By Larry Sandell, Intellectual Property Attorney
When it concerns generative AI, Intellectual Property (IP) law is murky and far from settled, but some answers, legal strategies, and best practices can be derived from existing law. Copyright law in particular is already struggling with how to fairly allocate creator attribution and ownership of AI outputs, whether and to what degree AI outputs infringe upon underlying work, and even what liability may attach to the use of copyrighted dataset inputs. While the questions on this topic are legion, this article seeks to provide an overview of the current state of US IP law as it applies to generative AI and some insight as to where it is headed.
Who owns the content created by generative AI?
Unsurprisingly, the answer here is very fact-dependent—but also without established legal parameters. A good place to start is with the user terms governing a generative AI system. As a prominent example, OpenAI purports to assign ownership of text and art outputs to the user only if the user complies with its terms of service. However, OpenAI’s terms prohibit the use of generative AI “in a way that infringes, misappropriates or violates any person’s rights.” Here, OpenAI appears to pass the buck on IP infringement liability, which, as discussed below, cannot be readily or reliably assessed. As a result, a user cannot be assured of her ownership of OpenAI’s output.
Are generative AI outputs protectable by copyright?
It is an open question whether AI generated text and images are “original works of authorship fixed in a tangible medium of expression” and therefore copyrightable.
The US Copyright Office has determined that generative AI cannot be an “author” under copyright law. It unequivocally rejected the attempted registration of a digital painting, where the would-be copyright registrant declared a generative AI platform to be the work’s “author.” The registration request was denied because the work lacked “human authorship.” This rejection is now being litigated in Federal District Court in DC. Nonetheless, it appears unlikely that non-humans can be “authors” under US copyright law.
A more practical question on this topic is whether a human can be an author when generative AI is heavily used as a creative tool. The issue has been squarely raised by Zarya of the Dawn, a graphic novel prepared with substantial generative AI assistance. A human creator filed for copyright registration on a comic book made with MidJourney-generated images While the Copyright Office initially registered copyrights on both the images and the overall compilation of the book, it ultimately canceled the copyright on the AI-generated images. The arguments that AI’s contribution was merely “assisting”; that the book reflected the human author’s creative, iterative AI prompts and overall artistic vision; and that the AI outputs were manually modified before publication were all rejected by the Office. An appeal is likely.
The law is unsettled, but it seems likely that generative AI may ultimately be accepted as an “assisting” tool that does not preclude copyright. The advent of photography notably raised a similar debate: some argued that photographers merely engaged in a rote process without creative input. But, in the late 1800s, Congress made photographs copyrightable. Since then, the Supreme Court has held that copyright merely requires “at least some minimal degree of creativity” “no matter how crude, humble or obvious” the “creative spark” might be. Generative AI requires at least a minimal degree of creative human input, such as ,drafting input prompts or even selecting a training dataset. Ultimately, the courts or Congress may need to definitively weigh in on this important issue.
Are generative AI outputs protectable by patent law?
This question has been answered with clarity—at least for now. In August 2022, the Federal Circuit Court of Appeals affirmed “‘inventors’ must be human.”
A more practical issue is whether generative AI’s contribution to technological invention can undermine patentability. Under US patent law, inventorship is split into two parts—(1) “conception” of the idea underlying the claimed invention and (2) “reduction to practice,” that is, bring the idea to technical fruition. Only persons involved in “conception” are “inventors”; those that merely reduce the invention to practice—regardless of the technical skill and effort contributed—are not.
While no case has addressed what level of AI contribution precludes patentability, human inventors should be able to patent their inventions at least when a generative AI contribution is limited to “reduction to practice.”
Where is the line between “fair use” and copyright infringement via creation of derivative works?
Generative AI training datasets commonly include copyrighted works—including art, photos, writings, sound recordings, video, and code. These inputs are inherently used to generate AI output. Yet, whether and to what degree these uses of copyrighted works are permissible are open questions.
On one hand, copyright holder possesses the exclusive right to create derivative works—namely, creative works based on the original (copyrighted) work . AI synthesizes and utilizes images, literature, and code when it generates output—but does this mean that generative AI output is a derivative work? If so, each AI platform (and its owner) may be infringing copyrights on a regular basis. On the other hand, the “fair use” doctrine can immunize certain uses of copyrighted works based on, for example, (1) the commercial vs non-profit/educational nature of the use; (2) if the original works are more factual/technical vs. artistic/creative (because copyright protects expression, not underlying data); (3) the amount of the original works used and how identifiable the elements may be; and (4) the effect on the commercial value of the original work.
