same words to write a different book is emphatically not the same as using samples of spoken/sung words or played notes or riffs to create another song. These arguments have all been tried in court (see that article linked above) and failed.
And that's where sites like this:
https://www.whosampled.com/Us3/Cantaloop-(Flip-Fantasia)/sam...
come in handy, if you listen to those other songs you'll realize how brilliant this Cantaloop version really is, in spite of being made using a very high fraction of samples from other songs.
But they did credit them, and they did ask for permission.
Finally, it's good form: artists are usually part of the same eco system and you tend to not do to others what you don't want done to you. If you feel that having your artwork sampled for inclusion by SD is the way forward then that should be your right to decide. But not for SD to decide (in my opinion...).
Also, from my time working in the music industry I know for a fact that kick drums and other sounds are taken from songs and used in other songs without permission or attribution all the time and it is done in a manner that is completely kosher because no artists think their artistry extends to someone trying to make the crack of a snare snappier in a mix. The lawyers think otherwise of course. Recognizable bass line? I mean, I guess pay up, but it’s not like the bass player who wrote that bass line in the session will see a dime. He never even got a publishing credit and why would he?
It's pretty thin ice, and if you mess up it will get costly especially with something that sees high rotation.
I’m not talking about even noticeable sampling of a drum hit but the drum hit being used just to add to the existing drum track in the mix.
This is where transformative allowances come in as well, if not logically because no crime can be committed without any evidence.
I mean, IMO, copyrights on music samples are egregious. Hip hop is significantly transformative and was stymied by lawyers for decades. I’m sure that the best hip hop songs were never allowed to be made. I bet someone put together some absolute best of all time bangers with The Beatles, Micheal Jackson and Prince but they couldn’t get the rights. That trade isn’t worth it my opinion. Luckily the visual arts are more open minded. Richard Prince can be an artist by taking someone’s photo, doodling over it, and selling it for millions while owing nothing to the original photographer.
> This is where transformative allowances come in as well, if not logically because no crime can be committed without any evidence.
I really don't share that particular view. Crimes definitely can occur without evidence, there are plenty of examples. The fact that nobody gets convicted, that possibly the crime goes completely undetected does not mean that no crime occurred.
Let me give you one example of how this could happen: I was in charge of the audit of the 'RSB', a system that controlled a few 100B guilders way back in the day, mainframes were in, the PC/AT was 2 and ran at a whopping 8 Mhz and cost as much as a really good second hand car, Madonna releases 'True Blue'. The bank I worked for had for compliance reasons two physically separate entities, one to produce the code and the data, and another to keep track of it.
Initially I worked on the side that did the daily audits to make sure that everything was good. The process was specified thus: you will receive about 1/8th of a cubic meter of large format (132 column) printout at 7:30 AM. You have 30 minutes to manually review this (feel free to use a pocket calculator that can handle 13 digit numbers if you can find one and be sure to let us know where you found it) and release it, you better have a bloody good reason to not release it because it means that technically there is no opening balance for the day if that should happen. Clearly this is impossible, good luck.
On the first day of my new job, at 7:45 I realized that this is completely impossible so I went to see my boss. He told me that in practice the only known way to make sure that the most horrific of errors are caught is to go to the last page of the report first, check the two number all the way at the bottom of that page and see if they match up, this can be done in less than a minute. And if they match - which they always do - you can safely release the system, it's never been wrong.
So much for the oversight department. About a year later I got accepted as a junior programmer on the other side, and guess what, while there was a formal division between the personal side and the business side of the bank you had access to everything. This was well before ISO27001, security was something someone once may have read something about but we mostly focused on physical security. Once you were 'behind the red rope' you could do just about anything. Sure we had logins and passwords. Three letter login, three letter passwords were pretty common. The three letter logins were typically just the initials of the employee because there wasn't room for much more so those were trivial to guess and there was no upper limit on the number of tries on a password. So you could access pretty much all of the code and while there was a test environment there was no code review worth mentioning. And frankly, with very few exceptions most of the programmers weren't all that good. So it would have been trivial to:
- use someone elses account in the right group
- make a bunch of changes to the RSB system
- ensure that the report that went out to the audit department had matching numbers on the last row of the report
- and finally, to trigger all this say a year after leaving the bank
This likely would never have been discovered because their whole idea of threat modeling did not for one second stop and look at the 'rogue employee' angle, it was strictly a single line of defense: to get on to the 'floor' you had to have a badge that allowed access. The badges were an older Motorola system, brown plastic casing with a sticker on the front with your picture on it. Inside was a beefy pickup coil all the way around the case and a board with a little computer on it that would respond to being powered up by sending out a serial stream ID'ing the badge. That was it. No encryption, no challenge/response. And the ID? It could be set with a nice 8 position dip switch. Guess what ID #001 did?
So all of the requirements for an absolutely perfect crime were there, there would be no evidence and yet a crime would have been committed. There were many other such opportunities, that bank really relied way too much on trust and I would not have been surprised to find out that they had been hacked horribly from the inside without being able to point at the culprit.
The bank was founded in 1918, well before the age of computers and it processed a small mountain of paper every day. On that front their procedures were pretty good. But on the digital side they were - frankly - absolutely hopeless. The gear they used then was old enough that some of it was still labelled 'Sperry Rand' which the company was no longer fielding for about a decade at that time. IT was a cost, security non existent and any half competent coder would have seen 1001 possibilities to defraud that bank, and get away with it cleanly. Maybe some even did.
Obviously a witness testimony of a person being tossed into a volcano is evidence of a body, etc.
So maybe they are locking up and trying people in Europe without evidence of crimes as being committed and having people shift the burden of proof to the accused but here in the United States we really do assume that everyone is innocent until proven guilty.
Our courts don’t expect to waste time either. Why would they hear complaints without evidence? Why would they hear a trivial complaint?
Part of the disconnection I have in these conversations is that people skip over the legal details and want to go right to some deep philosophical discussion about originality or justice, which is fine, but also super annoying because our legal system has developed a pretty sophisticated understanding of those concepts through literal millennia of trial and error.
The current legal system works so much better than whatever I hear commonly described in these comments. If you let nerds get involved the courts would be nothing but people nitpicking over the finest of useless details in some vain attempt to score argument points, as if they’ve been trained on upvotes for decades.
There’s real wisdom in legal practice and I’m always going to lean heavily on those practices when talking about legal matters like copyright.
Originality in art is a different conversation entirely and frankly even if the courts want to view 10ms audio samples as worthy of copyright (which I’m pretty sure they would not concern themselves with), from an artistic standpoint I’m fully onboard with the idea that the musician who chopped up that sample into indistinguishable parts and then made new music is the sole author of an original work.
Ethically, copyright goes too far the moment is begins to choke the public domain and that yes, this weighs the needs of a given individual against anyone else. I’m sure the first person to record a I-IV-V progression would prefer to get a dollar every time a future composer used that but that would clearly be stifling to art in general. Copyright law in the US has evolved to capture this same ethical understanding and it’s not the only embedded wisdom.
https://en.wikipedia.org/wiki/Murder_conviction_without_a_bo...
Other observations about how legal systems elsewhere work are ignored on account of the opening sentence.
So at this point have you run out of any ammo other than meaningless nitpicks?
