District Court Judge for the District Of Columbia Beryl A. Howell endorsed the United States Copyright Office’s decision to refuse copyright protection to the owner of an artificial intelligence system that generated visual art. Stephen Thaler owns a program called the “Creativity Machine,” but he was denied a copyright by the office for a piece of visual art his system created. It “lacked human authorship,” Howell decided, meaning that it lacked the “bedrock requirement of copyright.”
“[T]his case presents only the question of whether a work generated autonomously by a computer system is eligible for copyright,” Howell wrote. “In the absence of any human involvement in the creation of the work, the clear and straightforward answer is the one given by the Register: No.”
Wait: the thing didn’t just wake up one morning and decide to create something. A human being was involved at some level. If a human enters the perimeters that an AI program follows and the result is unique art, isn’t that sufficient “human involvement”?
Your Ethics Alarms Ethics Quiz of the Day is…
Is that ruling competent and fair?
My guess is that this is another (of many) examples of courts misunderstanding new technology, with rulings that will eventually look silly in the rear-view mirror.
Ethics Alarms 
I don’t know from a legal perspective if it is competent, but I’m pretty sure it is the only option unless you want to take a deep dive down a legal rabbit hole. AI in its current form works by scraping the internet for images and using a rule set derived from human training to make composite images. It is already probably scraping copyrighted material to make the images it is making. Can the copyright holders of the originals sue the copyright holders of the composites?
Then move in a step, AI images are now all over the web and AI is now scraping its own output as input. So if you have the original copyright, and then a secondary copyright, then can people copyright twice scraped copyrights? It’s only going to continue in a chain of copyrights until all possible combinations of pixels are owned by someone. Then mega corporations will buy up all the copyrights and make it so art is no longer legal without renting the right to imitate their inventory. Preschoolers will be getting slapped with takedown orders for sticking drawings on the kitchen fridge.
Ok, that last bit is a little over the top. I really don’t know how copyrighting AI generated material would be good in practice, though.
Excellent observations.
jvb
1. A deep dive is what new technology always requires. We’re still figuring out the internet.
2. “Can the copyright holders of the originals sue the copyright holders of the composites?” Collages using nothing but pre-existing images are considered original: so are “mash-ups” of literature.
3. Sure. Music, literature, art all borrow (steal) from multiple sources. it’s just along a spectrum.
4. Current copyright law prevents the last scenario.
I don’t know enough about law to opine on the legal side of things so I won’t try. I only know how things work from observation, and my observations are of DMCA takedowns being abused to remove things that ought to fall under fair use from the internet. People playing a video game to make a how to play guide get slapped by DMCA because the in-game music was playing. People review movies and get slapped because a 10 second snippet of the movie was played. There is what is legal, and then there is what people can afford to fight large corporations. There is a lot of things getting smacked down by lack of ability to fight with deep pockets. I doubt art will be treated any differently than music once the lawyers get involved.
Those are very good points regarding DMCA. I’ve been told a number of stories where content creators on YouTube were SLAPPED (pun intended) with an illegitimate take down order. In some cases the rights holder seems to just do it routinely without examining the content involved.
What they count on is that these people’s income is threatened, an especially potent threat when you are a full time content creator. It does not appear to be easy to contest these claims, and the big platforms are also somewhat to a lot biased in favor of the rights holders, as these tend to be the folks with deep pockets.
And, of course, another motive for those folks is to deter negative reviews and publicity — well, that may be the hope but I suspect they’re unaware of the Streisand effect.
I can judge for myself whether someone is producing legitimate content, I think, and it irks me when that type of person gets unfairly hit by one of these actions.
And as an added note, we see shenanigans with the DMCA and it causes us to mistrust those in power a bit more. Just another incremental loss of trust and respect, and extra added litigiousness.
Collage might be legal, but what happens when the AI spits out the exact original? That can happen with generative AI.
And that can’t be copyrighted.
Ok, so 100% copy is not copyrightable as a unique copyright. Now, what if it’s a 99.99% copy? Or total copy but tinted in new shades. Shadows added, modified or removed? At what point does it become ethical to copyright a piece of art as a “unique” piece of art? How does someone define that from a legal standpoint? Isn’t it going to fall into that nebulous territory of knowing it when you see it? You cannot even consider intent in this case because the machine doesn’t have any intent one way or the other. The person generating the images has intent to make a new piece of art but lacks the ability to determine whether the algorithm actually did or not without running some kind of comparison algorithm against all known images.
