AI-assisted art – A mixed decision from the copyright office
processmechanics.com
processmechanics.com
I think this is shortsighted; this opens up a space for "AI copyright trolls" who generate images for popular prompts in an automated fashion to get copyright, then go after people using AI art who happened to hit the same seed and prompt. Admittedly unlikely, but it could happen, and might eventually even be worth the GPU time depending on how popular AI artwork becomes (and how fast GPUs get).
In any case, I don't see why the unaltered, or in this case extremely minorly altered, image output itself should be copyrightable. It's like two people going to Venice Beach to record the waves at the same time. They each have copyright over their specific recordings, but they can't copyright the sounds of the waves itself; the other person is free to do what they want with their own recording of the same sound. The same way that if you generate a Midjourney image with a specific prompt/seed, I should be able to use the same settings to generate the same image and do what I want with it.
That seems extremely unlikely to make any sense given that the random seeds in question would be, what, FP32 or FP64 by default? So that's 4 billion random seeds per prompt. Better hope your GPUs are really cheap and you're targeting a really popular prompt+model+everything-else...
Why not copyright them all?
Sure, it’s irritating, and troll behaviour; but that’s literally what patent trolls do right?
If I wanted the look of my product to be "ball point pen on post-it note" it would have been a lot less work and effort and it would get copyright protection.
Effort is the wrong measure.
Sorry if I’ve missed your point here, but the purpose of copyright in the United States, as written in its constitution, is to promote the progress of science and useful arts[0], not to protect the interests of creators. The exclusive license to creators is a means to an end.
[0] https://en.wikipedia.org/wiki/Copyright_law_of_the_United_St...
That sounds like protecting the interests of human creators so they keep creating, to me at least.
Even as someone extremely pro-AI art, I don't think AI-generated content should have copyright in the usual sense. Every art model was trained on massive amounts of copyrighted data, it would be very hypocritical to suddenly claim strong copyright on the output of AI generated items.
I think AI generated copyright should be 'pay-to-register', people can pay say $5 per image to register their image, if they believe it to be valuable enough to be copyrighted. In this sense, its more like trademarks, rather than the automatic moral right of copyright.
Your average AI art (which is worthless) will not be copyrighted. The ones with significant amounts of attention, post-editing, and economic value, the AI artist can simply register at the copyright office.
The extra funding for the copyright office, will also give them more staffing to deal with the torrential tide of AI-related copyright issues, and encourage them to build a healthy ecosystem where manual art and AI art co-exists.
However, even minor changes (such as changing the tint of the photo) have been found to be enough for copyrightability.
I don't think past-decisions are that important here. There'll definitely be new laws around AI generated content, that set clear standards. Those standards have to consider the social impact of copyright and the livelihood of artists. The whole point of copyright is not some abstract golden moral standard, but something that keeps artists and writers employed.
That's why I think pay-for-copyright is the best balance in terms of liberty for AI art but preserving the incomes of traditional artists.
The generative art is not set in fixed pixels, but instead is represented in a fuzzy form (equations, algorithms etc) that slightly varies each time you look at it. The value of it is also constrained by the author's intent and work. Only humans can be subject to copyright, so it at least makes common (if not legal) sense.
ML models are the same as the generative art that previously existed - they only add the captured essence of the prior data to the mix of human artistic intent and hard algorithms.
Even leaving aside that some people think coming up a prompt or choosing from multiple generated images is itself work enough to justify copyright you'll have so many very similar images registered that copyright trolls could still intimidate people into forking over settlement money and people would still argue that even the smallest changes (say to brightness level/saturation) justifies their copyright.
Leaving AI generated works free from copyright would be ideal, since it'd free those images for others to build off of and remix and reuse in new ways, it encourages commercial projects to hire human artists so that they can gain the protection of copyright, and it doesn't stop anyone from doing what they want with the technology. Comic book authors who illustrate their work using AI can still copyright their stories, trademark their characters, etc.
How is this relevant to AI copyright?