That’s just silly. If no one can hear the difference then who cares? Why would it even matter what song I was getting tiny snippets from? It’s like asking a specific poet for permission to use a certain word.
> Ok, so I have this little piece of software here, it allows me to extract the notes, durations and volumes of individual notes of a piano piece and then I can re-play those notes using any instrument from an enormous bank of synthesized sounds. Do you feel that this should count as original work because 'it has never existed before' or do you feel that this is a mechanical transformation and that the original creator should be able to sue me for creating a derived work? Does the mechanism of the derivation count?
Does it sound like the same song or not? You know, would someone listen to it and be able to recognize the melody? That’s what matters!
A court, your lawyer, and then finally, you. After all, if there is no audible difference that is exactly what this would hinge on and in the age of digital music it is fairly trivial to prove that something was sampled on snippets as short as a couple of 10's of ms. Whether that amounts to infringement or not is another matter, but to prove it is not all that hard from a technical point of view.
> Does it sound like the same song or not?
No, because the instrument would be changed, the timing could be changed (slowed down, sped up). Even a musician would have a fairly hard time associating a piece manipulated like that with their own work.
> You know, would someone listen to it and be able to recognize the melody?
We could use a melody that is itself in the public domain for this (say, a piece by JS Bach played by some famous pianist). The melody may be PD but the performance definitely is not.
> That’s what matters!
No, what matters is whether or not it crosses the bar for being recognized as an original work, and I don't think it does. But you've got me thinking about this because I've been toying for a long time with the idea to clean up some extremely bad historical recordings of absolute masterpieces and this is one of the reasons I've been holding back from that, the copyright situation around those would be quite murky.
It's 2:20 am here and I really should be getting some Z's, thank you very much for the interesting exchange, much food for thought.
It is possible for a prompt to create a copyright violation, just like if I recreate an existing iconic drum loop it would be. But, just because a tool can infringe doesn't mean the tool itself is infringing.
That's fair. But if you would sample Phil Collins (just to name one famous drummer) to create a Phil Collins sample pack and you'd then create your own drum loop with it that wouldn't mean you can't expect some expensive mail incoming. And it will likely stick.
The easiest way to create original art without having this shadow hanging over you would be to make your own sounds, use licensed sounds or to attempt to license the sounds you want to use (but that are not available for public use). Any other path is likely going to be a legal minefield and may lead to you losing your rep and a bunch of $.
There are plenty of examples of people that did this, went to court and lost.
What is original to begin with? What's original about Bob Dylan's Blowin in the Wind? Certainly not the form! It's a standard AB folk song. Certainly not the chords! The melody? Sure, but very bounded by Western music theory and containing a number of common American melodic tropes. The words and specific rhymes have all be used before in previous poems and light verse. He used a standard 6-string guitar with standard tuning with standard guitar chords that have been strummed in similar patterns on hundreds if not thousands of previous recordings.
Beyond the technical skill required to create something there's nothing left but just a series of choices about how to rearrange what culture has provided for you. A truly original work would be incomprehensible to an audience in a way that a truly original language would be incomprehensible to an audience. The originality, the agency of the artist, stems from the choices being made, regardless of if that tool simulates a drummer based on the placement of notes on a grid or if that tool simulates a photographer based on the input of some key words.
Now, you can certainly say that if everyone has access to Stable Diffusion that the value of its output is relatively diminished and that is of course true. The same thing happened to drum machines. No one is that impressed by a four-on-the-floor beat coming out of a laptop and they soon won't be impressed by simulated painters but this is on a different axis of examination than originality.
The fact that he claims he made it, and that this went uncontested for decades is fairly strong proof that it really is his.
But I thought you were interested in originality outside of just the legal perspective? You know, that crimes can be committed without evidence and all?
What makes Blowing in the Wind different from another folk song of the time? What makes them the same? Why are some non-original aspects allowed in a work considered original?
Both show that there is no intent for individual infringement, with along with no infringing material being produced and the significant non-infringing commercial use like family photo touch-ups, it makes for a very strong narrative for the defense!
To change lanes to art for a second…
When I’m using a tool like Logic I can coax drums, Hammond Organ, 70s analog synths, out of my laptop. I can’t make any of those sounds with my body. I need a tool.
Without Logic I would have to own a drum kit (which I happen to do), know how to play it, know how to record it… with logic I just click a dozen times and I’ve got a drummer playing four-on-the-floor.
So when I do work with a real drummer don’t tell them exactly what to play on a grid like with a drum sampler in Logic, rather they just listen in contribute. Logic has a virtual drummer that does basically the same.
How different is it to fire up another piece of software and employ a virtual painter?
Ok, shift back over to technology and creative work… what are the pros and cons of drum samplers? Don’t they put drummers out of business? Isn’t the trade-off that now a lot more people have access to nice sounding drums and that the world has more music?
This is provably false: if any specific piece can be removed and there is no difference then you can keep doing that until there is no art left. I'd bet a very substantial amount of money that the output at that point will definitely have changed. Induction is a powerful thing.
Or if we want to do thought experiments, randomly remove any thousand images and the tool is still just as useful. If it’s random, not specific images, that are removed, then it is non-specific images that power the tool… yes, millions of those images, so a specific quantity is required but no specific works.
The problem as I understand it is in all the likely "precedent" cases for this, what was being done with the scraped data was in some identifiable way different than the purpose of the source data itself. Authors Guild v Google for instance, the argument was that Google wasn't reproducing whole texts, it just used that data to make the texts searchable. Meaning the purpose of the consumption of the text by Google was to essentially make a searchable index, rather than to reproduce a book, and thus that isn't harming the authors.
In this case, it would seem a very key difference is that this is Art being consumed and Art being produced, with no different purpose.
Sure, I mean, anyone can sit on the sidelines and imagine any sort of fantasy legal doctrines.
The one you’re imagining would have the courts deciding the purpose of art works.
In this case an example is easy. If someone uses Stable Diffusion to remove a person from the background of an image they are clearly using the tool for a separate purpose than copying Sarah Anderson’s works and making competing desk calendars.
But I guess your silly argument is that the doctored photograph is a “picture” and Sarah Anderson made “pictures” and all pictures have the purpose of “being looked at” and you want courts to decide this based on this reasoning?
Look at the extreme case, then. What if that one image is your only input, and your output is identical to it? What if your output is your input reflected over the x-axis? What if your output just crops the input? What if your output is your input cut into irregular pieces and randomly rearranged? Which outputs violate copyright?
Slightly less extreme: suppose your input is two images, and your output is those two images next to each other in a single image? Or your output is the second image, reduced in size and placed in the center of the first? What if both of the inputs are human figures, and your output is to cut out the face and hands of one image and put it onto the other?
> images that are derivative of that original image, the same way a counterfeiter would make a copy of it.
Only one of these outputs are anything a counterfeiter would do. Are any of the others copyright-violating?
Would they be able to use your photos for Adobe Stock without permission ?
With this sort of model's "creation" process, is something close to everything it generates derivative of everything it ingested, since had you ingested a different set of images you'd presumably have a different model with different weights?
That's kinda sorta analogous to human creation, but a human can much more actively choose what to think about, what to ignore, what to filter out.
