The Input Question in AI Copyright Law
by Thomas James, Minnesota attorney
In a previous blog post, I discussed the question whether AI-generated creations are protected by copyright. This could be called the “output question” in the artificial intelligence area of copyright law. Another question is whether using copyright-protected works as input for AI generative processes infringes the copyrights in those works. This could be called the “input question.” Both kinds of questions are now before the courts. In this blog post, I describe a framework for analyzing the input question.
The Getty Images lawsuit
Getty Images is a stock photograph company. It licenses the right to use the images in its collection to those who wish to use them on their websites or for other purposes. Stability AI is the creator of Stable Diffusion, which is described as a “text-to-image diffusion model capable of generating photo-realistic images given any text input.” In January, 2023, Getty Images initiated legal proceedings in the United Kingdom against Stability AI. Getty Images is claiming that Stability AI violated copyrights by using their images and metadata to train AI software without a license.
The independent artists lawsuit
Another lawsuit raising the question whether AI-generated output infringes copyright has been filed in the United States. In this case, a group of visual artists are seeking class action status for claims against Stability AI, Midjourney Inc. and DeviantArt Inc. The artists claim that the companies use their images to train computers “to produce seemingly new images through a mathematical software process.” They describe AI-generated artwork as “collages” made in violation of copyright owners’ exclusive right to create derivative works.
The GitHut Copilot lawsuit
In November, 2022, a class action lawsuit was filed in a U.S. federal court against GitHub, Microsoft, and OpenAI. The lawsuit claims the GitHut Copilot and OpenAI Codex coding assistant services use existing code to generate new code. By training their AI systems on open source programs, the plaintiffs claim, the defendants have allegedly infringed the rights of developers who have posted code under open-source licenses that require attribution.
How AI Works
AI, of course, stands for artificial intelligence. Almost all AI techniques involve machine learning. Machine learning, in turn, involves using a computer algorithm to make a machine improve its performance over time, without having to pre-program it with specific instructions. Data is input to enable the machine to do this. For example, to teach a machine to create a work in the style of Vincent van Gogh, many instances of van Gogh’s works would be input. The AI program contains numerous nodes that focus on different aspects of an image. Working together, these nodes will then piece together common elements of a van Gogh painting from the images the machine has been given to analyze. After going through many images of van Gogh paintings, the machine “learns” the features of a typical Van Gogh painting. The machine can then generate a new image containing these features.
In the same way, a machine can be programmed to analyze many instances of code and generate new code.
The input question comes down to this: Does creating or using a program that causes a machine to receive information about the characteristics of a creative work or group of works for the purpose of creating a new work that has the same or similar characteristics infringe the copyright in the creative work(s) that the machine uses in this way?
The Exclusive Rights of Copyright Owners
In the United States, the owner of a copyright in a work has the exclusive rights to:
- reproduce (make copies of) it;
- distribute copies of it;
- publicly perform it;
- publicly display it; and
- make derivative works based on it.
(17 U.S.C. § 106). A copyright is infringed when a person exercises any of these exclusive rights without the copyright owner’s permission.
Copyright protection extends only to expression, however. Copyright does not protect ideas, facts, processes, methods, systems or principles.
Infringement can be either direct or indirect. Direct infringement occurs when somebody directly violates one of the exclusive rights of a copyright owner. Examples would be a musician who performs a copyright-protected song in public without permission, or a cartoonist who creates a comic based on the Batman and Robin characters and stories without permission.
The kind of tool an infringer uses is not of any great moment. A writer who uses word-processing software to write a story that is simply a copy of someone else’s copyright-protected story is no less guilty of infringement merely because the actual typewritten letters were generated using a computer program that directs a machine to reproduce and display typographical characters in the sequence a user selects.
Contributory and Vicarious Infringement
Infringement liability may also arise indirectly. If one person knowingly induces another person to infringe or contributes to the other person’s infringement in some other way, then each of them may be liable for copyright infringement. The person who actually committed the infringing act could be liable for direct infringement. The person who knowingly encouraged, solicited, induced or facilitated the other person’s infringing act(s) could be liable for contributory infringement.
