BY TAYLOR FAIRCHILD
ChatGPT was unveiled to the public for testing in November 2022.11 https://openai.com/blog/chatgpt. Ever since then, opinions on generative AI—ranging from hesitant curiosity, to full-throated embrace, to abject fear—have been endlessly written about and discussed.22 www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai. While generative AI has become ubiquitous, being utilized in everything from high-powered super computers33 www.ibm.com/quantum?utm_content=SRCWW&p1=Search&p4=43700067950530792&p5=p&gclid=Cj0KCQjw_5unBhCMARIsACZyzS1aRrJgeLrf22hbz8uoDU68iyL92m6IyK0jLw0OkHeJKRLD4RW2-e4aAh4UEALw_wcB&gclsrc=aw.ds. to TikTok filters,44 www.dexerto.com/entertainment/how-to-use-tiktoks-ai-portrait-filter-2025769/. it is not necessarily well understood, which only adds to the intensity of public reaction.
IBM has defined generative AI as “deep learning models that can take raw data … and ‘learn’ to generate statistically probable outputs when prompted.”55 https://research.ibm.com/blog/what-is-generative-AI. These algorithms have the capability of generating models based on the data they are fed to create works that are similar but not actually identical. Most of the generative AI we think of, such as ChatGPT and DALL-E, uses what is called “supervised” learning; that is, it requires labeled data sets and a prompt from its user to generate outcomes. With resulting works such as images, text, videos, sounds, and schematics, a host of intellectual property issues arises. Who owns the works created by these algorithms? Can the generated works be copyrighted, trademarked, or patented? Can a generated work infringe on existing intellectual property? In the last few years, courts have battled with these questions, come to some definitive answers, drawn from historical IP philosophies, and muddied the waters.
“I’m Sorry, Thaler, I’m Afraid I Can’t Do That”: Some Answers on Ownership and AI
Perhaps the only question regarding IP and generative AI where some clarity can be found is: Can an AI algorithm be an author or an inventor? In both the patent and copyright spheres, the answer from the United States Copyright Office and the courts to date has been a resounding “no.”
In 2022, Stephen Thaler, a champion of AI, created an image, A Recent Entry to Paradise, with his “Creativity Machine,” a high-powered AI algorithm configured to generate original works of visual art.66 www.dazeddigital.com/art-photography/article/60654/1/can-ai-ever-be-creative-new-court-case-law-creativity-machine-stephen-thaler. He sought copyright protection for this image, disclosing that the work was created with his algorithm and submitting that the copyright should be held by the machine. The United States Copyright Office disagreed. Thaler posited in his request for reconsideration, “AI should be acknowledged … as an author where it otherwise meets authorship criteria, with any copyright ownership vesting in the AI’s owner.” Drawing from historical copyright cases, such as Naruto v. Slater7 and Kelley v. Chicago Park District,78 Kelley v. Chicago Park District, 635 F.3d 290 (7th Cir. 2011). the District Court ruled in August 2023 that the image was not copyrightable most fundamentally because “[h]uman authorship is the bedrock requirement of copyright.”89 Thaler v. Perlmutter, No. 1:22-cv-01564, __ F. Supp. 3d __, 2023 U.S. Dist. LEXIS 145823, at *11, 2023 WL 5333236, 2023 U.S.P.Q.2d (BNA) 980 (D.D.C. Aug. 18, 2023).
The United States Copyright Office has applied this rule unilaterally, revoking a copyright registration when it discovered a comic book, Zarya of the Dawn, used Mid-journey-generated images that had only been slightly tweaked by prompt generator and author Kris Kashtanova. In its revocation, the Copyright Office left open the possibility of conferring copyright to images generated by an AI if they are more substantially modified by a human author, but the amount of modification needed remains unclear. Since Zarya, the Copyright Office has proposed new registration guidance requiring the duty to disclose AI-generated content, noting that it will eventually publish a notice of inquiry soliciting public comments on the matter.910 www.copyright.gov/newsnet/2023/1004.html.