The human process involves an explicit creative judgement step that I don't think the image-generation-by-model process can - and that creative transformation is key, legally, to a derivative work being able to itself be copyrightable and to not be infringing.
in order to create 5 very different illustrations you need to talk with 5 people. in the end 5 people will get money when they finish with their work.
an AI consumes these artists past output and instead of paying to these artists it will gather income to the owner. So by using the output of 5 people who have spent decades on perfecting their craft, the AI generates income by stealing their work, and the money flows to the owner only, who doesn't give back anything to these people.
so in essence AI in this form kills income stream for humans, since it gives back nothing.
What specifically is the defining reason that people can learn by copying other peoples styles but ai cannot?
Are we supposed to halt technological progress to avoid antiquated job destruction?
So you want an image? For 5 bucks? You get an image that's worth 5 bucks, but not an image that costs 1000 dollars to make in real life.
The problem here is you giving a simpleminded person access to an AI, and for a few bucks, this person can generate something that uses thousands of man years of expertise for that given work.
I hope you see the potential slippery slope here.
Since reasonably simplified information about SD is available and/or the plaintiff could have involved an expert to review his claims - it does raise a question if the function of the lawsuit is more about rattling chains rather than the merits of their argument. I.E. A deliberate ploy to extract a settlement.
I think one issue is just that of scale. I personally tend to agree that there's something icky with just slurping up literally everyone's content, then producing a tool that will then proceed to put them out of business en masse. But proving that illegal under current law is certainly going to be a challenge.
I have not read the original complaint but it surprises me that the lawsuit doesn't have a much stronger focus on this aspect. Copyright law is very concerned about not destroying the market for a given work through infringement, but this is a case about destroying the market for entire artists at a stroke.
But that's a hard argument in court. There's no legal basis for claiming damages because the entire market itself is being destroyed.
Though I'm not sure it's any weaker than the other claims trying to be made. The basic problem is, this isn't illegal in any sense. I don't just mean "illegal in that it must be banned" but any level of gradation in between, in the licenses, in requiring compensation, in any sort of regulation whatsoever. Technology has simply outrun law again.
If an AI art engine outputs a frame of solid blue, is it infringing the copyright of Yves Klein's solid blue "IKB 79"?
I think that some artists' styles can be accurately replicated without training on any of their work: because the artists' style is generic enough that it can be exhaustively encoded via the works of others.
This seems like a bad test because generic barely-creative works are much more easily generated by AI engines regardless of the source training data. I wonder if we're going to see IP troll style behavior from artists drawing many obvious things so they'll have standing to sue (and negotiate a 'fuck off' settlement) with AI art engines.
Probably not, though even this may be debatable given the specific prompt and specific similarities (for example, if it generated the exact color and exact aspect ratio for a prompt like "Yves Klein IKB 79", I could see an argument for infringement; if it generated the same thing for a prompt like "filled-in 16:5 rectangle with color #0000FF", it would be arguable that it isn't).
> I think that some artists' styles can be accurately replicated without training on any of their work: because the artists' style is generic enough that it can be exhaustively encoded via the works of others.
The important question in copyright is, theoretically, if you actually copied the specific piece or if you happened to create a similar-looking piece by accident. The difficulty of proving one or the other varies with the specific circumstances. For example, it's rather hard to claim you happened to paint a picture almost identical with Picasso's Guernica. It's rather easy to claim that signing an empty piece of canvas was entirely your idea and you weren't copying Dali's signed empty canvases.
Similarly for AI, the copyright discussion will come down to how much of its output is identifiable pieces of other's artworks, especially when prompted for such. I personally haven't played with it, but if it's possible to get SD or other similar models to produce identical or very similar copies of (pieces of) some artist's works using relatively simple prompts*, it should be a pretty slam-dunk case of copyright infringement.
* by this I mean prompts that don't themselves encode the information, like I showed in my "16:7 rectangle with solid color #0000FF" example.
Solid blue is superlative example, but I think the amount of creativity in different artworks has a HUGE amount of variance. You can kind of "score" any artwork using and AI Art engine in terms of "what is the minimum number of terms in a prompt that I can use to recreate a substantially similar image (that wasn't a part of the training set)"
Some artists' styles can probably be articulated in no fewer than 600 words. Other artists can probably be articulated in 6. The quantifiable amount of interesting decisions in a piece of art has orders of magnitude of variance.
If someone (manually) copies your style and lets AI train on their works, your 600-word masterpiece could instantly drop to 15-words once (human created non-copyright infringing) derivatives are in the training set.
We are explorers of the frontiers of latent space, and theres going to be all sorts of new things for people to get mad about along the way.
Consider that even cooking recipes (the particular descriptions) are copyrighted, even though they are mostly technical descriptions of a simple process (the recipe itself is not copyrighted, the specific wording is).
Disney can train a model from every frame of their video library as well as whatever they can find which is unambiguosly public domain. Then they could hire a few hundred artists to draw whatever the model is bad at by the end of this process for finetuning.
Art is the creative aspect not the skill in creating things. Art will be fine, those who just are scribes for painting are doomed.
e.g., there's copying Dali's color tone, realism level, and general look, and then there's understanding the why of what he is doing and being able to translate that out with some skill. You may never 100% match Dali directly, but a human doesn't necessarily need a lot of data to reach this level. And of course there will be a valid market in copying his tone, realism level, and general look even without a deeper level of semantic copying. A lot of the AI art is already this, but the person crafting the prompt and then performing a conscious selection of the result can bring their own semantics to the art in the process, it is not strictly speaking necessary to have copied them from the original artist.
This is actually part of what is going through my mind when I expect this will be solved via legislation. Well... for a given value of "solved". I expect whenever Congress gets around to this, any law they will end up passing will have HN up in arms and I expect to be armed with a torch & pitchfork myself, such is my confidence in Congress on this front. But even given that Congress is going to royally mess it up, it may yet be less messed up than the inevitable hodgepodge of conflicting precedents that will precede Congress' efforts, because while in my opinion there is no coherent legal principle that will hold the decisions together, the lawsuits will arise, the decisions will be made, and the precedents will be set. The mere inability to retain logical coherence does not stop our legal system.
I hope it returns when they win and get rid of this legal bullying.
Information comes with many different rights: copy-right is the right to make copies; "moral rights" were mentioned in a few of my UK job contracts and that's "the right to be identified as the author of a work"; database rights are for collections of statements of fact that are not eligible for copyright but which were deemed to be worth protecting anyway for much the same reasons.
Even if copyright is totally eliminated from law by the mere existence of these AI[0], we may well retain the aforementioned "moral rights". And even if it is totally legal, there's also a strong possibility of it being considered gauche to use an AI trained on the works of those that don't like this.
[0] https://kitsunesoftware.wordpress.com/2022/10/09/an-end-to-c...
I don't think it means the author has a right to all similar styles. If I can legally ask somebody to paint me something in the style of a famous (living) artist, that person presumably having seen and studied their famous works for a while, why should I not be able to ask the AI to do the same thing?
