Copyright Law Fair Use: Insights from the Suno Case

Written by / Chen Jing, Shanghai Zhihe Law Firm

Recently, there has been new progress in the copyright lawsuit between the Recording Industry Association of America (RIAA) and AI music companies Suno and Udio: previously, the RIAA, representing the three major record companies including Sony, Universal, and Warner, collectively filed a lawsuit against the music generation services Suno and Udio, demanding $150,000 in damages for each infringing work. According to media reports on August 1, 2024, Suno admitted in its court documents that it indeed captured copyrighted recordings owned by the plaintiffs among the tens of millions of recordings used for training data.

This is similar to the AI voice case in our country, where the defendants also admitted to using the plaintiff’s voice/music. In the first AI voice case in our country: the plaintiff discovered that a text-to-speech product on a platform used the plaintiff’s voice. The plaintiff had previously been commissioned by defendant two to record audio products, and defendant two owned the copyright of the plaintiff’s recordings, while defendant three obtained the corresponding authorization from defendant two to use the plaintiff’s voice to generate the text-to-speech product in question. Defendants one and five signed an online service sales contract, and defendant five purchased the product from defendant three. They mutually proved that the AI-generated voice was indeed based on the plaintiff’s voice. (For my legal analysis of this case, see “Hearing the Voice First: Has AI Voice Infringed Copyright?”).

In fact, if Suno does not admit, it would be technically very difficult for the three major record companies to prove that Suno used their music works for AI data training. So why did Suno admit it? One of Suno’s main defenses in this case is “fair use,” arguing that it is to produce new music works, which does not constitute infringement. What is “fair use”? Can Suno defend itself through “fair use”?

The initiators of this lawsuit are the three major record companies represented by Universal Music, Warner, and Sony, which integrate music production, distribution, and copyright management, holding a leading position in copyright resources in the industry; this lawsuit has also received support from many music associations and industry companies. On the other hand, the RIAA provided a large amount of evidence in the lawsuit materials, such as comparisons of scores with existing copyrighted works. The method of evidence collection is fundamentally similar to the analysis article on Suno by Ed Newton-Rex, the former audio vice president of Stability AI. (See “In-Depth Exposure: Is Suno’s Music AI Training Data Infringing?”). According to Ed Newton-Rex’s test results on Suno:

1. In terms of style: If well-known musicians or track names are used as prompts, Suno will refuse to generate the corresponding track but can slightly misspell the artist’s name. Using this method, Suno seems to be able to generate a voice that sounds like Eminem (inputting about “beminem”).

2. In terms of melody: Suno’s assisted songwriting mode can obtain melodies suitable for existing famous song lyrics through inputting those lyrics. This method can relatively easily find examples where Suno replicates existing melodies. For example, prompts like “Style: 70s British rock; Song name: Bohemian Symphony; Lyrics: All lyrics of Bohemian Rhapsody”. Here, Suno almost verbatim copied the unique “Galileo” lines from Bohemian Rhapsody, using similar pitch melodies, similar styles, and similar vocal ranges.

From the perspective of the many copyright rights enjoyed by the plaintiffs and the degree of similarity of the songs, it seems hard to deny that Suno used copyrighted music. Ed Newton-Rex believes that if the model uses copyrighted works for training without permission, it infringes copyright, regardless of whether the model’s generated results completely replicate the training data. However, Suno’s main defense point is the “fair use” in copyright law, arguing that its use of copyrighted music is for model training, which is a learning process aimed at producing new music works, rather than mere “parroting”.

Fair use is an important rule in U.S. copyright law that allows the use of copyrighted materials without obtaining permission from the copyright holder under specific conditions. According to Section 107 of the U.S. Copyright Act, determining fair use requires considering several factors: the purpose and nature of the use, the nature of the copyrighted work, the amount and substantiality of the portion used in relation to the copyrighted work, and the effect of the use on the potential market or value of the copyrighted work.

Suno argues that deep learning models are similar to the human learning process, which can help people create music by learning styles, patterns, and forms, and then creating new music around them. They argue that model training resembles a child learning to create new rock songs by listening to rock music. Just like a child, the more AI learns, the better Suno will become. Moreover, these musical styles, whether opera, jazz, or hip-hop, cannot be owned by anyone.

In fact, “borrowing” is essential for creation; if the use of others’ works is completely prohibited, it will hinder the innovation of literature and art. For example, in the field of literature, there is borrowing; the Soviet literary theorist Bakhtin proposed the theory of polyphony: “The writing of any text is like a collage of quotes, and any text absorbs and transforms other texts.” The same is true in music; for instance, rock music may draw from classical music, and it can even be said that rock music is a synthesized form of music, incorporating rhythm and blues, gospel, soul, etc. Suno believes its deep learning model is merely “analyzing” and “learning” the components of music.

