I’ve generally been against giving AI works copyright, but this article presented what I felt were compelling arguments for why I might be wrong. What do you think?
I’ve generally been against giving AI works copyright, but this article presented what I felt were compelling arguments for why I might be wrong. What do you think?
Copyright itself is weird.
That said, a simpler way to handle this would be that the image generator model is a tool. And it doesn’t really matter if your input is through a paintbrush or a prompt in an image generator, you’re responsible for that piece of content thus you have copyright over it.
You can further couple it with the argument @luciole@beehaw.org used, that AI art is derivative work, and claim that the authors of the works used to train the model shall have partial copyright over it too.
To me this is a potential can of worms. Humans can study and mimic art from other humans. It’s a fundamental part of the learning process.
My understanding of modern AI image generation is that it’s much more advanced than something like music sampling, it’s not just an advanced cut and paste machine mashing art works together. How would you ever determine how much of a particular artists training data was used in the output?
If I create my own unique image in Jackson Pollock’s style I own the entirety of that copyright, with Pollock owning nothing, no matter that everyone would recognize the resemblance in style. Why is AI different?
It feels like expanding the definition of derivative works is more likely to result in companies going after real artists who mimic or are otherwise inspired by Disney/Pixar/etc and attempting to claim partial copyright rather than protecting real artists from AI ripoffs.
[Warning: IANAL] The problem is that you guys assume that those generator models are doing something remotely similar to humans studying and mimicking art from other humans. They don’t; at the end of the day the comparison with music sampling is fairly apt, even if more or less complex it’s still the same in spirit.
Furthermore from a legal standpoint a human being is considered an agent. Software is at the best seen as a tool (or even less), not as an agent.
Quantifying “how much” of a particular artist’s training data was used in the output is hard even for music sampling. Or for painting, plenty works fall in a grey area between original and derivative.
Following this reasoning (it’ll get misused by the American media mafia), it’s simply better off to get rid of copyright laws altogether, and then create another legal protection to artists both against the American mafia and people using image generators to create rip-offs.
Certainly no disagreement from me on this point
Copyright is absolutely a can of worms. It always has been.
But it does matter whether your input is a brush or a prompt.
If you physically paint something with a paintbrush, you have a copyright over your work.
If someone asks you to physically paint something by describing what they want, you still have copyright over the work. No matter how picky they are, no matter how many times they review your progress and tell you to start over. Their prompts do not allow them to claim copyright, because prompts in general are not sufficient to claim copyright.
You’re ascribing full human intelligence and sentience to the AI tool by your example which I think is inaccurate. If I build a robot arm to move the paintbrush for me, I would have copyright. If make a program to move the robot arm based on various inputs I would have copyright. Current (effective) AIs prompts are closer to a rudimentary scripting rather than a casual conversation.
It’s not a matter of intelligence or sentience. The key question is whether the output of a prompt is fully predictable by the person who gave the prompt.
The behavior of a paintbrush, mouse, camera, or robot arm is predictable. The output of a prompt is not (at least, not predictable by the person who gave the prompt).
Predictable? How are people ‘predicting’ those abstract paintings made by popping balloons or spinning brushes around or randomly flinging paint around. Where does predictable come in? Humans have been incorporating random elements into art for ages.
After you’ve spun enough brushes or popped enough balloons, the results will be fairly predictable. And some elements, for example the color of paint in the brushes/balloons, would be under full control.
Even if the final result is not completely predictable, an artist only needs to establish that a significant part of it is a form of creative expression.
@FlowVoid
So what happens if you use a voice assistant to direct a vacuum cleaner though snow or trace a pattern with paint on a floor?
@lvxferre
A prompt is more than a command. It is a command with an immediate output that is not fully predictable by the prompt-giver.
So for example the copyright office might ask, “This image includes a person whose left eye is a circle with radius 2.14 cm. Why is it 2.14 cm?”
Traditional artist: because I chose to move the paintbrush (or mouse) 2.14 cm. The paintbrush (mouse) can only go where I move it.
Photographer: because I chose to stand 3 meters from the subject and use an 85mm lens on my camera. The magnification (size) of the eye depends only on those factors.
AI-assisted artist: because I asked for larger eyes. I did not specify precisely 2.14 cm, but I approved of it.
In your example, if you can fully predict the output of the vacuum by your voice command, then it is no different than using a paintbrush or mouse.
Counterpoint: imagine that I used physical dice to decide the size of the eyes, I just rolled 4d6 and got 21 so the eye is going to be 21mm large. From a legal standpoint, I believe that I’d be still considered a trad artist; however effectively I’m doing a simpler version of what the image generator-assisted artist does.
The output is still fully predictable by the artist.