(I understand there might be people who think even a human person emulating the style of another artist is morally wrong, but at least that's a consistent argument)
I'm told that for people who think Rolex watches are important, a fake Rolex is much worse than not having a Rolex at all: the item is a signal much like a peacock tail, the signal is de-valued by making things easy.
but this isn't enforceable, they cannot be transferred
Comics are one example of an area where individual artists might develop a large body of work in a very distinctive style. You probably know what a Tintin comic (by Belgian artist Hergé) looks like. And lots of Manga artists have very specific and instantly identifiable styles. Individual artistry is a little less obvious with popular western comics because the best-known titles tend to be superhero franchises where the characters/story world are owned by a corporation and individual artists come and go.
Upside: the brilliant artist Rohan Kishibe falls under a bus, but while his loss is tragic his artistic legacy lives on - hooray!
Downside: up-and-coming artist Rohan Kishibe is on the verge of breaking through commercially, but the publishers of Shonen Lump, who invested heavily in AI, floods the market with work from Kohan Rishibe, and our hero is derided as a mere imitator.
In fact, I generated a bunch of Tintin panels between writing my earlier comment and this one, and they're bad but not terrible - mainly because I asked for 'Tintin riding a bicycle [...]' and it's having trouble with things like the bicycle spokes. Two out of the 4 'feel' right in terms of the line drawing style, color palette, foreground-background composition, level of background detail etc.
Any use of that work without permission (and thus attribution/compensation) is the problem.
Copying an artists style is legal in every jurisdiction in which Stability operates.
I've seen the collage tool argument several times, and I don't agree with it. But I can understand why people believe it.
You see, there's a very large number of people who use AI art generators as a tracing tool. Like, to the point where someone who has never touched one might believe that it literally just photobashes existing images together.
The reality is that there's three ways to use art generators:
- You can tell it to generate an image with a non-copyright-infringing prompt. i.e. "a dog police officer holding a gun"
- You can ask it to replicate an existing style, by adding keywords like "in the style of <existing artist>"
- You can modify an existing image. This is in lieu of the random seed image that is normally provided to the AI.
That last one is confusing, because it makes people think that the AI itself is infringing when it's only the person using it. But I could see the courts deciding that letting someone chuck an image into the model gives you liability, especially with all of the "you have full commercial rights to everything you generate" messaging people keep slapping onto these.
Style prompting is one of those things that's also legally questionable, though for different reasons. As about 40,000 AI art generator users have shouted at me over the past year, you cannot copyright a style. But at the same time, producing "new" art that's substantially similar to copyrighted art is still illegal. So, say, "a man on a motorcycle in the style of Banksy" might be OK, but "girl holding a balloon in the style of Banksy" might not be. The latter is basically asking the AI to regurgitate an existing image, or trace over something it's already seen.
I think a better argument would be that, by training the AI to understand style prompts, Stability AI is inducing users to infringe upon other people's copyright.
This is incredibly disheartening. Who knows how long will it take to progress the tech to the point where anyone will be able to train and run models unrestricted without dealing with lawyer nonsense.
This is really Venture Capital News, and accordingly they've appropriated the whole "hacker" image in an attempt at authenticity.
Open licenses are not the same as abolishing copyright.
https://blog.ipleaders.in/permissive-license-copyleft-possib...
https://blog.ipleaders.in/permissive-license-copyleft-possib...
Their weight is irrelevant. They would carry pretty much as much meaning in the complete absence of copyright.
In the world of sensible defaults they wouldn't need to exist at all.
> Their weight is irrelevant.
I think we're having different conversations. The only thing I'm talking about is whether terms can be legally enforced without copyright.
I'm not sure what you're getting at, but it seems to be something like "Permissive licensing is basically like public domain." I don't agree with that line of thinking because of the attribution requirement in permissive licenses, but if that's what you're getting at, I get what you mean.
By the way, it would also kill proprietary software as a concept. Source code leak? It's no longer a crime to use it. We'd never have to read licensing nonsense ever again.
Corporations R&D is great at one thing, making things cheaper to produce and thus more widely available. And that's wonderful. But actually we want corporations to steal that tech from each other because then the consumer benefits the most.
Your claim is… simply utterly untrue.
Those inventions were mostly self funded by individual inventors who in some cases had a great trouble getting the business interested in the invention at all.
Also how does velcro compare against LED, laser or microchip in terms of importance?
https://en.wikipedia.org/wiki/List_of_songs_subject_to_plagi...
https://en.wikipedia.org/wiki/Cuphead
> The game's creators, brothers Chad and Jared Moldenhauer, took inspiration from the rubber hose style of the golden age of American animation and the surrealist qualities of works of Walt Disney Animation Studios, Fleischer Studios, Warner Bros. Cartoons, MGM Cartoon Studio and Walter Lantz Productions.
I'm probably on the opposite side of the fence. I do find it disheartening that it's opt-out instead of opt-in. The training set should be limited to public domain and CC-0 until such a time it can comply with attribution; then other CC works could be incorporated.
To continue blindly applying historical understanding to fundamentally new technologies creates huge blind spots, and I’d argue similar to pretending that the creation of ever more destructive weaponry requires no changes to the rules of engagement in warfare.
The game has changed.
So many artists styles could have gone viral and actually bring those artists some work from the people who tried the AI commercially and got results that weren't completely satisfactory. Now barely anyone will ever have any contact with their art (relatively speaking vs scenario of virality).
Basically the only people who win are the lawyers and handful of artists that were mislead by lawyers primitive argumentation. Everybody else looses. First and foremost artists and art lovers but also AI researchers and hardware manufacturers.
I didn't know of any artist (except for long dead ones) that I saw the names of in prompts.
> And therefore people would abandon the AI to pay the artist individually?
If you have particular commercial needs you discovered through the use of AI you might want to go to the source, not for everything, but for some things. It might be easier to explain some of the stuff you need to a live human than to AI. AI images are still very imperfect and prompts are not easy to create.
But who knows how the numbers turn out in the long run. I don't think the artists is the group that's going to get the most annoyed in the next couple of years - copy writers and programmers, 3D artists, basically anybody who's doing grunt work..
I think the main problem with all of this is that it has entered our lives so fast, compared to other revolutions.
I think it's quite presumptuous to unequivocally state that artists lose with this. Is it a complicated situation? Yes, of course, too complicated for such certainty. That's a wonderful thing to believe, but it's just as plausible (I'd argue far, far more plausible as the tech improves) that clients who would formerly pay for their work no longer have to.
When you consider laws regarding organ donors you might appreciate how socially harmful might be wrong defaults.
With the volume of AI art generated was an even single actual case of that?
As a photographer, I can’t claim to have or require a fraction of the skills used by creators of hand-made art. And even I am not excited about some AI slurping up my best work and commoditizing it.
> So many artists styles could have gone viral and actually bring those artists some work
I’ve seen this sentiment, but it does not match the reality we see play out on the web every day.
The amount of literal content stealing and “creative reposting” that happens with absolutely zero attribution to the actual artist is quite extensive.
It makes no sense to me that the introduction of an AI tool would suddenly solve the problem of attribution instead of just make it far easier to steal content while making it harder to detect or take action against such theft.
I think I know that little better than lawyers do. Even if only because I had zero financial incentive when I formed my opinions.
> As a photographer, I can’t claim to have or require a fraction of the skills used by creators of hand-made art. And even I am not excited about some AI slurping up my best work and commoditizing it.