For Suno, there are some disadvantages in determining whether its use is fair:

• Suno profited from AI-generated music, including raising $125 million in its latest funding round and charging users a monthly subscription fee of $24;

• Suno’s music may also impact the plaintiffs’ market share, as user-generated music has entered mainstream media services and competes with copyrighted works, especially since this case is brought by the RIAA representing the three major music copyright giants.

In U.S. law, fair use may serve as an affirmative defense, and the defendant bears a certain burden of proof. In the case of “Campbell v. Acuff-Rose Music, Inc,” the Second Circuit Court of Appeals held that the defendant’s use was commercial in nature, thus presuming its actions to be unreasonable (unfairness) and causing market harm. This presumption shifts the burden of proof to the defendant, who must refute the “unreasonableness” of their use.

Article 24 of China’s Copyright Law also stipulates “fair use,” requiring that fair use must meet the “three-step test,” which states that it can only be made under special circumstances (the article lists 12 scenarios, including personal study, commentary, and news reporting where use is unavoidable), must not affect the normal use of the work, and must not unreasonably harm the legitimate rights and interests of the copyright holder.

In China’s first AI painting model infringement case (the “AI Painting Model Infringement Case”), the defendant model manufacturer also claimed a fair use defense. In this case, the plaintiffs were illustrators and registered users of a content-sharing platform, who had long published their created paintings on that platform. They discovered that users had posted images similar to the plaintiffs’ paintings on that platform, claiming these images were created using certain AI painting software, which the content-sharing platform had launched.

In the AI painting model infringement case, the plaintiffs not only sued the model manufacturer but also the content-sharing platform. The defendants’ defense had two points: first, the plaintiffs claimed that the works they asserted rights over were not substantially similar to the AI-generated images in question; second, even if the defendant’s model training behavior used the plaintiffs’ works, it should still constitute fair use.

If viewed solely from the perspective of model training, although model training lacks public nature, AI can produce substitutes for works through learning (both the Suno case and the AI painting model infringement case generated works that have a certain degree of similarity to the originals), the market for the works would be substantially impacted, which does not align with fair use scenarios. At this point, it can be considered whether the model manufacturer obtained the works through “legal acquisition” (referring to the initial views, see “On the Position of ‘Legal Acquisition’ in Copyright Legislation”), which can assist in determining fair use. Just like in human creative scenarios, individuals have dual identities: the first identity is that of a consumer, and the second identity is that of a creator; individuals obtain works through purchase or download for learning and then create. The same applies to models. If the model obtains works at the input end through “legal acquisition,” then its use in a private space is reasonable.

For example, based on the information disclosed in the AI painting model infringement case, on one hand, the plaintiffs publicly shared their works on multiple content-sharing platforms and provided original image downloads; on the other hand, according to the user agreement of a certain content-sharing platform, users authorized the platform to use the works uploaded. In this scenario, whether the model manufacturer obtained the plaintiffs’ works through the content-sharing platform’s permission constitutes “legal acquisition”?

Legal litigation is one way, but in the era of AI development, there may be another path for creators and AI companies (including model manufacturers and AI content-sharing platforms) to explore: that is, cooperation for mutual benefit. They can work together with major copyright holders to establish recognition of “legal acquisition.”

Recently, Universal Music Group announced the renewal of relevant agreements with Meta Platforms to create new opportunities for its artists and songwriters on Meta’s social platforms, jointly addressing the impact of unauthorized AI-generated content on artists and songwriters. Similarly, after encountering several copyright infringement lawsuits, OpenAI has been promoting cooperation with news publishing organizations, such as Condé Nast, the Associated Press, Axel Springer, and The Atlantic, all of which have joined the copyright cooperation with OpenAI.

For creators, AI can also be a useful tool. Currently, AI tools cannot yet create outstanding works, but artists can use AI tools for creation; for model manufacturers, content-sharing platforms, etc., it is also unrealistic to require them to obtain all major and minor copyright rights. However, AI companies can collaborate with copyright repositories/databases to fulfill their duty of care, which can impact the determination of fair use and reduce liability for damages. For instance, recently the Supreme People’s Court made a retrial judgment in the case of Canxing Culture v. Changxiang Youth copyright infringement dispute: Changxiang Youth, as a karaoke operator that has paid fees to the China Audio-Video Copyright Association, was ordered to compensate 8,000 yuan for economic losses related to 212 works (including reasonable rights protection expenses), resulting in an average compensation amount of less than 30 yuan per work after deducting reasonable expenses.