If AI art is copyrightable, they'd still need someone to generate the AI art. That person is an 'artist' in the legal sense, regardless if they can draw or not. If AI art is not copyrightable, they can still hire a non-artist (fully outsourcable), to do some minor post-editing of the AI art, and copyright it.
I'm arguing that they shouldn't be able to do that either. If a company wants to create something protected by copyright they should hire a human to do it. If they don't care about the copyright, then they can take advantage of AI generated art and everyone else can use the result for whatever they want. It makes it easy. If an AI made it, no copyright. If a human made it, copyright. There's no bickering over whether or not someone spent enough time post-editing or coming up with the perfect prompt. Artists stay employable, more artistic works gets into the public domain and there are no restrictions on how AI generated art can be used and no opportunity for copyright trolls to abuse the system. Everybody wins.
Nobody will ever know if a human was involved or not.
Removing AI generated works from copyright entirely means that if you can demonstrate your work was created by AI you wouldn't have to worry that it violates someone else's copyright. People would be free to use the technology for whatever they wanted as long as they were willing to give up the idea of copyrighting whatever the output was.
It's really not helpful to try and define what is an AI image and what is not. What if an artist repaints the background, or photoshops out an extra finger, or just re-colors some icons.
Whats different about using text in Midjourny and a brush in photoshop. What about if you use both on every image?
When using AI generated images no one should be an infringer since copyright shouldn't apply to AI's output. People should be able to use AI generated images already created without fear of infringement, but you're right in that it requires you to be able to identify a work as being AI generated. Edit's to an AI generated image shouldn't make it copyrightable so photoshops and recolors wouldn't matter, but repainting an AI generated image entirely by hand could arguably give an artist the ability to copyright the specific reproduction, to the extent that if I repainted the Mona Lisa by hand I should have copyright protection for my re-painting.
The difference between using Midjourny to create an image and using a brush in photoshop is that the line generated by the brush tool in photoshop wasn't dependent on countless other copyrighted works while the image generated by Midjourny was.
Since copyrighted works were exploited in the creation of images created by Midjourny in a way that makes it impossible to credit or compensate the authors of those works we can resolve the issue entirely by preventing any of the art it creates from being protected by copyright. Artists whose work was used to train AI can rest assured that no one is exclusively profiting off of their work without compensation. The sum of all art created by human hands throughout history goes in, art free for all of humanity to use without restriction comes out.
Anyone can use both Midjourny and photoshop to make an image but the resulting image wouldn't be able to have copyright protection, still, the image wouldn't be at risk of being accused of copyright infringement either. That frees everyone to use the powerful new AI technology to create artistic works without fear, but also without the protection of copyright.
Once again though, that'd require us know when an image was generated by AI even if edits were later made after the fact which may or may not be possible causing some people to get protection for works when they shouldn't, but even that would mean people who could prove their works were AI generated would be safe from being accused of infringement which solves one the big legal problems AI images have caused currently.
Companies have repeatedly been caught suing people for violations of copyrights they didn't actually own while paying zero consequences for it once they were caught. For examples see https://www.techdirt.com/2020/03/26/riaa-realizes-it-sued-ch... and https://blog.ericgoldman.org/archives/2010/03/viacom_v_youtu...)
I'm not saying this is an unsolvable problem, but it will take actually imposing severe and meaningful consequences for companies who lie.
If you want to read more about, it I recommend Legal Handbook for Photographers ( https://www.amazon.com/Legal-Handbook-Photographers-Rights-L... ) and note that this isn't just a random person writing a book...
> Bert P. Krages, Esq., is an attorney who specializes in intellectual property. He is the author of "Handbook for Photographers" and "Heavenly Bodies "and Photography: The Art of Composition"
---
Registering works is "pay to register" and has been for some time. Though you can do it in bulk.
https://www.copyright.gov/registration/visual-arts/
---
My takes on your examples:
Img2img : The AI created art is a derivative work of the source image. By itself it is not copyrightable as there is no human creativity at work.