I know it doesn't feel great. But your art has already been commoditized. There are hundreds photographers perfectly capable of replicating your style and many of them do it completely accidentally. The value of your art is a personal element not the content itself. What's valuable is your service and the name you made for yourself.
AI gives an artist a ticket to a lottery that can strongly boost their name without doing any additional service.
> The amount of literal content stealing and “creative reposting” that happens with absolutely zero attribution to the actual artist is quite extensive.
I wonder how much money you've lost due to that. Besides, attribution is naturally built into those "plagiarist" prompts for AI.
As for prompts that don't mention specific authors ... you shouldn't kid yourself that AI won't be able to completely naturally recreate your style from styles similar to yours even it it never seen yours. After all that's what you did to create your style. You created a variation on similar styles you saw during your education as an artist.
> ... just make it far easier to steal content while making it harder to detect or take action against such theft.
steal, theft ... what do artists actually loose in those brazen robberies?
Copyright conglomerates created language that doesn't reflect reality. But it reflects most primitive human instincts evolved in the world of scarcity not abundance.
This still doesn't give you standing to speak on behalf of artists, and "because I know better than lawyers do" is generally a problematic form of argument. It continues to ignore the key people that matter: the individuals with the creativity and skills to create the content that started this whole IP conundrum in the first place.
> I know it doesn't feel great. But your art has already been commoditized. There are hundreds photographers perfectly capable of replicating your style and many of them do it completely accidentally. The value of your art is a personal element not the content itself. What's valuable is your service and the name you made for yourself.
This is a very one-dimensional view of what makes art, and how the broader community plays a role. I have no illusions about where I stand as an individual photographer among the multitude of photographers in terms of raw technical talent and capability. But I'd argue that you are deeply misinterpreting the implications of that reality, and imposing your own definition of value on a category of human expression that is by definition deeply subjective and far more complex than a simple formula of exposure and conversion rate with some resulting monetary return.
> I wonder how much money you've lost due to that. Besides, attribution is naturally built into those "plagiarist" prompts for AI.
This assumes the only reason I would be upset is because of lost sales. I take photos for the love of it. I don't currently sell them. If someone else starts making money on my work, it takes on a different meaning entirely. And even if I turned this into a business, "lost sales" is still only one of multiple factors.
Regarding prompts, how is attribution built in? Nothing requires an individual to reveal their prompts, currently. There are AI-art sharing communities emerging where prompts are held tight, because the authoring of the prompt is the only thing the "AI artist" brings to the table. Even if prompts were universally provided, that doesn't solve the issue of permission, or imply that this is automatically an acceptable form of attribution to all artists overnight.
When video game companies use stolen artwork, they are ridiculed and derided for blatantly profiting from the work of individuals. Even if it was an honest mistake, this kind of misuse is always a headline.
And yet, when we talk about a system that unlocks a seemingly limitless portal through which the life's work of every artist is made systematically available to the entire world without limit, with no consultation with the original creators, those worries about unattributed benefit just disappear.
I'm curious how you feel about the video game scenario?
Sure. That's why I don't speak on their behalf. I'm just voicing my opinion about harmful silliness of the scheme they allowed themselves to be coaxed into.
> It continues to ignore the key people that matter: the individuals with the creativity and skills to create the content that started this whole IP conundrum in the first place.
Silently ignored in lawyers arguments are all the consumers of culture. All the people who wrote the prompts and all the people who drew great joy from looking at AI creation. They'd get literally nothing if the case of strict copyright so their collective loss is great because they are many.
> But I'd argue that you are deeply misinterpreting the implications of that reality, and imposing your own definition of value on a category of human expression that is by definition deeply subjective and far more complex than a simple formula of exposure and conversion rate with some resulting monetary return.
Sure, opinions may vary. Only actual data can resolve who's wrong. And the number for compensation of artists in copyright industry are not great when compared to viral gains from attention based, open economy.
> If someone else starts making money on my work, it takes on a different meaning entirely.
If you haven't lost anything why do you care? Why do you want to devoid others of joy they draw from availability of artwork?
> Regarding prompts, how is attribution built in? Nothing requires an individual to reveal their prompts, currently.
It's the internet. People talk. Nothing stays secret. And at any point in time original artist or their fan can chip in and say, "yeah, that's exactly like mine, see?". And no-one can do anything about it.
> When video game companies use stolen artwork, they are ridiculed and derided for blatantly profiting from the work of individuals. Even if it was an honest mistake, this kind of misuse is always a headline.
Yeah. Using AI artwork in a very specific style would generate same kind of news. And those games are not very good and they don't make much money. So not only there's no harm, severe ostracism. There's also not much opportunity to have any gain if copyright was strictly observed. And as you noticed it already happens. Darkest Dungeon had very fresh and attractive art-style. Now it's very easy to randomly encounter in Play Store cheap clicker games blatantly ripping off that esthetics. It's not directly stolen, but it's pretty much what AI would do if someone was hell bent on replicating the esthetics. Yet humans did it. What's the loss to Darkest Dungeon graphic designer? Exactly zero.
And I think games that would start with AI art and got really popular ... they'd invite original artist for DLC or to get on the news or for just good will of the public. Public relationships is very important when selling games.
Fame has a very short half-life and unless you have all the licensing/contractual machinery in place beforehand, you probably won't be able to cash on that boost. The line of thinking you articulate here is extremely familiar to anyone who does creative work. It's the same argument that producers use to get people to work for free or cheap on films, that broadcast or streaming services use to justify very low payouts to content creators, that commercial commissioners use to try and get art for free etc. The creative field is absolutely full of promoters who offer to match artist to audience, with the promoters getting the first cut of ticket sales and the artist getting the last or none.
https://theoatmeal.com/comics/exposure
Besides, attribution is naturally built into those "plagiarist" prompts for AI.
Only if you are already kinda famous. Suppose you have a distinctive visual style that's a great fit with a genre, like ghost stories. I, an unscrupulous publisher, note that the market for ghost stories is currently booming and decide to buy, or perhaps generate from AI, and bunch of mediocre ghost stories, and then publish them with 'art in the style of scotty79.' I make a little app offering 'best new ghost stories every day!!' for $1, put it in app stores, and make $7 million before the ghost story fad runs its course. You get nothing, and consumers who got familiar with your style by paying $1 or looking at ads to use my app don't care about you because I never gave you credit and in their mind the style is associated with Best Daily Ghost Stories, not you.
Maybe a few of them will do the work of combing back through the history of the fad and to find which artists influenced the 'daily ghost story' aesthetic. Maybe this will lead to a revival of interest in your work even though the fad it was associated with has come and gone. Good luck with that.
The dirty secret of the creative industries is that if you don't get paid up front for your contribution, you will probably never get paid at all.
I'm not sure how did you manage to miss thousands of artists able to capitalize on sudden and accidental fame for decades without any prior arrangements. I'm not saying it's easy. I'm saying it's possible. Also to put this in context compare this with how often very popular artists get completely screwed by huge copyright behemoths earning a score of money but mostly for someone else, someone completely uncreative.
> https://theoatmeal.com/comics/exposure
Trying to buy something for exposure is absolutely abhorrent because you try to coax someone into doing work for no money. And for nothing basically because people who try to pay with exposure don't really provide any significant exposure 99.99% of times.