Patent law has also grappled with the questions proposed by Thaler, who also filed a patent application with another of his AI software listed as the sole inventor. After the United States Patent and Trademark Office (USPTO) required a human inventor, Thaler filed a similar suit, Thaler v. Vidal.1011 Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023). In that case, the Federal Circuit held “there is no ambiguity: the Patent Act requires that inventors must be natural persons; that is, human beings.”1112 Id. at 1210 (emphasis added).
Thaler has yet to submit a registration for an AI-generated trademark, but surprisingly, current understanding seems to reflect that AI-generated trademarks are eligible for registration.1213 https://gust.com/blog/rights-ai-generated-trademarks-trade-secrets/. At least Entrepeneur.com claims, “one way to [protect AI-generated works] is by registering them as trademarks” so long as existing trademark registration requirements are met.1314 www.entrepreneur.com/growing-a-business/how-to-protect-chatgpt-content-with-trademark-registration/444057.
These seemingly definitive statements have not dissuaded industry from coming up with their own solutions, however. For example, OpenAI includes in its license, “OpenAI hereby assigns to you all its right, title and interest in and to Output” effectively conferring ownership, if possible, to its users.1415 https://openai.com/policies/terms-of-use.
Does ChatGPT Dream of Infringing IP?: The Tension Between Creators and AI
If the output of generative AI is not necessarily protectable, is it capable of infringing existing IP? At least in the realm of copyright law, plenty of artists claim it is. As stated by IP attorney Anne Lanteigne in her article, “Generative Artificial Intelligence Creates Copyright Issues,” posted on her firm’s website, “[w]ith [generative AI], there are two points where copyright issues arise: at the point of input of the training data … and at the point of output of AI-generated content.”1516 www.jdsupra.com/legalnews/generative-artificial-intelligence-5888315/. Because these AI algorithms are often monetizing their produced content, a question arises as to how these algorithms are trained. And with what? One generative AI algorithm, Stable Diffusion, scraped billions of images from “everything from personal blogs hosted on WordPress and Blogspot to art platforms like DeviantArt and stock imagery sites like Shutterstock and Getty Images” to train its algorithm.1617 www.theverge.com/23444685/generative-ai-copyright-infringement-legal-fair-use-training-data. Generative-AI companies have alleged that the use of these copyrighted images and works fall squarely into fair use. The courts remain unsure.
In Authors Guild v. Google, Inc,1718 Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015). the Second Circuit found that Google’s digitizing of the entirety of copyrighted books and providing access to snippets of text fell under fair use. Google created a search engine with the ability to “discern fluctuations of interest in a particular subject over time and space … [and] allow researchers to comb over … books Google has scanned in order to examine ‘word frequencies, syntactic patterns, and thematic markers.’”1819 Id. at 209. It was held that the use of these snippets was sufficiently transformative, with the court stating, “[c]omplete unchanged copying has repeatedly been found justified fair use when the copying was reasonably appropriate to achieve the copier’s transformative purpose and was done in a manner that it [does] not offer a competing substitute for the original.”1920 Id. at 221. The parallels between the facts of Authors Guild and the use of copyrighted work for training data is apparent. The court also stated,“[w]hile Google makes an unauthorized digital copy of the entire book, it does not reveal that digital copy to the public.”2021 Id. (emphasis added). Similarly, these generative AI algorithms do not distribute or even in some cases reveal the works they are using to generate their outputs.
This tactic can make it difficult to prove copyright infringement. In J. Doe1 and J. Doe2 v. GitHub, Inc.,2122 J. Doe1 and J. Doe2 v. GitHub, Inc., No. 4:22-cv-06823-JST (N.D. Cal). Order Granting in Part and Denying in Part Motions to Dismiss, 2023 U.S. Dist. LEXIS, 2023 WL 7132064 (N.D. Cal. Oct. 30, 2023). two developers alleged that GitHub used code written by them and hosted on GitHub’s site to train their generative AI, CoPilot. The developers alleged that they could tell their code had been used because it appeared in CoPilot’s generated results. CoPilot “auto-fills” lines of code as a developer writes it, providing suggestions to achieve desired functions. The court dismissed the copyright infringement claim. Not only was the code uploaded to GitHub subject to an opensource agreement, but there was also no evidence that GitHub used or suggested the developers’ code specifically. This case has since been dismissed, based on inadequate pleading. This seems to be a pattern in these AI cases; while the harm may be real, plaintiffs have not yet provided enough evidence to prove it.
