University of St. Thomas Law Journal

Artificial Intelligence, Real Musicality: Reconciling Generative AI with Copyright Law

*Jack Thram

In April of this year, “Heart on My Sleeve,” a song featuring vocals from pop megastars Drake and The Weeknd, went live on various social media sites and streaming platforms to tremendous fanfare.1 The song quickly amassed fifteen million views on TikTok and six hundred thousand streams on Spotify.2 Then, almost as abruptly as it had appeared, it was taken down.3

The problem was that neither Drake nor The Weeknd had actually been involved in the project.4 In fact, the sole artist behind the song was an anonymous, bedsheet-donning songwriter known simply as Ghostwriter.5 Ghostwriter wrote the lyrics for the track and used a generative artificial intelligence (AI) program to make it sound like it was sung by Drake and The Weeknd.6

While the result was a viral sensation, not everyone had praise for Ghostwriter. On April 17, Universal Music Group (UMG), the parent company of the label to which both Drake and The Weeknd are signed,7 sent DMCA takedown notices and succeeded in getting the song removed from many sites.8 UMG also issued an accompanying statement, in which it asserted that “the training of generative AI using [their] artists’ music” constitutes “a violation of copyright law” and labeled songs made using generative AI programs “infringing content.”9

This spelled the end for “Heart on My Sleeve” for the moment,10 but what transpired brought to the fore a much broader issue: the applicability of copyright law in the era of AI-generated music. AI-generated music existed before the advent of “Heart on My Sleeve,”11 and will be around long after. Indeed, Ghostwriter has since released a follow-up song, “Whiplash,” in the style of rappers Travis Scott and 21 Savage.12 It is clear that, in the words of Stanford music professor Ge Wang, “[t]he cat is not going back in the bag.”13 Bearing all this in mind, lawmakers will inevitably have to reckon with AI-generated music and its effects on copyright.

Specifically, “Heart on My Sleeve” highlighted a trio of important questions regarding the interplay between AI technology and copyright law. First, do the creators and owners of generative AI programs infringe on copyright holders’ rights by “training” their programs, using copyrighted works as input? Second, does the resulting output—that is, the music created by the programs—constitute infringement? And third, is the output eligible for copyright protections?

Article I, Section 8, Clause 8 of the U.S. Constitution grants Congress the power “[t]o promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Investors the exclusive Right to their respective Writings and Discoveries.”14 This so-called “Copyright Clause” provides the basis for copyright law in the United States.15 The Copyright Act of 1790 was the first piece of federal legislation to recognize copyrights and define their scope.16 Over the centuries, there have been multiple expansions of and modifications to copyright law, including the Copyright Act of 1976 and the Sonny Bono Copyright Term Expansion Act of 1998.17 In sum, copyright law now covers “original works of authorship fixed in any tangible medium,”18 and identifies both musical compositions and sound recordings as such.19 The effect of copyright protection, as applicable here, is that copyright holders are given the exclusive right to distribute, reproduce, and create derivatives of their protected works.20

Two other important concepts related to copyright are derivative works and fair use. Derivative works are those “based on or derived from one or more already existing works.”21 For example, a movie adaptation of a book is considered a derivative work.22 If the movie is created by the book’s author (who is the copyright holder), there is obviously no infringement—the author has the right to create derivative works using their own copyrighted works. However, if the movie is made by anyone other than the author, it will be seen as infringement, unless they receive consent (often in the form of licensing agreements) or it is found to be a fair use.

Fair use is the right to use copyrighted works without the permission of the copyright holder.23 This is only permitted under certain circumstances and is an affirmative defense.24 The factors that determine whether a use is a fair use are (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the substantiality of the portion used in relation to the copyrighted work as a whole, and (4) the effect of the use on the value or marketability of the copyrighted work.25 With regard to the first factor, the focus is primarily on whether the copyrighted work is being used for a commercial or noncommercial purpose; for example, education is almost always a fair use.26 However, the first factor also considers the character of the use, which generally translates to the transformative nature of the derivative work—that is, whether it results in new expression, meaning, or message.27 With regard to the second factor, when a work is unpublished or highly creative, it will generally be afforded more protection than a work that is published or more fact-based.28 With regard to the third factor, the substantial similarity standard is used to determine whether infringement has occurred.29 Finally, with respect to the fourth factor, courts look both at whether a derivative work resulting from the use directly harms sales of the copyrighted work, as well as whether it usurped the copyright holder’s right to create a derivative work of their own; in deciding on the latter consideration, courts must analyze how likely it was that the copyright holder reasonably would have—or could have—done so.30

So how does all of this bear on the questions posed above? The first question turns on whether the training of AI programs is fair use. The second question turns on whether the output from such programs constitutes a derivative work, and, if so, whether that is fair use. The third question turns on the interpretation of the term “author” as it appears in the Constitution.