If AI people were forcing artist to create new art and paying them with the promise of exposure in their generated works I'll be completely on your side. However it requires zero work from artist to have their already published work to be used as learning material. That's why they never opposed it when other artists were learning from their art. That and of course that target of their wrath would be basically the entire rest of the art community which wouldn't make them very popular.
> Suppose you have a distinctive visual style that's a great fit with a genre, like ghost stories. .... I never gave you credit and in their mind the style is associated with Best Daily Ghost Stories, not you.
That's completely fine in my book. And if those generated stories get really popular so I learn about them I might do just a little bit of online marketing to inject my name in the discussions about them and publish new ones to my fresh new subscribers ahead of time. Heck, I could create my own generated and fine-tuned manually content and sell it just like that guy does since he's already proven a business model for me.
Compare now this with the world of strict copyright where this guy doesn't even know I exist, same goes for swaths of fans of ghost stories. Or let's assume he knows and wants to deal with me. Since he's the one with the money I'll be severely dependent of him and strongly disadvantaged in any deal. But let's assume we struck a deal that's nice for me. There's no way I'll be able to produce new ghost stories every day. Not to mention I wouldn't wish that workload on my worst enemy. So no business happens and many people have their love for ghost stories un-satiated, many didn't discover their love for ghost stories and I am 100% still poor struggling author who's known by nobody.
World without copyright and zero publishing cost is the one where authors and consumers are in control and negotiate through attention economy. World of copyright is the world where copyright hoarding dragon starve both artists and consumers.
> The dirty secret of the creative industries is that if you don't get paid up front for your contribution, you will probably never get paid at all.
And yet you vehemently defend the system that created this situation and refuse to even consider alternatives.
This seems like an odd way to frame this. The reality is closer to "artists were never included in the conversation to begin with". Arguing that "no one forced them to create anything new" seems irrelevant when you consider that without the content, none of this exists to begin with.
The problem is the assumption that artists are or should be universally fine with this.
> World without copyright and zero publishing cost is the one where authors and consumers are in control and negotiate through attention economy. World of copyright is the world where copyright hoarding dragon starve both artists and consumers.
If you want to argue against copyright, that's fine, and I have plenty of issues with the current iteration of this framework of rules. But that is not the same argument as "Tools like Stable Diffusion aren't infringing because xyz technical reasons".
I think it'd be helpful to be clearer about arguments for/against the spirit of the rules themselves vs. arguments about why generative AI tools do or do not create content that infringes those rules as currently designed or require an entirely new framework of thinking about the problem.
They are important but distinct problems.
I don't think anybody assumes all artists will be fine with it. But being disgruntled doesn't automatically mean you should be the one who makes the decisions. Virtually every human has a stake in this because nearly everybody consumes some art. It's time we put stronger emphasis on the rights of everybody else, instead of just mostly people who bought copyright, and artists those people think have the right to exploit.
> "Tools like Stable Diffusion aren't infringing because xyz technical reasons"
I don't think I'm saying there's some technical reasons those tools aren't infringing. Just overwhelming moral, practical and economical reasons that they should be allowed to operate and treated just like human artists who too can mimic and mix styles and nobody can deny them authorship unless they copy specific complex elements nearly verbatim.
Btw if human artist think AI got to close to one of his works he can prove it beyond all doubt by registering their creations on some blockchain with his key and timestamp and I wouldn't be against banning this specific AI artwork that got too close.
Banning use of all published art as teaching material by default is not the way to go in my opinion. Banning imitating specific style is bad too. Imagine portrait painters banned photography from using classical portrait compositions to protect their jobs. Or denied photographers even looking at portraits so they can never learn to imitate.
> I think it'd be helpful to be clearer about arguments for/against the spirit of the rules themselves vs. arguments about why generative AI tools do or do not create content that infringes those rules as currently designed or require an entirely new framework of thinking about the problem.
Maybe, but AI is the wonderful opportunity to discuss the spirit of horrible rules we currently have. Best opportunity we had since creation of social networking sites that chipped those rules away a bit.
Yeah, it's disheartening. There's also no good way to fix it; the cost of storing copies of their art is negligible, and AI trains the same whether the material is copyright or creative commons. If you get Stability AI to omit your art, then Unstable Diffusion will be trained on your likeness. Opt-out of that one, and some guy in Nevada personally sponsors a bespoke model for making copies of your art.
So, I agree with the parent. The most tragic part is not the short-term fight, it's the long-term consequences. Artists will have to internalize what software developers realized decades ago; creating takes work, and copying is free.
This is very dismissive, the scale is what makes the difference. You can get away with pirating all types of content for personal use.
AI companies are essentially trying to legalize that, but in reverse - taking from small creators to enrich the shareholders of their billion dollar corporations.
If you try to play by these rules and create a million dollar business that distributes various copyrighted content from the internet (e.g. AAA games, movies or music) you'll very quickly realize how much their stance on copyright changes.
Power ultimately belongs to whoever can wield it with the least friction. It's not fair or right, but it's the way they play in the business world.
People are allowed to view private art, draw inspiration and ideas from it, and execute on those to create new things.
Why should we limit AI any differently?
If the end result is too close to the original - apply the same guidelines you would for any other artist who copied your work.
Otherwise... you're not allowed copyright over a particular style (for damn good reason). While I would like to see artists retain some form of revenue, I don't think this is really the most pressing issue on that front.
What we need is enough computation power to run these models on our own computers, on our phones even. Then we'll be able to do whatever we want and there's nothing they can do about it.
The technology isn't.
The content is.
This is the crux of the issue for me. It's a different set of rules for AI companies than everyone else. If I started selling pirated copies of Nintendo games they would send an army of lawyers after me and this "opt-out" reasoning would not be a valid defense in court. These AI companies are trying to get away with stealing art and other content with a simple "whoopsie, we promise we won't do it again" when people demand that their own rights be respected.
It's not copying, it's breaking down work to it's foundational features and recombining those features with others to make new things. Literally exactly what humans do when they make art.
If a person was doing what these models are doing it not only wouldn't be illegal, it would be laughable if we even had the discussion.
It doesn't matter what the law says or what is "right." It comes down to who has the power and who doesn't.
(The other difference between what humans do and what AIs do is a matter of scale. A human imitates by spending many hours to duplicate a work of art. An AI can churn out millions in a second. That's a separate issue, though.)
Current trajectory will only harm original creators, there is zero consideration or benefit for them. On the other side you have companies that stand to make billions off of their work.
Arguing in favor of such a system is, in my mind, appalling. Either ensure that copyright law prohibits unauthorized machine learning from valuable original art or abolish it completely.
First, whether particular features like watermarks end up in the end product or not is not terribly relevant. The particular model implementation that produces that behavior doesn't understand that watermarks are not a desirable part of the transformed end products, whereas humans do. In that way, the AI is certainly a little more 'honest' about what it's doing. It would be trivial to make the AI understand this, and stop reproducing watermarks.
You claim that the model performing this work is somehow different because AI doesn't 'think' about problems, as evidenced by strange artifacts, whereas presumably humans do think about them because of the absence of these artifacts. I'm curious why this makes a difference. If a model was sufficiently advanced that you were unable to differentiate a painting made by a human along certain themes and an AI along certain themes would it be somehow less objectionable? Why or why not? If the end product is of identical quality why should it matter the route you take to get there, in the eyes of the law?