This year, class action lawsuits have been filed against both OpenAI and Stability AI alleging similar infringement. In Tremblay v. OpenAI, inc.,2223 Tremblay v. OpenAI, Inc., No. 3:23-cv-03223 (N.D. Cal). horror author Paul Tremblay alleges OpenAI acquired copies of illegally published books on the internet and trained ChatGPT to create accurate summaries. Similarly, in Andersen v. Stability AI Ltd.,2324 Andersen v. Stability AI LTD., No. 3:23-cv-00201 (N.D. Cal). cartoonist Sarah Andersen alleged that at least 16 collections of her copyright-registered cartoons were potentially used by Midjourney as training data.2425 Opposition to Motion to Dismiss filed by Sarah Andersen, available at https://docs.justia.com/cases/federal/district-courts/california/candce/3:2023cv00201/407208/67. Stability AI has filed a motion to dismiss2526 Amended Motion to Dismiss filed by Stability AI, available at https://cases.justia.com/federal/district-courts/california/candce/3:2023cv00201/407208/73/0.pdf. under very similar grounds as those that were successful in GitHub. In October, the U.S. District Court dismissed all claims expect for the direct infringement claim against Stability AI.2627 Order on Motions to Dismiss and Strike, 2023 U.S. Dist. LEXIS, 2023 WL 7132064 (N.D. Cal. Oct. 30, 2023). See also www.reuters.com/legal/litigation/us-judge-finds-flaws-artists-lawsuit-against-ai-companies-2023-07-19/.
Analogous arguments regarding fair use have been presented about the output of generative AI algorithms.2728 www.npr.org/2023/08/16/1194202562/new-york-times-considers-legal-action-against-openai-as-copyright-tensions-swirl. However, as some artists have pointed out, if the training data used to train generative AI includes copyrighted material, the product generated could include portions of these materials. Some have argued that even the output of an algorithm could be considered a derivative work. After all, a common prompt when using generative AI algorithms includes the clause, “in the style of,” followed by an artist’s name.2829 www.byteside.com/previously/2022/09/ai-art-named-artists-monet-picasso-rutkowski/.
In Andy Warhol Foundation for the Visual Arts v. Goldsmith,2930 Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 143 S. Ct. 1258, 215 L. Ed. 2d 473 (2023). Lynn Goldsmith alleged copyright infringement of one of her photographs, used as an artistic reference to produce a collection of paintings by Andy Warhol (the Prince Series). The court held that aesthetic or expressive changes do not necessarily make a work transformative. Further, if the purpose and character of the work is not “sufficiently distinct” from the original, it may not fall within fair use. The court declined to extend this ruling into the world of AI, but many point to it as a potential basis for alleging infringement in the future.
Issues around trademark infringement have also arisen. In the recently filed Getty Images, Inc v Stability AI, Inc,3031 Getty Images, Inc v Stability AI, Inc, No.:23-cv-00135-UNA (D. Del. Feb. 3, 2023). plaintiff Getty Images alleges, among other things, infringement of its trademarks based on a series of images produced by Stable Diffusion bearing their logo in both slightly modified and identical forms. In their amended complaint, Getty Images states, “[o]ften, the output generated has contained a modified version of a Getty Images watermark, creating confusion as to the source of the images and falsely implying an association.”3132 Amended complaint against Stability AI filed by Getty Images, available at https://dockets.justia.com/docket/delaware/dedce/1:2023cv00135/81407. Getty Images also alleges the sometimes “grotesque” results denigrate Getty Images’ trademarks and reputation. The question of infringement remains open, but Getty Images has recently requested oral argument, and it will be interesting to see how they choose to proceed.3233 Getty Images request for oral argument, available at https://dockets.justia.com/docket/delaware/dedce/1:2023cv00135/81407.
Conclusion
Pending litigation filed since generative AI’s entrance onto the world stage may provide more questions than answers at this point. Based on what has been decided in the last few years, it appears that works created with AI algorithms cannot be protected as copyrights or patents, but potentially human intervention could tip those scales. Further, using copyrighted images to train generative AI may be fair game, even if the outputs could potentially infringe.