Txt2img : The text prompt is copyrightable, though that's rather boring. The AI art generated by it is a derivative work of the text prompt and is not copyrightable.
Post processing : Just as DuChamp's L.H.O.O.Q. was copyrightable in its day and the source material for it was not (the work of an old master), so to the human modified image that was original AI (and uncopyrightable) is now a work that can be copyrighted.
Note that these are my interpretation of current copyright law and (to me) seem rather reasonable. See https://www.copyright.gov/comp3/chap300/ch300-copyrightable-... sections 306, and 313.2. Changing this is a change in the law (e.g. need congress to do it) rather than a change in how the copyright office interprets the law.
Copyright office funding : https://www.copyright.gov/about/budget/2022/house-budget-tes... aside from managing the capital request to implement the CASE act, they are already pretty much self funded from royalties and fees.
Consider also the painting "Edmond de Belamy" and the surrounding debate. https://en.wikipedia.org/wiki/Edmond_de_Belamy . While this isn't the final say, you can be sure that lawyers have looked at it - https://commons.wikimedia.org/wiki/File:Edmond_de_Belamy.png and note the licensing section and its link to the UK copyright office - https://www.gov.uk/government/consultations/artificial-intel... ... and that copyright in the UK is based on sweat of the brow (which was rejected with Feist v. Rural Telephone Service in 1991 in the US to instead apply original works of authorship)
I like this idea so much, it should be extended to all copyright.
Not only does it make very clear when something was "published", we could have a central repository of copyrighted works you could search and review.
All work published after 2025 is public domain unless resisted with the global copyright office.
AFAIK the opinion was that that doesn't count as copyrightable art.
If I go to a human and ask them to draw me an image, I will iterate and collaborate with my prompt just as much if not more than I would for an AI generated image. I'll look through multiple pictures and point out things I like and dislike. But I won't get joint copyright over the final image unless the artist gives me a contract assigning it. We recognize that collaboration with a human to describe a final image isn't something that usually falls in the narrow range of copyright.
So the argument around prompt generation seems like it has much wider implications than most copyright-expansionists are saying. I don't understand how to grant AI images copyright without granting a bunch of other stuff copyright too. And traditionally, we don't think of commissioning as a copyrightable act, even though it arguably has very similar elements of creativity that are being talked about here.
Is there a creative human input into an AI-generated image that isn't present when commissioning or working with a human artist? Because otherwise we're talking about a frankly massive expansion of copyright that should probably be approached with a lot more caution. I mean, some of these arguments I see for granting copyright are getting really close to outright saying that deciding what to draw should be treated as creative enough to warrant protection. That's a wild thing to say, that has so many implications beyond just AI images.
It's not that different to a company paying artists.
In a way that's even worse; building a tool that someone else uses to produce a creative work shouldn't grant the original builder copyright over the user's output. That would also have a ton of implications beyond AI.
It's bad enough that many software tools come with license agreements around their usage that reassigns copyright and restricts output, but at least in those cases the agreement rests on a license that the user is signing to get access to the tool. But we wouldn't claim that the person who's made an artist's paints owns the painting made with them.
> The website is fairly circumspect around the whole issue as you can imagine.
I remember it making the news when one of the companies in the imagegen space (I don't think it was Midjourney, but it might have been) said that users would keep copyright on images they made, and I remember similarly thinking at the time, "well, that's very nice of you but I'm pretty sure that's not your decision to make."
But agreed, I think that the company is probably not going to go out of their way to really clarify how much IP control they think they have over what other people do with their tool.
----
There's implications beyond AI to saying that AI images are copyrightable, but I think they're small potatoes compared to the implications of saying that inventions confer copyright of artwork made with inventions back to the inventor, even if someone else was using the invention to create the artwork.
When studios pay artists, they don't get copyright, the artist does. The studio gets copyright if they think ahead and make the artist sign a contract to assign their copyright over. And negotiation over assigning copyright is part of the payment process and contract process, the studio doesn't get that for free.