You bring up scale, but scale is also not relevant. Say that I create a school devoted to training legions of artists to produce paintings in the style of a particular artist, while maintaining a transformative aspect. Is this illegal because I'm doing it at scale? No. If that's not illegal, why is it illegal to do it with code? Because it's more efficient? In what other domain is producing creative work more efficiently by transforming existing work illegal?
How's that different from a human artist shackled to a company by some secret agreements in which he's not the stronger side?
No, it is transformative just like a computer would. A person would be inspired and create some kind of unique work. A prompt would - given identical start-up conditions - result in the exact same image over and over again no matter who entered the prompt.
Even a trained monkey giving the prompt would result in the same output, as would 'thumper' or the closest local equivalent.
The ends of having an useful model like stable diffusion doesn't really justify just ignoring the IP rights of tens of thousands of creators who were already having a pretty rough time making ends meet. That's just a shitty thing to do.
Copyright law isn't friendly to small creators, and big creators use it as a cudgel with absolutely no consequences.
This reminds me of the backlash against the wacom community on deviantart in the early days.
The model learns concepts from images, not the images itself. It has developed general solutions explaining light, colors, composition, objects and their relation to one another, facial features and too many more concepts to even begin enumerating them.
How is this different from a human studying art, literature, music, etc. to learn concepts and then apply them in creating new pictures, novels or songs?
Further, just because we can talk about how artists, at a high level do the same thing as AI image generators, the actual mechanism is not exactly the same and is therefore still subject to distinct regulation.
Even if you were able to somehow establish that computer programs should have the same rights as people (since they are made and used by people,) you’re still not out of the woods. Much debate remains about what creativity and originality means when talking about human generated content in an IP sense, and adding the programmatic aspect doesn’t simplify things. (eg The Sina Qua Non of Copyright is Uniqueness, Not Originality https://tiplj.org/wp-content/uploads/Volumes/v20/v20p327.pdf)
So? An excavator clearly isn't a human, but it digs holes in the ground by the same principles that a human using only his bare hands would, only faster and more efficient.
In Washington you must dig for clams with human power, you cannot use a hydraulic clam pump. There was a time where you could down 50 birds with a single firing of a punt gun. This was made illegal, even if you could still eventually bag 50 birds with a normal shotgun.
I can setup stable diffusion on a computer, I can put a brush next to a roll of canvas, or hammer and chisel next to a block of marble. Neither of these setups will create art on its own.
These are orthogonal issues at this point.
The one concern I do have is that the “lawyer nonsense” (read: AI companies playing fast and loose with current laws) will stack the regulatory deck against AI technology unnecessarily - essentially because of an unforced error that brings negative attention to the technology.
Put another way, these companies are asking to have a spotlight put on them by being so flippant about copyright and ethics issues. This spotlight could have been avoided with better behavior, and the tech would still appear magical and remain one of the most impactful jumps in tech in decades.
The computing requirements enforce the current reality that the training of models will be centralized.
This places a larger ethical burden on those central entities, IMO.
It's an area where there are no existing laws. We're not going to stop AI because some furry deviant art artist complains loudly online.
Whether or not one believes those laws apply to generative AI seems to be based on one's belief in how similar that AI software is to humans.
I'd argue that systematically ingesting 2.3 billion images is not remotely human (one of a myriad of reasons the comparisons break down), and that it is a long stretch to claim that this falls into the realm of fair use as originally envisioned.
It is this insistence that the software is human enough to be granted human-like status that is playing fast and loose with the definitions of things, ranging from consciousness, to learning, to the interpretation of those concepts relative to current laws.
I believe new laws will be written, and old laws will be updated. There's no question that the current legal system is not well equipped for various generative AI systems. But I don't think the current laws have nothing to say.
And I'd still argue that this conversation can be separated from the one about indiscriminately slurping up artist's content.
> We're not going to stop AI because some furry deviant art artist complains loudly online.
Please don't argue against straw men. There are legitimate concerns from artists across disciplines and genres, and this isn't just isolated shrieking.
Artist backlash is frankly one of the most natural outcomes I could imagine from a system that uses their work without permission. Many of the people who are complaining loudly are not against AI, just against the use of their work without consent or attribution.
I'm both extremely excited about the possibilities the software unlocks and concerned about the implications. AI can exist without ignoring the rights of artists.
There's implicit assumption that if you can get a hold of a copy and manage to learn from it you are free to use what you learned in your creations.
But there are explicit laws about how you acquire copies of things, and whether they apply seems to be based on what someone believes “learning” to be.
Your claim relies on the belief that a computer ingesting images is similar to a human learning from those images.
In the United States, a legal derivative work, which isn't a parody, needs to make three substantive changes from the original. It's fair to say that creating new works in the same style or 'look-and-feel' of an artist would satisfy that at prima facie.
The price floor on art commissions is already very low and AI effectively makes that cost zero, while providing zero compensation to the thousands of artists. Without their work, there's no Stability AI. From an ethical standpoint Stability is in the wrong, and from a legal one I think the class has a very strong case to recover damages.
The aesthetic subset is .05% Artstation:
https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/im...
Not sure if that's a large amount or not. They could've used robots.txt if they didn't want to be indexed.
That dataset only covers "aesthetic" clip terms as well. Not to mention a lot of images come from Pinterest and other aggregators.
[1] https://waxy.org/2022/08/exploring-12-million-of-the-images-...
I can see see a future dispute arising over outpainting (beginning with an existing copyrighted work) but there infringement and identity of the infringer (the user, not the toolmaker) is more clear.
Stable Diffusion is equivalent to hip-hop sampling in the 80s and 90s. The outcome is obvious.
Are there specific similarities that make you believe these are equivalent scenarios? Not just “it feels thematically similar”.
Hip-hop originally recorded and transformed vocals, instruments, and beats to create something new from pieces of something old. The practice occurred without permission and obviously ended up in court. Now sampling requires a licensing agreement. The additional cost has fundamentally changed the genre (over the last 40 years).
Hip-hop and tech both ignored IP rights because neither started with a legal framework and both would have found the additional cost prohibitive.
If I'm understanding you correctly, you see the similarity more in how the initial side-stepping of copyright eventually gave way to new licensing rules (or adherence to existing rules).
I've heard similar sentiment trying to make another point entirely - something closer to arguing that the AI is creatively inspired the way humans are, and therefore is by definition not infringing.
I suspect this might be where the flurry of downvotes came from.
The systems are similar too despite having completely different internal processes. Both transformed existing IP without permission, to produce sufficiently remixed art as an output. A sufficiently generic abstraction would look very similar despite the disparity of domain.
The primary difference between hip-hop and Stable Diffusion is that AI cannot rationalize, explain, or attribute inspiration to a final product.
There was no aha moment and thus no creativity. It has no vested interest in its work.
Remember, input is fundamentally required. Without that dataset, Stable Diffusion delivers exactly nothing.
That's true of lots of things that are legally considered transformative though. E.g. without all those books that Google scanned, Google Books would be completely useless.
More moderately, all art is derivative at the end of the day. None of it was created ex nihilo. We already have legal guardrails for direct reproduction of specific characters and specific pieces, and that is plenty. The fact that maximalists want to kill a nascent technology by restricting the right to learn from, something that even our extremely slanted laws have carveouts for, is nothing short of offensive.