Many of the images are generated with commands like “x in the style of y”, so intent to mimic is at least a factor. The degree to which the machine does so and in what ways isn’t in the control of the creator.
I still think it’s ethically murky.
It’s funny this was posted now, as I have been playing around with AI image generation in the last few days. I’m certainly no expert in how everything works under the hood or how the system I am using compares to the one in this case, but I’ll give it a shot.
The system I have been using is Stable Diffusion. Essentially it is a machine learning model that is trained by showing it a few thousand images, which then allows it to generate its own images. These images are created by entering a text prompt. Below the system level are checkpoints; these are trained to produce a specific type of image, such as trees, buildings, human faces, etc. At the core of the image generation process is the “seed”. A seed is simply a string of numbers used by the AI to generate an image. Level 2 is the sampler, which determines how these numbers translate into an image. Level 3 is the step count, which determines how many iterations the seed to image process takes before the image is finished. Level 4 is the CFG number; the higher this number is, the more strictly the image generation is tied to the text prompt. There are also various adapters, which change how a certain part of the image generation takes place. An adapter might make eyes with a realistic green color, or operate more like a photoshop filter and make the image look like an oil painting, for example.
If all of that is confusing and makes no sense, don’t worry. The core idea is, all of these are objective steps that will always create the same image. While there are a lot of complex variables, it boils down to translating a number into a picture. A consistent translation process will only ever make a single image from a single seed. If Jack used the same software and all the same settings and the same seed as me, he would create exactly the same image.
Is the process of writing the text prompt and choosing all of the settings an act of human authorship? I would argue not, as the human is not the one who conceives the image. The conception occurs when the AI translates the human input into an image. To argue otherwise would be to claim that something like “A tree on a grassy hill, Realistic Checkpoint afd7e94, Euler-a, 20 steps, CFG 7, #521887” is an identifiable image. It’s not, rather it is a set of instructions for a machine to create an image of a tree on a grassy hill in a certain way.
Whether a set of instructions to create a piece of art can be copyrighted is certainly an interesting question. However, it was not the question before the court. The plaintiff was claiming that the computer created the image, he owned the computer, therefore any creation of the computer transferred to him. Thus the judge made the right decision in this framework, which is that a computer is not a human and therefore can’t create material eligible for copyright.
If anyone knows more about AI image generation or how this Creativity Machine works, feel free to correct me. I am by no means positive I got everything here right.
Reading the decision further, the judge admits that the law on copyright will probably need to be seriously overhauled to handle cases where a human and non-human are both involved in the creative process. And while the plaintiff did in fact bring up his role in prompting the AI at trial, he never did so in his filings for the patent office. Rather he only ever claimed that the AI autonomously created the images. As the judge is required to rule on the contents of the patent filings, he had to dismiss the case. The plaintiff shot himself in the foot on this one.
Also, there was another interesting case cited that was similar. Apparently, someone tried to copyright the layout of a garden. While the court admitted that the way the garden was cultivated was original, the primary force responsible for the garden layout was nature rather than the gardener. I think there is a clear parallel where the image prompter is the gardener and the primary force that creates the image is the AI.
I can understand the argument for allowing people to copyright things they make using artificially intelligent tools, so I will argue the opposition, since it is less obvious but still deserves consideration. I also believe it is not without merit.
An artificial intelligence is qualitatively different from other tools. It is effectively a sliver of a mind. An AI collects data, infers patterns from it, and then extrapolates from those patterns based on requests in order to generate new data. Roughly speaking, those processes of pattern inference and extrapolation are what separate rote data storage from actual learning, and mere computation from intelligence. An intelligence builds a model of reality and uses that model to navigate situations towards particular goals.
Thankfully, the AIs we currently have have no goals of their own. They are also not as versatile as human brains; they only build models of very limited contexts. However, an art-generating AI is very intense in its narrow focus. We have to consider the implications of people’s ability to buy very powerful, narrowly focused mind-slivers and instruct them to create art. This poses some problems for the copyright system.
I’m ambivalent about how intellectual property is implemented, but I know the purpose is to incentivize intellectual effort by making people pay to use others’ intellectual creations. There are two situations protected by copyright that have worked just fine with this principle, but AI may disrupt that accordance.