The answer to the first question is unclear. For starters, the argument that AI training is infringement relies on the copying of copyrighted material for “ingestion” by the AI. However, it is not apparent whether generative AI programs actually copy input works during the scraping process, rather than simply viewing them.31 “If an [AI program] is working as intended, it does not copy the training data at all,” said Matthew Sag, Professor of Law in Artificial Intelligence, Machine Learning, and Data Science at Emory University School of Law, before a Senate subcommittee.32 Even if copying does occur, then the AI programs’ creators and owners can still assert that it is fair use. Their arguments would likely rely on the first and fourth factors of the fair use analysis. As Louis Tompros, an intellectual property partner at WilmerHale, explains, “[t]he argument might be that AI training is not something that’s being done to make a copy for commercial purposes; it is instead deliberately transformative and trying to create something new, and there is no direct impact on the market for the original by allowing an AI to hear or see a copy of the original to be able to create to train its algorithm.”33 Many academics and commentators seem to agree with this assessment, and argue that this is no different than a student internalizing material from a textbook, or an artist perusing the work in their local art museum to gain inspiration, another point Sag raised before the Senate subcommittee: “Rather than thinking of an [AI program] as copying the training data like a scribe in a monastery, it makes more sense to think of it as learning from the training data like a student.”34 However, not everyone subscribes to this theory of AI learning. Critics of this viewpoint contend that the human learning process is inherently transformative, whereas the AI learning process is not.35 But also worth considering is that copyright protects only the fixed expression of facts and ideas—not the underlying facts and ideas themselves. So, if it is that AI programs are simply learning from copyrighted material, and not copying it, then it seems likely there has not been infringement. However, in the absence of legislation, this determination will be left to the courts.

The second question is also complicated. As noted above, to constitute a derivative work, a work must be “based on or derived from” a copyrighted work. With regard to AI output, the argument could be made either way. One could say that the output is “based on or derived from” the entire catalog of the artist the AI program seeks to emulate, because at a fundamental level, it is; the AI program evaluates the copyrighted works, then bases its output off of patterns it has detected.36 On the other hand, songs “in the style of” an artist do not usually qualify as derivative works.37 Says Tompros, “If I, as a person, listen to a whole bunch of Drake music and write my own song, having been inspired by him, but I don’t use the same lyrics or the same music, everybody would agree that’s not a copy. Why wouldn’t the same standard apply to an AI?”38 The latter viewpoint appears particularly strong when the output does not include any material from the copyrighted works, since it would be near-impossible to argue substantial similarity when there is no portion of the original used. However, the use will still be scrutinized under the fourth factor of the fair use analysis as to whether the artist could have realistically made the song themself. In theory, this is also a hard argument to make. To say that the artist could have generated the song themself is to say that they would have come up with the exact same lyrics, arrangement, and other elements. However, courts will determine exactly what this burden looks like for artists trying to rebut claims of fair use, and it may not be quite so high.

As for the third and final question, it will come as a relief that there is some degree of clarity. Federal courts have consistently held that the meaning of the term “author” in the Copyright Clause is limited to humans.39 The U.S. Copyright Office has also made its position clear, stating that in its view, “it is well-established that copyright can protect only material that is the product of human creativity. Most fundamentally, the term ‘author,’ which is used in both the Constitution and the Copyright Act, excludes non-humans.”40 This does not preclude the application of copyright to all music made with the assistance of AI. The Copyright Office explained that the test is “whether the traditional elements of authorship in the work ([]musical expression or elements of selection, arrangement, etc.)” originated from the AI program or its user.41 In the former case, the application to register the work would be denied; in the latter case, it would be accepted. As the Copyright Office acknowledges, “[t]his is necessarily a case-by-case inquiry.”42 However, while the Copyright Office may decide whether or not to accept a copyright application, its position is not codified by any means, and will be tested in the courts. This means that, at least for the time being, case law will inform exactly what level of human input or intervention is required for AI-generated music to be eligible for copyright protections.

The true issue at the heart of all three questions is that while many have tried to apply the current copyright law framework to AI music, it does not fit squarely within that framework. New, comprehensive legislation is needed that specifically addresses AI-generated music. While this will require time and effort on the part of legislators, industry advocates, and others, it is the only way forward without relying solely on the case law developing around this issue on an ad-hoc basis. Regardless of what the future holds for AI music, copyright law, and their interplay, one thing is certain: “[T]he future of music is here,”43 and it will require copyright law to adapt.

* Jack Thram, J.D. Candidate, University of St. Thomas School of Law Class of 2024. Publications Editor, University of St. Thomas Law Journal.