It feels like we need some distinction that's a little more qualitative than quantitative.
That's not how the diffusion process works. You can pick any number of interesting ways to describe it but if they're technically wrong, it doesn't really matter how poetic they are right?
Diffusion models do use random noise.
As I understand it, every 'step' is composed of three parts: a) the previous output, b) the latent generated from the prompt and c) random noise.
As you move further up, the scheduler changes the weights of a, b, and c that get mixed in.
...but from the article:
> The subtle error comes in a misunderstanding about the "randomly generated noise."
It's not an error. You're just focusing on what you want to focus on.
Let's be 100% blunt: The author of an AI art image is pressing the random generator button. Every time. The output is random.
It's not a matter of debate; the initial seed to the diffusion model is random noise.
The prompt guides the diffusion process, which basically denoises the random noise added to the image certainly... but saying there's no random component to it is completely and utterly wrong.
Perhaps that first sentence could be more precise, but by the end of the paragraph the author’s meaning is clear: the court has a misunderstanding about the “randomly generated noise” when it believes there is randomly generated noise in both the pixel and the latent - this is not the case, there is no randomness in the latent, that exact handcrafted prompt picks out a precise spot in the model’s giant table of embeddings, that prompt will always pick out that spot in that model, and the random noise is only on the pixel side of things. The author believes the court has this misunderstanding because the court uses the analogy of “a patron makes a suggestion to an artist”, which is a scenario that DOES have random noise involved in producing the latent (the brain is an inherently noise place; an artist’s brain likely even more so).
No.
This is factually incorrect.
The random noise is applied in the latent.
The VAE doesn’t add random shit when you convert to pixel space.
Having randomness in pixel space would make no sense at all.
You don’t ever pick a deterministic point in the latent space (unless you fix the seed to the random number generator, and then you're still picking a random point in the latent space, unless you're prepared to argue you somehow know what a specific seed is going to do before you use it... you're just saving the point for later).
Ultimately, it comes down to this:
- You have a function that generates a bunch of random images.
- You pick one.
Did you create the image? No. You didn't.
Did you apply creativity? Should it be copyrightable? Maybe? You picked the one you liked most out of a set. You certainly applied your sense of aesthetics.
The practical question is:
What stops someone generating every possible image and copyrighting it?
Sorry! I know you typed that prompt out and got a random seed, but it turns out I've actually copyrighted every image for that prompt for seeds 200000000 - 300000000. It's only a 100 million images. Easy.
Right? Who cares if it's random or not? (I do, pedantically, it is), but practically, the copyright office doesn't care. They care about the practicality of preventing the system being abused.
We can argue about the semantics of where the noise is applied, but it doesn't actually matter.
How do you support people by letting them copyright their 'human scale' generated content, but avoid abuse from trolls who apply 'industrial scale copyrighting' with the same process?
For stable diffusion, you can actually just set the seed to a fixed number.
After that you can always get the exact same image for the exact same prompt.[1]
In general, there are some interesting philosophical debates to be had about pseudo-random number generators.
[1] YMMV: in some configurations you may still have other sources of noise.
The Jackson Pollack example cited in the article is apt. Lots of artists use randomness-infused techniques, and select their final product from among many random versions based on what came out the best.
*However I am pretty confident that the result will eventually land on "Works created by big media companies, even done entirely using AI, are fully copyrightable."
Info: https://arstechnica.com/information-technology/2023/02/netfl...
There's certainly a blurry line here, it's never going to be 100% cut and dry. I can copyright a figure drawing I do, but I can't copyright my drawing of a straight line. Maybe text-to-image can't be copyrighted, but image-to-image results can. Going to take a long while to resolve this.
> First, that's not the right legal standard. The standard is whether there is a "modicum of creativity," not whether Kris could "predict what Midjourney [would] create ahead of time." In other words, the Office incorrectly focused on the output of the tool rather than the input from the human.
how does "elephant with blue skin" or any other prompt meet the criteria of 'modicum of creativity'?
in the end, its the tool that is doing the heavy lifting and being able to copyright its output sounds against the spirit of copyright (allow a human to get proper compensation for their creative work and incentivize creativity) imo.
is there something obvious i am missing?