What is left out of these motions to dismiss and written opinions, however, is the attention these cases have generated. Users and consumers remain troubled by these questions, and we are seeing AI companies act of their own accord in response. After the lawsuit filed against them was dismissed, GitHub introduced a new filter that prevents displaying code that matches 150 characters of publicly uploaded GitHub code.3334 https://techcrunch.com/2023/01/27/the-current-legal-cases-against-generative-ai-are-just-the-beginning/. Stability AI has proposed an opt-out system for artists not wanting their art to be used as training data in the next iteration of Stable Diffusion and also set up the website “HaveIBeenTrained.com” to better alert artists to which of their works have been used. OpenAI has decided to license works from stock image hosting sites like Shutterstock, instead of broadly scraping the internet. The true battleground for and against generative AI may, in the end, be outside the courtroom and instead in the court of public opinion.
About the author
Taylor Fairchild is a patent attorney at Christensen O’Connor Johnson Kindness. She has a bachelor’s in biochemistry and a master’s in electrical engineering. She is also a board member of Washington Lawyers for the Arts, and an avid crocheter, musician, and hobbyist in her spare time. She can be reached at:
NOTES
1. https://openai.com/blog/chatgpt.
2. www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai.
4. www.dexerto.com/entertainment/how-to-use-tiktoks-ai-portrait-filter-2025769/.
5. https://research.ibm.com/blog/what-is-generative-AI.
7. Naruto v. Slater, No. 16-15469 (9th Cir. Apr. 23, 2018).
8. Kelley v. Chicago Park District, 635 F.3d 290 (7th Cir. 2011).
9. Thaler v. Perlmutter, No. 1:22-cv-01564, __ F. Supp. 3d __, 2023 U.S. Dist. LEXIS 145823, at *11, 2023 WL 5333236, 2023 U.S.P.Q.2d (BNA) 980 (D.D.C. Aug. 18, 2023).
10. www.copyright.gov/newsnet/2023/1004.html.
11. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023).
12. Id. at 1210 (emphasis added).
13. https://gust.com/blog/rights-ai-generated-trademarks-trade-secrets/.
15. https://openai.com/policies/terms-of-use.
16. www.jdsupra.com/legalnews/generative-artificial-intelligence-5888315/.
17. www.theverge.com/23444685/generative-ai-copyright-infringement-legal-fair-use-training-data.
18. Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).
19. Id. at 209.
20. Id. at 221.
21. Id. (emphasis added).
22. J. Doe1 and J. Doe2 v. GitHub, Inc., No. 4:22-cv-06823-JST (N.D. Cal). Order Granting in Part and Denying in Part Motions to Dismiss, 2023 U.S. Dist. LEXIS, 2023 WL 7132064 (N.D. Cal. Oct. 30, 2023).
23. Tremblay v. OpenAI, Inc., No. 3:23-cv-03223 (N.D. Cal).
24. Andersen v. Stability AI LTD., No. 3:23-cv-00201 (N.D. Cal).
25. Opposition to Motion to Dismiss filed by Sarah Andersen, available at https://docs.justia.com/cases/federal/district-courts/california/candce/3:2023cv00201/407208/67.
26. Amended Motion to Dismiss filed by Stability AI, available at https://cases.justia.com/federal/district-courts/california/candce/3:2023cv00201/407208/73/0.pdf.
27. Order on Motions to Dismiss and Strike, 2023 U.S. Dist. LEXIS, 2023 WL 7132064 (N.D. Cal. Oct. 30, 2023). See also www.reuters.com/legal/litigation/us-judge-finds-flaws-artists-lawsuit-against-ai-companies-2023-07-19/.
29. www.byteside.com/previously/2022/09/ai-art-named-artists-monet-picasso-rutkowski/.
30. Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 143 S. Ct. 1258, 215 L. Ed. 2d 473 (2023).
31. Getty Images, Inc v Stability AI, Inc, No.:23-cv-00135-UNA (D. Del. Feb. 3, 2023).
32. Amended complaint against Stability AI filed by Getty Images, available at https://dockets.justia.com/docket/delaware/dedce/1:2023cv00135/81407.
33. Getty Images request for oral argument, available at https://dockets.justia.com/docket/delaware/dedce/1:2023cv00135/81407.