Vicarious infringement occurs when the law holds one person responsible for the conduct of another because of the nature of the legal relationship between them. The employment relationship is the most common example. An employer generally is held responsible for an employee’s conduct, provided the employee’s acts were performed within the course and scope of the employment. Copyright infringement is not an exception to that rule.
Programmer vs. User
Direct infringement liability
Under U.S. law, machines are treated as extensions of the people who set them in motion. A camera, for example, is an extension of the photographer. Any images a person causes a camera to generate by pushing a button on it is considered the creation of the person who pushed the button, not of the person(s) who manufactured the camera, much less of the camera itself. By the same token, a person who uses the controls on a machine to direct it to copy elements of other people’s works should be considered the creator of the new work so created. If using the program entails instructing the machine to create an unauthorized derivative work of copyright-protected images, then it would be the user, not the machine or the software writer, who would be at risk of liability for direct copyright infringement.
Contributory infringement liability
Knowingly providing a device or mechanism to people who use it to infringe copyrights creates a risk of liability for contributory copyright infringement. Under Sony Corp. v. Universal City Studios, however, merely distributing a mechanism that people can use to infringe copyrights is not enough for contributory infringement liability to attach, if the mechanism has substantial uses for which copyright infringement liability does not attach. Arguably, AI has many such uses. For example, it might be used to generate new works from public domain works. Or it might be used to create parodies. (Creating a parody is fair use; it should not result in infringement liability.)
The situation is different if a company goes further and induces, solicits or encourages people to use its mechanism to infringe copyrights. Then it may be at risk of contributory liability. As the United States Supreme Court has said, “one who distributes a device with the object of promoting its use to infringe copyright, as shown by clear expression or other affirmative steps taken to foster infringement, is liable for the resulting acts of infringement by third parties.” Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd., 545 U.S. 913, 919 (2005). (Remember Napster?)
If AI-generated output is found to either directly or indirectly infringe copyright(s), the infringer nevertheless might not be held liable, if the infringement amounts to fair use of the copyrighted work(s) that were used as the input for the AI-generated work(s).
Ever since some rap artists began using snippets of copyright-protected music and sound recordings without permission, courts have embarked on a treacherous expedition to articulate a meaningful dividing line between unauthorized derivative works, on one hand, and unauthorized transformative works, on the other. Although the Copyright Act gives copyright owners the exclusive right to create works based on their copyrighted works (called derivative works), courts have held that an unauthorized derivative work may be fair use if it is “transformative.: This has caused a great deal of uncertainty in the law, particularly since the U.S. Copyright Act expressly defines a derivative work as one that transforms another work. (See 17 U.S.C. § 101: “A ‘derivative work’ is a work based upon one or more preexisting works, . . . or any other form in which a work may be recast, transformed, or adapted.” (emphasis added).)
When interpreting and applying the transformative use branch of Fair Use doctrine, courts have issued conflicting and contradictory decisions. As I wrote in another blog post, the U.S. Supreme Court has recently agreed to hear and decide Andy Warhol Foundation for the Visual Arts v. Goldsmith. It is anticipated that the Court will use this case to attempt to clear up all the confusion around the doctrine. It is also possible the Court might take even more drastic action concerning the whole “transformative use” branch of Fair Use.
Some speculate that the questions the Justices asked during oral arguments in Warhol signal a desire to retreat from the expansion of fair use that the “transformativeness” idea spawned. On the other hand, some of the Court’s recent decisions, such as Google v. Oracle, suggest the Court is not particularly worried about large-scale copyright infringing activity, insofar as Fair Use doctrine is concerned.
To date, it does not appear that there is any direct legal precedent in the United States for classifying the use of mass quantities of works as training tools for AI as “fair use.” It seems, however, that there soon will be precedent on that issue, one way or the other. In the meantime, AI generating system users should proceed with caution.