The tension between the music industry and the rapidly growing artificial intelligence music generation industry has flared up once again. On June 24, 2024, the Recording Industry Association of America (RIAA) announced that it had filed two copyright infringement lawsuits in the United States on behalf of three major record labels against the operators of Suno AI and Udio AI, two popular AI music generation platforms. The lawsuits add many interesting questions to the already complex debate taking place in both the US and Canada regarding the intersection between copyright law and AI.
The Copyright Infringement Lawsuits
RIAA brought the lawsuits on behalf of the world’s three largest recording companies, Universal Music Group, Warner Music Group, and Sony Music Entertainment. Collectively, these plaintiffs operate record labels that represent thousands of recording artists, including many of the world’s most popular performers.
The defendants, Suno AI and Udio AI, are the operators of music generation platforms that rely on generative AI to create sound recordings. These platforms allow users to provide a short, written prompt describing a type of music and/or a topic for a song (for example, “an electronic pop song about going to the beach”) and will then instantly produce a recording of an AI-generated song that sounds like it is performed by a human. While neither platform accepts written prompts that contain the name of a specific existing recording artist, both platforms allow for highly specific genre and style prompts that can result in a comprehensive description of a specific artist.
The Suno and Udio platforms both rely on generative AI, which is trained on large data sets of pre-existing human-made music. Essentially, the AI models are “fed” with thousands of pre‑existing songs from across a wide range of music genres, styles, melodies, vocal techniques, and lyrical content, which the models use to “learn” standard characteristics of different types of music. The AI models then aggregate the training data to produce new recordings that are reminiscent of a general type of music, or that may sound similar to pre-existing songs, but, at least as claimed by Suno and Udio’s operators, do not directly reproduce any pre-existing recording.
The RIAA-managed lawsuits focus on the input data sets. Specifically, they allege that the data sets likely contain thousands of pre-existing recordings whose copyright is owned by the plaintiffs and that the unauthorized reproduction of those recordings as part of the data sets constitutes copyright infringement. To date, Suno and Udio have insisted that the pre-existing recordings that make up the input data sets were all legally acquired.
Notably, the plaintiffs clarify explicitly in the lawsuits that they are not currently alleging that the output of Suno and Udio—i.e., the “new” AI-generated recordings—are themselves infringements of copyright. While the lawsuit refers to the fact that the AI-generated recordings produced by these platforms can often sound similar to specific pre-existing songs, the plaintiffs allege that these similarities are evidence that the input data sets on which Suno and Udio is trained contain the plaintiffs’ copyrighted works.
In other words, the plaintiffs argue that, even if Suno and Udio will not disclose which specific pre-existing recordings are included in the training data sets, the similarities between the output recordings and the plaintiffs’ pre-existing recordings are enough to prove that the AI models must have been trained on those pre-existing recordings. This line of argument is necessary because Suno and Udio have historically refused to provide details on which recordings are contained in their data sets or how those data sets are populated. They have argued that those details constitute confidential business information.
Among other remedies, the plaintiffs seek statutory damages in the amount of $150,000 USD for each recording infringed. Since the data sets relied on by Suno and Udio may contain thousands of pre-existing recordings, the total damages award being sought by the plaintiffs could, theoretically, rise to billions of dollars if the courts were to find that each individual recording contained in the data sets constitutes a separately infringed recording.
Key Takeaways for Canada
Although a similar infringement action has not yet been brought in Canada, the issues that the RIAA-managed lawsuits raise are likely to be relevant to Canadian copyright law as well. The lawsuits will require the US courts to grapple with numerous unanswered questions about the intersection between artificial intelligence and copyright that directly relate to concepts under Canada’s Copyright Act, including whether using a copyrighted work or recording as part of a training data set constitutes an unauthorized reproduction, when a work or recording is acquired lawfully, whether AI training data sets can fall under the “fair dealing” exception to copyright infringement, and how damages are to be calculated if a data set containing thousands of copyright-protected works or recordings is found to be infringing.
More broadly, the RIAA-managed lawsuits illustrate the current complexities and uncertainties around the world, including in Canada, about the creation and use of AI-generated works for commercial purposes. These unresolved questions extend far beyond just the music industry. For example, as we have discussed in previous Cassels comments, courts in both the United States and Canada have already started ruling on whether copyright can exist in an AI-generated work and whether a company can be held responsible for statements made by its AI chatbot, while the landscape of AI regulation at large in Canada continues to develop across many industries.
The Cassels Intellectual Property Group will continue to monitor these rapid developments in the AI regulatory space, including in the copyright law context.