Credit: Boris Eldagsen/DALL-E2 (He [the human] refused to accept a Sony World Photography Award.)
By Eddie Garmat
When you create a book using AI, who owns it? What about a picture or a video? These may seem futuristic issues, but the future is already here. AI companies have been working for years to create text-to-image and text-to-video AIs. Today we have products like Sora from OpenAI and Veo2 from DeepMind, a Google subsidiary. These programs challenge humans with top-of-the-line video outputs. You may be inclined to say the person who prompted the AI gets the copyright, but does someone who commissions an artist get the copyright? Then you might revert to the company that made the AI. Does a company who never signed a contract with an artist get the copyright to that art? In this series, we’ll dive deep into the state of modern copyright law regarding AI and discover whether anyone owns AI art.
What do we mean by ownership? Copyright.
Defined straight from the Copyright Office of the United States, “Copyright is a type of intellectual property that protects original works of authorship as soon as an author fixes the work in a tangible form of expression.” Basically, if an author or artist creates a work that is original and creative, they have the exclusive rights to reproduce, distribute, or perform the work. Copyright does not protect an idea, but once something tangible materializes, it can be copyrighted. Generally, when a work is created, copyright is automatically guaranteed, and registration is optional. The benefit of doing so would be to file a lawsuit if and when your tangible work is co-opted.
How does an AI image generator actually work?
According to How-to Geek and MIT’s CSAIL when an AI is given a prompt for a toaster, contrary to popular belief, it doesn’t compare all of the images in its training data with the tag “toaster” then create an approximation. What actually happens is, as an AI model is fed images with tags, “noise” (what we perceive as static) is applied to the image over and over until it becomes unrecognizable. The AI is then taught how to undo this process. The undoing of this static is called diffusion. So if you ask an AI for an image of a toaster, it starts with an image of noise (static) and moves a few pixels so it thinks it looks more like a toaster. This process is done over and over until the AI thinks it’s created something that would be described as a toaster. The important part for our question is that the AI doesn’t learn how to copy, it learns how to create. It’s actually very similar to how humans create art.
Given this new definition of creativity, what say the law? Browse below to learn about US courts’ interpretations of the Copyright Act and how these have created precedent for the legality of AI-generated images.
Let’s say you’re on a wildlife trip in Indonesia and you see a beautiful monkey. You set up a camera near the monkey and it starts taking selfies. Who owns those images? It may seem outlandish, but this is the background for the case Naruto v. Slater. As US News & World Report, Arstechnica, and Lexology all report, in 2008 this very scenario happened to David Slater and in 2011 he licensed the image. In 2015, People for the Ethical Treatment of Animals (PETA) filed a lawsuit against Slater to request the monkey–whom they had named Naruto–gain copyright of the image, and that PETA become the administer of funds. The United States Court of Appeals for the Ninth Circuit ruled that Naruto had copyright of the image as he was the one who took it. The catch, however, was that only humans could have copyright, meaning the image was actually in the public domain. This case set the precedent that only humans can have copyright of something.
If you go through incredible efforts to recreate a public domain artwork – we’re talking dozens of hours of work to recreate this – and you recreate it perfectly, do you own that recreation? As mentioned in this Wikisource document, in Bridgeman Art Library v. Corel Corp, the courts say no. Bridgeman Art Library owned many photographs of public domain artworks and argued they were protected by copyright since they required creative effort to create. Corel Corp, a professional graphic design company, used some of Birdgeman’s photos without permission. The art library brought the matter to court claiming Corel Corp had illegally used their copyrighted works. The United States District Court for the Southern District of New York ruled that Bridgeman Art Library’s photos of old paintings did not fall under copyright because they were judged to have no creativity. Therefore Corel Corp was free to use the photographs however they wanted. This set the precedent that recreations have to be original to be copyrighted.
In this scenario, you have made a telephone list of a large rural area. You put the list in alphabetical order. Do you own the copyright to that list? You might say you do because it’s in alphabetical order and that constitutes originality. Well, the Supreme Court says no (6). You can’t copyright facts.
Kris Kashtanova wrote a comic book, arranged all the panels, wrote all the text, and made all of the creative direction decisions, but all the images were AI generated. This comic book was called Zarya of the Dawn. According to the Wall Street Journal, Kashtanova applied for copyright and it was initially granted; their next move was to post about it on social media, which proved to be a mistake. AI Business reported that in a Facebook post Kashtinova said the Copyright Office “overlooked” the work’s artist being AI despite being mentioned on the cover. The office initiated a proceeding to revoke the protection of the copyright and instead granted Kashtanova the copyright to all the text and layout of the panels but not the images. Kashtanova now holds the copyright to an unpublished literary work since what she has rights to is effectively scattered text. An unpublished literary work is usually a partial product that can’t be published due to its use of copyrighted material within the work itself.Kashtanova did fight this ruling. They argued that they didn’t simply use the first take in their comic book. They went through hundreds or even thousands of iterations until they got the image they liked. They effectively creatively directed the art. This notion was denied, and Kashtanova would never get the copyright for the images. This reinforces the precedent that the user cannot have copyright of AI generated art.
If you take 10,000 photos then build an AI completely from scratch and train that AI exclusively on your photos, who owns the photos that the AI generates? The answer according to the US Patent Office is no one, The Verge reported in 2020. Chiefly, the AI created the image with inspiration from 10,000 existing images, and since the AI is not human, it can’t own the copyright of an image. If I examine every single Van Gogh painting, then decide to paint an image in his style, does Van Gogh own the copyright to that image? Of course not. That’s half the argument against Stephen Thaler and DABUS. The other half is that, as we established in Slater, only humans can own copyright. Since an AI is not human, it can’t own the copyright to something.