If you're somehow at the helm of a plane, car, boat and you don't have the training, but it drives itself, you aren't a pilot, driver, captain, you're a passager.
AI Art is made by passagers, not an artist.
iPhone owners aren't photographers.
Pressing the button, on an already setup camera, is something that anyone can do as well, and they will get the exact same camera output as anyone else.
The decision about where to take a picture, what settings to use, and how to take it is the creativity that grants the image copyright. There are various arguments about which parts of prompt generation and refinement might register as copyright and there are various arguments about how tuning settings and tuning prompts is similarly creative to using a camera, but the point is still that unless the user input is sufficiently creative, the image wouldn't get copyright. The button doesn't matter; the button is not what gives you copyright.
In fact, numerous accidental photos have been denied copyright; most famously when PETA sued a photographer over a picture that was taken when a monkey stole the photographer's camera. The court's decision was that nobody owned the photo. It was an accident, it wasn't the result of a creative decision. There was not enough human creativity involved in the process of a monkey stealing a camera to warrant protection.
Of course, it's somewhat of an oversimplification of prompt engineering to phrase it as just saying "elephant with blue skin", but if that was the entire creative input, it's not clear at all to me that someone saying "I want an image of an elephant with a blue skin" is a sufficiently creative input that it should be copyrightable. What the AI does with that prompt is irrelevant, it's the human creativity that matters. Same with photography; the camera isn't really the important part. The machine isn't what is generating the copyright. The person making a conscious decision about where to stand, what settings to use, and when to press the button is viewed by the law as a creative act that requires creative skill and execution. The button press itself doesn't matter.
Again, to be clear, prompt generation tends to be oversimplified in these conversations and I'm not saying there's definitely nothing creative happening, but if we take that simplified version of prompt generation at face value, then just saying what image you want... does that really meet a creative standard?
Because saying that descriptions/requests on their own are sufficiently creative for copyright protection has implications far beyond AI art; it implies that commissioning a piece of artwork even from a human being should be enough of a creative act that I should get joint copyright over the final image. If the argument is that prompt generation is more than that, and that it takes more skill, then fine -- but if the argument is that even just a one line description of what image you want should be counted as creative... yeah, that's a pretty significant expansion of copyright that will affect a lot more than just AI art.
---
Edit: I think people also get a little confused about the difference between how people generally treat photography and what the law would decide if the copyright on every single image was challenged.
If I set up a camera mounted to a pole, focused on a static scene, and you walk over and press the button on the camera, legally you very likely don't have any copyright over that image. But practically, nobody is going to challenge you over it.
It's possible that some of the photographs where people say, "well that gets copyright, why doesn't mine?" might not actually get copyright if they were ever challenged. But people generally don't challenge copyright status in the first place. Recipes, APIs, monkeys, (and apparently now AI images) are the rare exceptions.
Indeed that is the case! And all the factors that you are talking about, could apply just as much to AI art. I am glad that you reject this idea, that "Lul, its writing a prompt, dude, therefore AI Art can't be copyrighted!"
> but if that was the entire creative input
> it's the human creativity that matters.
Sure, if we want to say that someone spending 10 seconds, writing a 3 word prompt isn't much human input, fine.
But, for anyone who has spent more than an hour, messing around with AI art, you will quickly learn that there is so much more to AI art than the degenerate case.
> The person making a conscious decision about where to stand, what settings to use, and when to press the button is viewed by the law as a creative act that requires creative skill and execution.
All of which can be applied to AI art. It can be so much more complicated, than the stupid argument that I see over an over again, which is "Lul, the computer did it. No human input".