The dice didn’t make the eyes, after all. They just showed 21, now it’s your job to actually make 21mm eyes. In doing so, you could mess up and/or intentionally make 22mm eyes. If someone asks, “Why are these eyes 21mm?”, you can answer “I decided to do what the dice asked”.
The dice are more like a client who asks you to draw a portrait with 21mm eyes. In other words, they are giving you a prompt. Nobody but you knows if they will get what they asked for.
And arguably, neither the image generator did. Who did were the artists of the works being fed into the model. In this sense, the analogy is like the artist picking an eye from some random picture and, based on the output of the dice, resizing it to 21mm.
The same reasoning still applies to Stable Diffusion etc., given that you can heavily tweak the output through your prompt. And you can also prompt the program to generate multiple images, and consciously pick one of them.
In which case, neither the image generator nor its operator are eligible for copyright.
The point is that the AI generator (or, if you prefer, its training data) exercised direct control over the image, not you. For that matter, giving extensive prompts or other iterated artistic direction to a human artist would not make you eligible for copyright, either. Even if the artist was heavily influenced by your suggestions. There is a fundamental difference between an art creator and an art critic.
Finally, choosing one among many completed works is not a creative process, even if it requires artistic judgment. That’s why choosing your favorite song does not in any way make you a song creator. Even if you know that all the songs you don’t choose will be destroyed.
Ultimately the only agent involved is the person commanding the vacuum cleaner through the voice assistant. If someone got copyright, that should be that person; and if we’re going to be consistent with the current (and rather shitty) typical copyright laws used by plenty govs, it should.
By “thus you have copyright over it”, I’m saying that it should apply equally to both (paintbrush vs. image generator), not that it currently does.
So if someone asks you to paint something and gives you detailed instructions about what they want to see in your painting, do you think they should have copyright over your work?
Under the spirit of copyright law (that, again, I criticise), and depending on how detailed those instructions are, that wouldn’t be “my work”. It would be “our work”, because the person in question is actively creating the work alongside me.
Transposing this reasoning to images generated by Stable Diffusion etc., you’d get co-authorship between the person inputting the prompt and the people who made the works used to feed the model with. You could even theoretically argue a third author - the person/people coding the model. (It’s a legal nightmare.)
No, under copyright law it would be your work and your work alone.
Someone who is providing suggestions or prompts to you is not eligible to share the copyright, no matter how detailed they are. They must actually create part of the work themselves.
So for instance if you are in a recording studio then you will have the full copyright over music that you record. No matter how much advice or suggestions you get from other people in the studio with you. Your instruments/voice/lyrics, your copyright.
Otherwise copyright law would be a constant legal quagmire with those who gave you suggestions/prompts/feedback! Remember, an idea cannot be copyrighted, and prompts are ideas.
In the case of Stable Diffusion, the copyright would go to Stable Diffusion alone if it were a human. But Stable Diffusion is not a human, so there is no copyright at all.
That doesn’t work with AI for a variety of technical and practical reasons.
Two people could, completely coincidentally, generate something that is so similar that it looks the same at a glance… even with dramatically different prompts on dramatically different models.
No, the output of an AI is fundamentally “coincidental” and should not be subject to copyright. Human intent and authorship MUST be a significant factor. An artist can still use AI in their workflow, but their direct involvement and manipulation must be meaningfully “transformative” for copyright to apply in a fair and equitable way.
As a rule, the law doesn’t care about technicalities or practicalities.
Sure. If two humans take a photo of the same sunset standing next to each other, they will be virtually identical and they will both own full copyright protection for the photo they took.
That protection does make the other person’s photo an illegal copy - because it wasn’t a copy.
Of course. By the way that applies to human created works too. Copyright doesn’t apply to everything created by a human, only certain things are protected. If someone asks you what 2 plus 2 is, and you reply “4”… you don’t own the copyright on that answer. It wasn’t creative enough to be protected under copyright. If you reply with a funny joke, then that’s protected.
If the image generator in question generated a piece of content that looks similar enough to some piece of content protected by copyright, odds are that it has been trained with it, or that both were produced with the same set of pieces of content.
That shows that the issue might be elsewhere - it isn’t the output of the image generators, but the works being fed into the generator, when “training” them. (IMO the word “training” is rather misleading here.)
Refer to the link in the OP, regarding photography. All this discussion about “intent” (whatever this means) and authorship has been already addressed by legal systems, a long time ago.
What does “meaningfully” mean in this context? It’s common for people using Stable Diffusion and similar models to create a bunch of images and trash most of them away; or to pick the output and “fix” it by hand. I’d argue that both are already meaningful enough.
If anything, this only highlights how the copyright laws were already broken, even before image gen…