  1. See, e.g., Chris Willman, AI-Generated Fake ‘Drake’/‘Weeknd’ Collaboration, ‘Heart on My Sleeve,’ Delights Fans and Sets off Industry Alarm Bells, Variety (Apr. 17, 2023, 11:02 AM), https://variety.com/2023/music/news/fake-ai-generated-drake-weeknd-collaboration-heart-on-my-sleeve-1235585451/. ↩︎
  2. Laura Snapes, AI Song Featuring Fake Drake and Weeknd Vocals Pulled from Streaming Services, Guardian (Apr. 18, 2023, 5:37 AM), https://www.theguardian.com/music/2023/apr/18/ai-song-featuring-fake-drake-and-weeknd-vocals-pulled-from-streaming-services; Igor Bonifacic, AI-Generated Drake and The Weeknd Song Pulled from Streaming Platforms, Engadget (Apr. 19, 2023), https://www.engadget.com/copyright-in-spotlight-after-streaming-platforms-pull-ai-generated-drake-song-183513972.html. ↩︎
  3. Willman, supra note 1. ↩︎
  4. Samantha Murphy Kelly, The Viral New ‘Drake’ and ‘Weeknd’ Song Is Not What It Seems, CNN Bus. (Apr. 19, 2023, 9:13 AM), https://www.cnn.com/2023/04/19/tech/heart-on-sleeve-ai-drake-weeknd/index.html. ↩︎
  5. ghostwriter977 (@imghostwrit3r), Twitter (Apr. 16, 2023, 10:31 AM), https://twitter.com/imghostwrit3r/status/1647623877214732290. ↩︎
  6. Id. ↩︎
  7. Artists, Republic Recs., https://www.republicrecords.com/artists/ (last visited Sep. 7, 2023). Republic Records is owned by UMG. See Our Labels & Brands, Universal Music Grp., https://www.universalmusic.com/labels/ (last visited Sep. 14, 2023). ↩︎
  8. Willman, supra note 1. For further information on the DMCA takedown process, see The DMCA Notice and Takedown Process, Copyright All., https://copyrightalliance.org/education/copyright-law-explained/the-digital-millennium-copyright-act-dmca/dmca-notice-takedown-process/ (last visited Sept. 7, 2023). ↩︎
  9. Tim Ingham, Universal Music Group Responds to ‘Fake Drake’ AI Track: Streaming Platforms Have ‘A Fundamental Responsibility to Prevent the Use of Their Services in Ways that Harm Artists’, Music Bus. Worldwide (Apr. 17, 2023), https://www.musicbusinessworldwide.com/universal-music-group-responds-to-fake-drake-ai-track-streaming-platforms-have-a-fundamental-responsibility/. ↩︎
  10. The song has now been submitted for Grammy consideration—and has been deemed eligible by the CEO of the Recording Academy. So, it appears the story is far from over. Carl Lamarre, A.I.-Generated Song that Mimics Drake & The Weeknd Submitted for Grammy Consideration, Billboard (Sept. 6, 2023), https://www.billboard.com/music/awards/ai-generated-drake-weeknd-song-grammy-submitted-1235407148/. ↩︎
  11. See, e.g., The Daily, The Ballad of ‘Deepfake Drake’, N.Y. Times (Apr. 28, 2023), https://www.nytimes.com/2023/04/28/podcasts/the-daily/ai-deepfake-drake.html?showTranscript=1. ↩︎
  12. Joe Coscarelli, Ghostwriter Returns with an A.I. Travis Scott Song, and Industry Allies, N.Y. Times (Sept. 5, 2023), https://www.nytimes.com/2023/09/05/arts/music/ghostwriter-whiplash-travis-scott-21-savage.html?searchResultPosition=1. ↩︎
  13. Chloe Veltman, When You Realize Your Favorite New Song Was Written and Performed by … AI, NPR (Apr. 21, 2023, 5:00 AM), https://www.npr.org/2023/04/21/1171032649/ai-music-heart-on-my-sleeve-drake-the-weeknd. ↩︎
  14. U.S. Const. art. I, § 8, cl. 8. ↩︎
  15. See A Brief History of Copyright in the United States, U.S. Copyright Off., https://www.copyright.gov/timeline/ (last visited Sept. 14, 2023). ↩︎
  16. Copyright Act of 1790 § 1, 1 Stat. 124 (1790) (current version at 17 U.S.C. § 302). See also The 18th Century, U.S. Copyright Off., https://www.copyright.gov/timeline/timeline_18th_century.html (last visited Sept. 14, 2023). ↩︎