To give a personal example, one thing that I did, was that I created a dreambooth, fine tuned model, for creating magic items, for the game Dungeons and Dragons.
https://huggingface.co/stale2000/sd-dnditem
It took me about 100 hours, over the course of 2 weeks, to get this art generator, exactly right. And the process involved, not just typing a few words like "Give me magic fire sword please".
Instead, it involved the process of collecting, modifying, captioning, and combining exactly the correct training set of images, into my training set, recreating and retraining my model over a dozen times, until I could get it to do even close to what I wanted.
Even one bad image or one bad setting, that I put in my training set, could screw up the whole model, or make the swords look weird, or mess up the painting style, or any numerous actual creative differences.
> If the argument is that prompt generation is more than that, and that it takes more skill, then fine
That is what it is, for anyone who has done anything in AI art. But people's brains for some reason, stop working when they are faced with the fact that there is so much more that you can do with this stuff, and the stupid strawman case of "lul, I just typed a couple words into a prompt! No human input!".
I'm just trying to point out, it doesn't have anything to do with the button, or AI, or technology. It's not a good analogy for people to say "you just point a camera at a thing, and that deserves copyright" -- if that was the case, the camera wouldn't get copyright either. Neither prompt-generation nor photography is getting granted copyright based on the decision to press a button, there has to be something more complex going on in both situations for copyright to get involved.
An accidental photograph, a photograph that's made by simply iterating over every possibility in a space, a photograph that's a reconstruction of an existing copyrighted work without changes, etc, etc... none of those are copyrightable, and that they're made with a camera doesn't make them copyrightable. IP law for photography is the same as for anything else; the machine isn't the important part.
In the scenario you describe, if someone mounts a camera someplace and two people press a button and get the exact same photo -- I would be surprised if that second photo has copyright protection, it probably doesn't meet a threshold for human creativity.
If "print 'hello world'" is not copyrightable does that imply that more complex programs are also not copyrightable?
> "elephant with blue skin" or any other prompt
I don't buy it. The space of prompts is pretty big just at the length of a few lines. You'll find guides and how-tos on how to write the prompts you want, seems quite the complex and creative endeavour to me!
Within a given domain, if you are very exacting with your prompting and start doing things like adding in weights, you can do things like eg. create portraits of your friends, or of characters you've imagined, or basically make things that you can either see or imagine.
There's a new problem though. You can easily make a picture of 1 thing/person/concept, but if you have more than one, they share the prompt.
Say you have an ever so slightly modded Stable Diffusion checkpoint, and say you want a picture of a
1boy with short_hair, 1girl with (long_hair:1.3)
* a few images with what you thought you wanted (a boy with short hair and a girl with long hair),
* in some cases you'll get a boy and a girl both with long hair (because the long_hair:1.3 overrides)
* and in some cases you'll get a single person, who happens to have long hair (because of the long_hair:1.3) , and who is more likely to be female (because of implication: long hair was more commonly associated with females in the original dataset).
Of course, if you're smart, you might think to exploit this ambiguity on purpose.
On the gripping hand, this is obviously insufficient control.
So now we're getting newer systems that allow you to chain/layer prompts and apply different prompts with different strengths to different domains of your illustration. This means that you can have eg. your male character on one side, the female character on the other side, and deal with things like room decoration, the view from the window, and specific props - all with separate prompts.
And this is just applications of txt2img. There's also things like img2img, inpainting, and... well... new tools are showing up every week it seems.
If this hasn't crossed back into the creativity sphere already, it probably will by next month, and else the month after. (By manner of speaking)
The amount of money you can make via copyright protection should be in direct correlation to the amount of effort that went into generating the creative work.
If you type a prompt into a machine and it spits out an image. 30 seconds of work, copyright will protect you until you have made a reasonable profit for your time.
If a movie company puts 10000 person months into a big blockbuster, copyright will protect them until the move has made a reasonable profit.
If you choose never to monetize an artwork, it remains in copyright for the maximum time.
I have no idea how you would value a photographer being in the right place at the right time.
Good article, but the landscape is a lot bigger than what this article would suggest. And, even more obviously, constantly growing.