No court has yet offered insight into where the line is, but there are three important cases to watch:
First, in November 2022, GitHub, Microsoft, and OpenAI were sued by anonymous coders who contributed to Github’s open source code repository under the open source GNU General Public License. The coders argue that training the for-profit Codex and CoPilot generative AI platforms on Github’s open source code repository both breaches the contractual terms of the GNU license and impermissibly removes copyright management text (e.g., human-readable references to the license in each code section). The tech companies have argued that the case should be dismissed because there is no evidence of actual violations in AI output—merely assumptions of wrongdoing based on the training dataset.
Second, in another class action suit, Stability AI, MidJourney, and Deviant Art were sued in January 2023 by a group of artists. The artists assert copyright, publicity, and unfair competition claims, arguing that “AI image generators are 21st-century collage tools that violate the rights of millions of artists” and, in particular, that that the AI generation of commission-free art made “in the style of” a particular artist erodes their commercial opportunities.
Third, in February 2023, Getty Images filed suit against Stability AI asserting copyright, trademark and unfair competition claims. Getty asserts “Stability AI has copied more than 12 million photographs from Getty Images’ collection, along with the associated captions and metadata, without permission from or compensation to Getty Images, as part of its efforts to build a competing business.” Getty identifies Stable Diffusion’s output images that substantially copy original works, remove or alter copyright notices, and mangle the Getty trademark.
These cases will begin to shape US law on generative AI and copyright law. Ultimately, however, Congress may consider resolving some of these issues by mandating a compulsory licensing that requires generative AI platforms to identify copyrighted works materially utilized in a particular AI output (something that might be difficult given that sources of output are often unidentifiable in current AI models)—and attribute and compensate the copyright holders accordingly. An analogous US law has long provided for compulsory licensing payments to compensate songwriters for others’ recordings of their covers.
What can content creators and businesses do to protect themselves in this uncertain and shifting legal landscape?
The prospect of widespread generation of infringing derivative works by AI platforms, has made it even more important to ensure the potential availability for statutory copyright damages. Statutory damages enable a copyright holder to receive monetary damages in a lawsuit without specifically proving the commercial value of damage caused by the infringement—a task for which evidence can be difficult or impossible to find. Eligibility for statutory damages requires copyright registration within three months of publication or infringement of a work. As a result, content creators of all stripes should consider filing with the US copyright office every three months to maximize enforcement possibilities in the future. Such filings are surprisingly low cost, and the Copyright Office even permits the registration of computer code in a manner that maintains trade secret status for the code.
Companies with generative AI platforms—and those that publish generative AI output—should also consider prominently offering members of the public the opportunity to directly lodge complaints about infringement. As a practical matter, this may stave off some lawsuits or demonstrate good faith in court. For example, OpenAI’s terms of service wisely include a procedure for receiving copyright complaints, potentially taking advantage of the copyright liability safe harbor provisions of the Digital Millennium Copyright Act.
Conclusion
Inevitably, the evolution of IP law on generative AI will continue to lag behind the rapid evolution of the technology itself. This poses strategic problems for generative AI firms and their clients because IP law—and copyright law in particular—may ultimately render certain products and services commercially unviable due to legal liability. For now, however, the best course of action appears to be: innovate, stay informed, be fair, be ethical, and hedge legal risks.
Key Takeaways:
- AI output under US IP law, there are currently more questions than answers
- Legal ownership of works created by generative AI platforms cannot be reliably determined in most instances, including, for example, under OpenAI’s terms of service.
- Under current US IP law, generative AI cannot be an author for the purposes of copyright or an inventor on a patent. The level of permissible contribution by generative AI before precluding patentability or copyrightability, however, is unclear.
- Courts are just beginning to set the contours of how generative AI output should be treated under US Copyright law—most notably defining the fine line between what is an infringing derivative work and what is “fair use.”
- The dawn of widespread generative AI has made copyright statutory damages — and therefore prompt copyright registration — more important.
Biography
Larry Sandell is an experienced IP strategist, patent attorney, litigator, and appellate advocate, with a technical focus that includes software—including generative AI and blockchain, consumer electronics, medical devices, cannabis, and food science. Larry has argued in the U.S. Courts of Appeal for the Federal Circuit, the Ninth Circuit, and the D.C. Circuit, and has litigated before numerous Federal District Court, the Patent Trial and Appeal Board (PTAB), the Trademark Trial and Appeal Board (TTAB), and the US International Trade Commission (ITC). As IP law catches up to the AI age, Larry is applying his skills and expertise to helping clients not only manage AI-related risk, but position themselves to thrive in this brave new world. Eager to experiment with new AI tech applications, Larry’s hobbies include using ChatGPT and DALL-E to write and illustrate love poems to his wife and stories for his young children. Larry practices law with Mei & Mark LLP and can be reached at [email protected].
