Harmonizing AI and Copyright: Fair Use in the Age of Generative Music

MBHB summer technical advisor intern Leyna Fougere co-authored this article.

In previous Snippets articles and AI News Roundups, we introduced the current lawsuits between the Recording Industry Association of America (RIAA) and two of the largest generative AI music companies, Suno and Udio. To recap, RIAA is suing on behalf of major music labels Universal Music Group (UMG), Warner Music Group, and Sony Music Entertainment (“the major labels”).[1] The major labels allege that Suno and Udio illegally used copyrighted material to train their AI models, which can output polished AI-generated music based on text prompts and audio “style references.” As the lawsuit continues, however, the plaintiffs have begun quietly negotiating licensing agreements with Suno and Udio.[2] The major labels are seeking damages for past usage as well as future licensing fees and potential ownership stakes in the AI music companies.[3] This tactical shift toward licensing and ownership reflects a rapidly evolving legal landscape around AI-generated content and the copyrighted works used to train the AI models.

The Current Legal Landscape of AI and Copyright Law

Recent rulings in AI-related copyright cases have delivered key wins for AI companies while also opening new opportunities for plaintiffs — making litigation an increasingly risky endeavor. AI companies have persistently argued that their use of copyrighted content, including music and lyrics, does not represent infringement under the fair use doctrine.[4] AI companies argue that their use is transformative, the first of four factors considered in a fair use analysis. See Campbell v Acuff Rose Music and Andy Warhol v. Goldsmith.[5] A recent fair-use ruling on a case between generative AI company Anthropic and several authors found that Anthropic’s use of the copyrighted works it lawfully purchased to train large language models (LLMs) is “spectacularly” transformative.[6] However, Anthropic also used pirated copies of works to train its LLM, and claims related to the pirated copies will proceed to trial to determine potential damages.

Plaintiffs may take lessons from Anthropic to tactically weaken AI companies’ defense under fair use. For example, another important fair use factor is “the effect of the use upon the potential market for or value of the copyrighted work.” 17 U.S.C. § 107. This factor focuses on whether the disputed use negatively impacts the demand for or value of the copyrighted works. In the context of music, AI-generated songs and lyrics would likely occupy the same market as human-made music, which may tilt the market effect factor in favor of plaintiffs.

A decision in the Suno and Udio cases will likely hinge upon the judge’s perception of those market effects. The Suno and Udio cases have similar facts to Anthropic, with some key differences. Suno has admitted to using “essentially all music files…that are accessible on the open Internet,” which may align with the piracy in Anthropic. Additionally, Suno and Udio argue the training of their models is fair use, emphasizing the transformative nature of the use.[8] The nuances of music copyright law, however, will make it more difficult for the plaintiff music labels to argue that Suno and Udio have infringed. A musical recording, unlike a book, generally includes at least two copyrightable elements. Those elements include the audio recording, which labels generally own, as well as the composition of the music itself, which musicians often own.[7] The Copyright Act makes it clear that a new recording of a song is not infringing on the original recording, even if the “sounds imitate or simulate those in the copyrighted sound recording.”[8] Therefore, even if Suno and Udio were reproducing the label’s songs note-for-note — which the major labels admit they are not — the labels’ position may be weak with regard to infringement of the recordings. The deciding factor may then be whether Suno and Udio’s AI-generated music will harm the major recording label’s market. The labels argue that AI-generated audio content undercuts licensing markets that the copyright owners have or could reasonably develop.

Recent court decisions indicate that plaintiffs must provide clear, concrete evidence showing that a licensing market existed in the first place. Without that showing, arguments describing a hypothetical market are unlikely to succeed. In a recent summary judgment decision in a case involving Meta (regarding its LLM LLaMa) and several best-selling authors, Judge Vince Chhabria rebuked the authors for not doing enough to show that their markets would be impacted, calling their arguments “clear losers.”[9] In another case between Anthropic and UMG, the judge rejected a market harm argument, citing UMG’s failure to establish the company had suffered market harm, especially because the “contours of a licensing market for AI training [are still] undefined.”[10]

Crafting a “Winning Argument”

In light of these recent legal developments, how can record labels and AI companies best position themselves to succeed in a copyright infringement case? Judge Chhabria has offered a “potentially winning argument.”[11] Plaintiffs should argue that market harm stems from AI-generated content diluting the market with similar works, shrinking demand for their own, rather than focusing on lost licensing or exact reproductions. Presenting substantive evidence of market dilution or substitution may be crucial in this approach. Additionally, copyright holders seeking to argue lost licensing opportunities should demonstrate prior attempts to license their copyrighted content to AI companies and other third parties. For example, if a company such as UMG has previously licensed their copyrighted music to generative AI companies, they can leverage those licensing agreements as evidence of an existing and active licensing market. This evidence helps quantify claims of lost revenue and may strengthen a rebuttal of a fair use defense.

As demonstrated by the recent rulings from Judge William Alsup and Judge Chhabria, licensing agreements can smooth over many issues for generative AI companies. If a company has come by the copyright material legally — either by license or legitimate purchase, they can rest assured; based on the Anthropic and Meta cases, their actions will most likely be deemed transformative use.

Early licensing negotiations may also be a way for copyright holders, like major music labels, to avoid the need for lawsuits altogether. Music licensing is already a thriving market, and licensing copyrighted content to AI companies is rapidly gaining popularity. For example, The New York Times recently licensed all of its news content to Amazon, allowing Amazon to legally train its LLMs on high quality data.[12]

Based on these recent examples, advice for generative AI music companies, major music labels, and individual musicians seems to take a very similar shape. Namely, whether you’ve recorded it, composed it, or if you are using it, try to license it! The aforementioned court decisions will likely spur AI companies to obtain more licensing agreements to ensure they are operating within the bounds of fair use, and in turn help copyright holders to proactively guard against future fair use defenses. Licensing agreements will also help individual musicians retain control over whether and how LLMs might be trained on their copyrighted works. If licensing is the standard, musicians could choose not to license their content and also know they have some recourse if their work is pirated. Anthropic cited difficulties in obtaining licensing agreements as a major reason they pirated content.[7] If licensing channels are more mainstream and easier to access, it may be a win-win situation for generative AI companies as well as for copyright holders.

[1] https://time.com/6991466/record-labels-sue-ai-music-generator-startups/

[2] https://www.wsj.com/business/media/ai-music-licensing-universal-warner-sony-92bcbc0d?reflink=desktopwebshare_permalink

[3] https://www.law360.com/articles/2351696?nl_pk=c1a50e9e-88c6-415f-a417-719fbb30196d&utm

[4] https://supreme.justia.com/cases/federal/us/510/569/

[5] https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_4.pdf

[6] https://www.courtlistener.com/docket/68878608/28/umg-recordings-inc-v-suno-inc/

[7] 1 Nimmer on Copyright § 2.05 (2025)

[8] 17 U.S.C. § 114(b) (Copyright Act)

[9] https://assets.law360news.com/2357000/2357703/https-ecf-cand-uscourts-gov-doc1-035125897019.pdf

[10] https://www.reuters.com/legal/anthropic-wins-early-round-music-publishers-ai-copyright-case-2025-03-26/

[11] https://www.law360.com/articles/2351696?nl_pk=c1a50e9e-88c6-415f-a417-719fbb30196d&utm