First, a person can learn skills from imitating others, and copyright new works inspired by others. Their compensation is limited by how much effort they put in to develop their skills and create their products.
Second, a person’s “intellectual creation” can be simply recording random events from inanimate sources. This takes very little effort, but their compensation is limited by the lack of skill that went into the product.
These two situations demonstrate balance. You can borrow creative style and ideas, but benefitting from them still takes effort to learn and apply the skills. Contrapositively, low effort yields low-quality work. AI breaks this balance, supplying high skill for low effort. Quality and quantity.
Here’s the first problem: an AI isn’t an ordinary tool, like a camera. Photographers take pictures of things they didn’t create, using cameras they didn’t design, and they can copyright their photos. That’s fine; photography takes skills. Furthermore, the camera inventor was paid for the skills they used to create the tool. Moreover still, the photographer can only take pictures of things that already exist. They can take pictures of someone else’s art, but the power to create more such art still lies with the original artist.
Creating art with AI also takes skill: you must know what data to feed the AI, how to pose your requests to it, and what your audience wants to see in the finished product (and in most cases, how to touch up the results).
But the AI isn’t just built on the skills of the people who use it, or those who design and sell it. It copies the skills of people whose art it’s fed. A person can learn to imitate other artists’ styles and make art that rivals theirs. They have that right, and that is good. An AI can do that too, but those skills don’t go into the AI owner’s brain. They stay within the AI itself. The AI is a repository of other people’s skills, and it can learn them with much less time and effort. How much credit can take for the creations of a machine that learns skills on your behalf, faster than you can?
The second problem is that the AI doesn’t just learn superhumanly fast. It’s superhumanly prolific. Artists spend a great deal of time getting good at their craft and still struggle to make a living. Now they’re not just competing with other human artists, but AIs that can create works thousands of times faster.
They’re facing down a device that collects decades of skill based on the creations of others, and then floods the world with similar creations, all within minutes. The AI looks at art, reverse engineers the art-making patterns the artist developed over years, and applies them endlessly on command. The quality is not limited by the user’s skill, nor the quantity limited by the user’s effort, for the user supplies neither. So what happens to the human artists?
As far as I can tell, that’s how and why people would draw the distinction between AI art and human-created art for the purposes of copyright. That said, if a person applies artistic skills to modify an AI-generated piece of art in a substantial way, I would find it difficult to say they couldn’t copyright it. Requiring every copyrighted piece to be personally modified by a human might also mitigate the issue of unlimited productivity, although I couldn’t say how much.
In any case, if we want to live in a world where people can make a decent living under good conditions as artists, we’ll have to get creative and explore options for making that happen. An AI won’t do that for us, because it has no examples to learn from.
“[T]his case presents only the question of whether a work generated autonomously by a computer system is eligible for copyright,” Howell wrote. “In the absence of any human involvement in the creation of the work, the clear and straightforward answer is the one given by the Register: No.”
If that is the question, then her judgment is wrong.
First of all, an AI cannot exist (yet) autonomously by any reasonable definition. It requires a man-made environment, it is designed, programmed, and its execution parameters put in place by men and/or women. It is effectively a creation of man with an ability to learn limited by its programming and operating environment which is delineated, controlled, and operated by humans.
In other words, it is a limited learning machine, and is not truly autonomous in any sense of the word. Human interaction is directly or indirectly required at every level of execution.
What Howell (who is doubtless the most unethical and technologically illiterate judge in the DC circuit) is attempting to do is characterize this limited AI as a true thinking machine in the mold of SkyNet or the Butlerian Jihad of the Dune series. That isn’t the case, and showcases her fundamental technological cognitive infirmities.
I agree with you, Glenn. I could see the same kind of opinion being issued to deny that photographs could be copyrighted: after all, its the machine that makes them.
I didn’t think of that, but it is a better analogy than some might think.
Having an idea about what you want something to be and bring able to do it is what separates the artists from the rest of humanity.
The fact that artists see what most of us can’t, and then present it in a form or fashion that shows it in (usually) an aesthetically pleasing manner is another trait of artists the rest of us do not have.
An AI generated image still can’t do that, it pulls from what actual artists have done.
The guy who asked AI to generate the image can’t do that, so what’s he copyrighting?
I’m sure that’s ignorant as all hell related to legal reasoning, but it’s the reason I agree with the ruling. I think it was correct based on that.