  17. Sonny Bono Copyright Term Extension Act, Pub. L. No. 105-298, 112 Stat. 2827 (1998) (amending 17 U.S.C. §§ 302–304). For a succinct overview of the Sonny Bono Act and its implications, see Kyla Krohn, Pooh and the Public Domain: Childhood Characters Turned Killers, U. St. Thomas L.J.: Blog (May 9, 2023), https://ustlawjournal.com/2023/05/09/pooh-and-the-public-domain-childhood-characters-turned-killers/. ↩︎
  18. 17 U.S.C. § 102(a). ↩︎
  19. Id. §§ 102(a)(2), (a)(7). ↩︎
  20. Id. § 106(1)–(3). ↩︎
  21. U.S. Copyright Off., Circular 14, Copyright in Derivative Works and Compilations (2020), https://www.copyright.gov/circs/circ14.pdf. ↩︎
  22. Id. ↩︎
  23. Harv. U. Off. Gen. Couns., Copyright and Fair Use: A Guide for the Harvard Community 8 (2023), https://ogc.harvard.edu/sites/hwpi.harvard.edu/files/ogc/files/ogc_copyright_and_fair_use_guide_bea_july_2023.pdf?m=1689173591. ↩︎
  24. Id. ↩︎
  25. 17 U.S.C. § 107. ↩︎
  26. Harv. U. Off. Gen. Couns., supra note 23, at 9. ↩︎
  27. Harv. U. Off. Gen. Couns., supra note 23, at 9–10; Richard Stim, Fair Use: What Is Transformative?, Nolo, https://www.nolo.com/legal-encyclopedia/fair-use-what-transformative.html (last visited Sept. 15, 2023). ↩︎
  28. Harv. U. Off. Gen. Couns., supra note 23, at 10. ↩︎
  29. Id. at 10–11. “[T]wo works are substantially similar where the ordinary observer, unless he set out to detect the disparities, would be disposed to overlook them, and regard [the] aesthetic appeal . . . as the same.” Castle Rock Ent., Inc. v. Carol Publ’g Grp., Inc., 150 F.3d 132, 139 (2d Cir. 1998) (internal quotations omitted). Substantial similarity also requires that “an average lay observer . . . recognize the alleged copy as having been appropriated from the copyrighted work.” Ideal Toy Corp. v. Fab-Lu Ltd. (Inc.), 360 F.2d 1021, 1022 (2d Cir. 1966). ↩︎
  30. Harv. U. Off. Gen. Couns., supra note 23, at 11–12. ↩︎
  31. Ariel Soiffer & Aric Jain, Copyright Fair Use Regulatory Approaches in AI Content Generation, Tech Pol’y Press (Aug. 8, 2023), https://techpolicy.press/copyright-fair-use-regulatory-approaches-in-ai-content-generation/. ↩︎
  32. Testimony Before the U.S. S. Comm. on the Judiciary Subcomm. on Intell. Prop., 118th Cong. (2023) (statement of Matthew Sag, Professor of Law in Artificial Intelligence, Machine Learning, and Data Science, Emory University School of Law). ↩︎
  33. Rachel Reed, AI Created a Song Mimicking the Work of Drake and The Weeknd. What Does that Mean for Copyright Law?, Harv. L. Today (May 2, 2023), https://hls.harvard.edu/today/ai-created-a-song-mimicking-the-work-of-drake-and-the-weeknd-what-does-that-mean-for-copyright-law/#:~:text=On%20March%2015%2C%20the%20Copyright,human%20involvement%20cannot%20be%20copyrighted. ↩︎
  34. Testimony Before the U.S. S. Comm. on the Judiciary Subcomm. on Intell. Prop., supra note 32. ↩︎
  35. E.g., Bernard Marr, Is Generative AI Stealing from Artists?, Forbes (Aug. 8, 2023, 2:36 AM), https://www.forbes.com/sites/bernardmarr/2023/08/08/is-generative-ai-stealing-from-artists/?sh=68b970235d1e. ↩︎
  36. See, e.g., What Is Generative AI?, Nvidia, https://www.nvidia.com/en-us/glossary/data-science/generative-ai/#:~:text=How%20Does%20Generative%20AI%20Work,generate%20new%20and%20original%20content (last visited Sept. 15, 2023). ↩︎
  37. Reed, supra note 34. ↩︎
  38. Id. ↩︎
  39. See Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 56 (1884); Urantia Found. v. Maaherra, 114 F.3d 955, 957–59 (9th Cir. 1997); Naruto v. Slater, 888 F.3d 418, 420, 426 (9th Cir. 2018); Kelley v. Chi. Park Dist., 635 F.3d 290, 304 (7th Cir. 2011). ↩︎
  40. 88 Fed. Reg. 16190 (Mar. 16, 2023), https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence#footnote-8-p16191. ↩︎
  41. Id. ↩︎
  42. Id. ↩︎
  43. ghostwriter977 (@imghostwrit3r), Twitter (Sept. 5, 2023, 2:32 PM), https://twitter.com/imghostwrit3r/status/1699143499159855241. ↩︎

Posted

in

by

USTLawJournal

Tags:

Comments

Leave a comment