Skip to Main Content.
  • AI chatbot usage and concepts

    Who Owns AI Created Content? The Surprising Answer and What To Do About It

This was originally published in Reuters Legal News (December 14, 2023).

Attempts to protect AI generated material, and to protect human generated works from being used in AI, have exposed serious issues in applying current law to generative AI. These issues may prevent the users of generative AI tools from protecting their creations, and could even render such users liable for infringement based on materials which were used to train the AI. This article discusses those risks, as well as steps which can be taken to mitigate them in taking advantage of this emerging technology.

If you made it with AI, you probably own less than you think you do…

On Sept. 20, 2022, artist Kristina Kashtanova posted to Instagram that she had obtained a copyright registration for a graphic novel she created using the Midjourney generative AI. The following month, the copyright office sent a letter to Ms. Kashtanova informing her that, now that it knew about the role played by Midjourney, it was initiating the process of canceling her registration.

Ms. Kashtanova’s attorney responded, pointing out that the copyright office had registered works created with mechanical assistance since at least the advent of photography, and arguing that Ms. Kashtanova’s work embodied her original conception and was therefore copyrightable regardless of the fact that she used Midjourney when creating its images. Ultimately, though, the copyright office was unpersuaded. It concluded that Ms. Kashtanova was the author of the graphic novel’s text, and that the selection and arrangement of its images and text were also protectable as a creative compilation.

However, the graphic novel’s individual images were treated as lacking sufficient human creativity and so a new registration certificate was issued which explicitly excluded “artwork generated by artificial intelligence.”

The basic issue which led the copyright office to exclude the images from the new registration certificate was that Midjourney’s specific output cannot be predicted by users. While artists have long created works using more or less unpredictable processes, the copyright office felt that Midjourney’s images differed from copyrightable artworks because it began the image creation process with a field of random noise.

The fact that Midjourney’s images are only created as a result of prompts from a user like Ms. Kashtanova was not sufficient to qualify Ms. Kashtanova as an author because the prompts were only influence, and there was a significant distance between what a user may direct Midjourney to create and what it actually produces. Essentially, the copyright office found that Ms. Kashtanova’s use of Midjourney was analogous to hiring a visual artist to produce images based on brief text prompts. Since Ms. Kashtanova would not be the author of images created by such a hypothetical visual artist, she also was not the author of the images created by Midjourney.

… and may even be infringing someone else’s rights.

While the analogy between Midjourney images and images created by a hired artist may be a bit muddy, there is a group that has embraced it — artists whose work was used to train Midjourney (and other similar tools) and who are allegedly no longer being hired because of the availability of Midjourney images.

In an ongoing class action lawsuit, these artists argue that, not only should Midjourney itself be treated as an unauthorized (and therefore infringing) derivative work, but also that when AI tools rely on an artist’s work in creating an image, the user of the AI tool is liable for infringement. In the class action complaint, this argument is limited to images created using prompts which include a particular artist’s name (referred to as “Fakes”), and therefore are more likely to pass as original works by the artist.

However, it is not clear that there is any principled reason why the argument, if accepted, would be limited to only that narrow class of images. Indeed, since the underlying theory of infringement alleged in the litigation is that Midjourney (and similar tools) is nothing more than a 21st century collage tool for combining copies of the artists’ works, and since similar allegations have been made regarding the output of large language models from OpenAI, it appears that any use of any generative AI tool could arguably be treated as infringing the rights of the authors whose works were used for training data.

There are limits to the risks

Fortunately, while the copyright status of creative works created using generative AI is highly uncertain, there are other types of protection which may be available. For example, in contrast to the conclusion reached by the copyright office, the U.S. patent office took the position that a human would not be disqualified from inventorship because he or she used an AI as a tool in making an invention. “Public Views on Artificial Intelligence and Intellectual Property Policy,” USPTO, October 2020.

Similarly, if the output of a generative AI could be held as a trade secret, then whether the user of the generative AI owns it in the copyright sense may not be relevant, since the ability to protect a trade secret depends not on ownership but on possession.

Additionally, while copyright protection may not currently be available for works created using a generative AI, in most cases the risk of those works giving rise to infringement liability is likely to be relatively low. This is because, to show infringement, a copyright holder must show that the alleged infringer copied original elements of the copyright holder’s work.

Given how thoroughly the training data is diffused and mixed in most generative AI systems, any original elements included in that data would be unlikely to be preserved in the generative AI’s output — at least outside of special cases such as the “Fakes” from the Midjourney litigation. Further, even if some original element was incorporated into a generative AI’s output, that could only serve as the basis for an infringement complaint if: (1) it was actually discovered; and (2) the author of that element registered the copyright in their work.

Accordingly, unless a generative AI is used in such a manner that its output would be recognizably linked to some person or entity who is likely to actively police the use of their works and whose works are likely to be registered, the risk of the generative AI’s users being sued for infringement seems low.

but smart governance is needed to minimize them.

In practice, the legal issues surrounding generative AI mean that its outputs should be handled in a manner similar to materials covered by open source or creative commons licenses — i.e., with policies and procedures which ensure use only in appropriate manners and cases. This includes determining if a project where generative AI would be used is something whose results would need to be protected and, if so, determining whether tools are available for that protection other than copyright.

It also includes avoiding high risk uses, such as using generative AI to attempt to replicate the work of a particular artist whose materials were used as training data. However, with intentionality and forethought, the risks associated with generative AI can be managed, and this new technology can bring tremendous benefits to those who deploy it intelligently.

For more information, contact any attorney with our Data, Digital Assets & Technology practice group.