If it is art, then it will be made both by patronage and by people saying: "Hey StableChatSiri, make me some art."
(People will still do the later even if it doesn't count as art).
Hm. Thinking of patrons, does anyone know how many artists and sculptors there were in 1710 England?
I keep programming computers just because I like it. Maybe they'll keep creating too for the same reasons. Maybe they won't. It's irrelevant either way.
The nihilistic feeling that it's irrelevant whether new art is created kindof proves my point.
Because when I have access to these tools I will make better art than the technocrat with access to these tools?
You’ve got a lot of knots to untangle.
The use of including copyrighted materials in the trained model was a choice, not some obvious fact about the nature of AI. All of this could've been avoided if the data set did not include unlicensed work in the first place.
If your AI is limited to only training on the paucity of explicitly permissibly licensed/public domain content (and as I think about it more, this would only apply to things that are permissibly licensed without an attribution requirement, which is something there is no meaningful way to do in a model like the one we are discussing) your AI will not be very useful. With that in mind, I would argue that yes, it absolutely is an obvious fact about its nature.
Best outcome in my opinion would be for the output to be judged on a case-by-case basis, like human works are, not for machine learning on data without "proper authorization from the owner" to inherently count as infringement.
How does this work? Do they retrain the model from scratch every week? Or is it somehow possible to retroactively remove specific training-set items from the already-trained model?
The Disney protection act rears its head…
As noted in OP, this is an outstandingly bad definition of Deep-Neural-Networks, and the lawsuit should fail when the court hears an explanation from any competent practitioner.
However, a correct definition would make the lawsuit far more interesting, imo. Diffusion models can be compared to a superhumanly talented artist that can be cloned in unlimited fashion by anyone having the software and hardware means. How does this entity affect social well-being, how should existing laws be modified--if at all-- with the welfare of humanity in mind, etc?
So how often does this happen? Somehow I'm too cynical to believe that a judge would rule against the intellectual property industry. The whole thing is based on absurd concepts to begin with, concepts that can be reduced to the ownership of unique numbers. Once a society accepts that, what difference do explanations make?
How can you claim with a straight face that this is a better explanation of what an NN is?
An NN is simply an approximation of a multi-valued function, whose parameters are adjusted by minimizing the difference between the output of the NN and the output of the real function for a certain input. It is much much closer to "a giant archive of compressed images being used to interpolate between them" (though it's not that) than it is to a "superhumanly talented artist".
Right, but that equally fits a biological NN if you zoom in that close. You'll need more than wikipedia to appreciate what deep-neural-networks are doing here, it's dimensional space that's key. What DNNs do that is similar to the human brain is that they order "concepts" in high-dimensional space. Colors, textures, shape and hierarchies of same are organized and cross-referenced with text in an incredibly complex connectome. It would be useless to memorize images with their textual descriptions as that would be horrendously inefficient/ineffective during inference. Rather, the model must do what we do and understand what makes an image a "landscape" or a "portrait" or a "cartoon". It needs to understand what is an artist's style and how to perform it on a work never before created.
"Understanding" can only mean ordering meaningless letters and pixels in multidimensional space so that they line up with human understanding (and human 'understanding', in turn, can only mean ordering meaningless sensory perceptions in the brain's multidimensional connectome such that reality turns out to be approximately predicted and controlled). The only systems that work this way efficiently are neural networks, biological and artificial.
First of all, we have no idea how biological NNs learn, how they represent information, how they reason etc. Given what we do know, there is no reason to assume any similarity with ANNs on any of those fronts. Just to give one example, we know very well that a single biological neuron encodes significant information and is capable of reasoning on its own. In fact, even non-neuronal biological cells are capable of such - especially looking at single-celled organisms, which display extraordinarily complex behaviors with no NN in sight.
Second of all, we don't exactly understand how the huge models we have actually encode the higher-level representations of the training set that they store. Of course, we can say for sure that they are not literally storing a copy of the data on simple space requirements. But we can also say for sure that their "understanding" of the data, as well as their capacity for inference, is significantly different from our own - since they make certain mistakes that are nearly impossible for a human to make, while showing super human abilities in other aspects. So, if anything, we must conclude that whatever it is they are doing, it is most certainly not a way of understanding the information the way we understand it.
What is laughable is that these abilities could come from interpolation or collage (not your claim but the plaintiff's). The only way these abilities could occur is if Stable Diffusion can represent image and text very similar to the way human brains comprehend them. The argument here is simple: what are the odds that StableDiffusion/DNNs have hit on a representational method that is totally different from human brains yet yields the same recognition, praise and admiration for the artist from everyone who sees it? Seems to me close to 0.
I very much disagree. Pretty much any halfway decent artist (say, anyone able to at least caricature recognizable people) is able to produce this kind of imitative art, when/if they are aiming for this type of copying. I've seen nothing coming out of SD that I couldn't expect to find on DeviantArt, at least if I commissioned it specifically. Most human artists of course don't do this, since people usually don't like reproductions (apart from posters) or copying others' style. Note as well the huge problem SD has with consistent fine details (especially text, but also often hands and even faces).
Of course, I fully agree on the speed factor (and would add scale/cost) in which SD without a question is far beyond humanity, obviously. That is a meaningful difference that is very likely to affect markets like decorative prints and other low-value art.
Perhaps this basic disagreement (which is of course ultimately subjective, unless someone is going to do a blind taste test) explains the difference of opinions for the rest of the points.
Now, sure, winning just 1 contest isn't going to settle the matter but I think it provides reasonable evidence that SD is heading to elite quality at a rapid pace. Above, the submitter Allen was responsible for directing and cleaning the results so deserves credit. However, Allen is also using a year-old MidJourney model that has likely improved dramatically already.
It seems uncontroversial that even if SD models aren't in the top-tier of imitative/derivative art (or "art from text" as that genre evolves) right now, they will be soon.
Now, that could work out. Major movie studios and recording companies do file copyright registrations and submit a deposit copy. But few others bother. It seems that you can send a DMCA takedown request without a copyright registration, but you can't enforce it in court without one.[2] This raises the question of, if you as a service receive a DMCA takedown request, should you ask the requestor to send proof of copyright registration, and if they don't, ignore the request?
[1] https://www.copyright.gov/registration/
[2] https://www.traverselegal.com/blog/is-a-registered-copyright...
That would mean that the vast majority of artwork posted online is essentially free to exploit in the USA, since I’m sure most people do not routinely register their works with the copyright office before posting them.
This suggests an online process which looks like this:
* US Service provider offers web page for DMCA notices.
* Web page requests that the user enter copyright registration info.
* If user fails to provide registration info, web page offers links to various national copyright registration sites to register a copyright. A payment receipt for copyright registration is acceptable as temporary proof of registration, but must be followed up within some period of time by actual proof of registration.
* Temporary proof of registration is enough for a takedown, but the material will go back up if full proof is not submitted later.
This would put a big dent in nuisance DMCA claims. The service provider might get sued occasionally, but for big providers, it's probably worth litigating this once or twice. The companies that have valuable IP file copyright registrations. Disney will be able to show a copyright registration on all their movies.