The New York Times (The Times or NYT) has sued OpenAI and Microsoft for “unlawful copying and use of The Times’s uniquely valuable works” The Harvard Law Review reports. OpenAI has used countless NYT articles in their training data for ChatGPT. The Times argues that since ChatGPT has this unauthorized data, it can summarize existing articles and create new ones in The Times’ style. The Times continues to argue ChatGPT and OpenAI not only stole copyrighted material, but serve as a market replacement for The Times. OpenAI has 2 major defenses to this.
Recall that Feist Publications, Inc. v. Rural Telephone Service Co. established that copyright does not apply to facts. The Times is about reporting facts, just facts. If their argument is to stand, they must provide evidence that they have been sufficiently creative to gain copyright – for every single article ever published in the newspaper. What OpenAI’s lawyers have done is ask for every bit of research, every single note, and every shred of creativity for every article The Times has ever posted. There is no way to convey how monumental of a task this is for a publication as large and historic and the New York Times.
The other defense OpenAI has is fair use. National Public Radio reports OpenAI is arguing that since they’re generating something new, not copying an existing article, they are protected under fair use. Now fair use protects a work that is transformative. Judging whether something is transformative is incredibly hard and is done on a case-by-case basis. For something to be considered transformative, it must modify the original work and not claim to be the original, but instead make a reference to, be in the style of, or take inspiration from the original work. If a piece satisfies these conditions it generally is transformative. OpenAI is claiming that since their AI doesn’t copy The Times’ articles, it either summarizes them or creates new ones in The Times’ style, that their work is transformative and therefore protected under fair use.
Another take on the fair use defense is that The Times’ articles only serve as an inspiration for ChatGPT. If we look back at the Van Gogh argument from Thaler and DABUS, we can say products of inspiration do not warrant copyright. Here, OpenAI can argue that NYT articles are simply an inspiration for ChatGPT.
This case is still ongoing. The result will be a landmark decision regarding the unauthorized use of copyrighted material in training data for AI.
The other major modern AI copyright case is Getty Images v. Stability AI. As reported in the Justia Docket, Getty images is one of the largest galleries of stock photos in the world. Over 3 decades they have gathered hundreds of millions of photos from photographers they employ and freelance photographers. Getty accuses Stability of illegally using Getty’s images as training data for Stable Diffusion, Stability’s AI image generation algorithm. Getty says that Stability has infringed upon Getty’s copyright and gained an unfair advantage against those who have licensed their images. Getty also claims that AIs using Stable Diffusion often produce images with recreations of the Getty Images watermark, which can create confusion among consumers and that tarnishes the reputation of Getty Images. Stability has a few available defenses.
The quick defense is fair use. As mentioned in an IPRMENTLAW article, Stability argues fair use as the purpose of their use of Getty’s images are transformative. Since Stable Diffusion’s goal is to generate new images, not recreate existing ones, they argue it falls under fair use. Stability AI can also argue that there are no substantial similarities to Getty’s images, despite their AI being able to recreate Getty’s watermarks.
The other defense Stability AI is employing is that if a user intentionally recreates copyrighted material, that’s on the user. This is similar to how YouTube operates. If something illegal or against the YouTube Terms of Service are uploaded to the platform, their defense goes along the lines of ‘We’ll do our best to control it but ultimately blame is on the user.’ Stability is arguing that they can try to control recreation of copyrighted material, but ultimately it is the user who generates copyrighted material.
Conclusion
So to go back to the original question, when you create art using generative AI, who owns it? Well Zarya of the Dawn proves to us that the user doesn’t own it. The companies that own the AIs don’t own it as most of them sign away their rights. Thaler shows those who don’t have any copyright to begin with. The defenses in New York Times v. OpenAI and Getty Images v. Stability AI show that the artists might not own anything because the AI works meet the benchmark for transformativity. Stephen Thaler and DABUS show us that the AI algorithm owns a work of art, but Slater and Thaler both show us that only humans can possess copyright. So every case we’ve looked at basically shows us this: no one owns anything! This effectively puts any AI generated art in the public domain and makes it completely free to use and consequence-less to generate.
Postscript
After finishing the initial draft for this blog, the AI model DeepSeek R1 came out and completely shattered the US’ monopoly on generative AI. A Cornell research article says the model was significantly cheaper to create at a reported cost of about $6 million vs. about $100 million for ChatGPT. OpenAI has accused the founders of DeepSeek of ‘distilling’ ChatGPT’s data and using it to train DeepSeek. The Atlantic tells us this means DeepSeek developers would have prompted ChatGPT billions of times and trained DeepSeek on the outputs. These claims are not entirely unfounded. DeepSeek has occasionally identified itself as ChatGPT, and OpenAI reported a mass export of data—although they have not reported how they know it is linked to DeepSeek. OpenAI claims this is theft although it’s not quite clear how it’s different from what they do to obtain data. As we covered, OpenAI uses seemingly endless quantities of copyrighted material to train ChatGPT in the same manner that DeepSeek would have trained their model. I believe the legality of distilling hinges on the outcome of New York Times v. OpenAI. If the New York Times wins, using copyrighted material to train AI, which serves as a market replacement, would be illegal without a license. This would mean DeepSeek could no longer use OpenAI’s data in any model commercially available in the US. However, as we’ve discussed, no one currently owns the copyright to the output of an AI. It is possible that even if the NYT wins it may be legal for DeepSeek to use the output of ChatGPT. The outcome I believe is more likely is that OpenAI beats NYT, implying that even if OpenAI had the copyright to ChatGPT’s output, it would be legal for DeepSeek to use that data as training data.
Leave a Reply