Udio and Suno are not trendy new eateries on the Lower East Side, but they are innovative AI startups that allow you to create lifelike songs. These platforms use AI to generate music from simple prompts, including instrumental and vocal performances. However, on Monday, major record labels took legal action against them, accusing them of copyright infringement on a massive scale. They claim these companies could only achieve their results by illegally using copyrighted music to train their AI models.
Growing legal challenges for AI
These lawsuits add to the mounting legal issues faced by the AI industry. Many successful AI firms have trained their models using vast amounts of data scraped from the internet without permission. For instance, ChatGPT was initially trained on millions of documents sourced from Reddit links.
The Recording Industry Association of America (RIAA) is leading these lawsuits, focusing on music rather than text. Similar to The New York Times' lawsuit against OpenAI, these cases raise a crucial question that could change the tech industry: can AI companies use any data they want, turn it into a valuable product, and claim it as fair use?
“That's the key issue that needs resolution, as it affects various industries,” said Paul Fakler, an intellectual property expert at Mayer Brown.
Understanding Udio and Suno
Udio and Suno are relatively new but have already made significant impacts. Suno, launched in December by a Cambridge-based team previously associated with Kensho, quickly partnered with Microsoft to integrate its technology with Microsoft's AI chatbot, Copilot.
Udio, launched this year, has raised millions from notable investors like Andreessen Horowitz and musicians such as Will.i.am and Common. Comedian King Willonius used Udio's platform to create “BBL Drizzy,” a Drake diss track that went viral after producer Metro Boomin remixed it and made it available for public use.
The RIAA's lawsuits use grand language, stating the litigation is about “ensuring that copyright continues to encourage human creativity and innovation, as it has for centuries.” However, the underlying issue is financial.
The RIAA argues that generative AI threatens record labels' business models. They claim potential licensees might choose to generate AI soundalikes at virtually no cost instead of licensing copyrighted sound recordings. This could flood the market with imitation tracks, disrupting the established sample licensing business.
The RIAA seeks damages of US$150,000 per infringing work, which could be huge given the extensive data typically used to train AI systems.
The RIAA's lawsuits provide examples of music generated by Suno and Udio and compare their musical notation to existing copyrighted works. Some generated songs include phrases similar to actual songs, such as a track beginning with “Jason Derulo” in the same style as the singer himself. Others mimic longer sequences from famous songs, like Green Day's “American Idiot.”
While this seems incriminating, the RIAA claims the infringement comes from the AI companies using copyrighted music as training data, not from the generated songs.
Neither Suno nor Udio have disclosed their training datasets, a practice common in the AI industry. For example, OpenAI has avoided questions about using YouTube videos to train its models.
The RIAA highlights statements from Udio CEO David Ding about training on publicly available high-quality music and comments from a Suno co-founder about using a mix of proprietary and public data.
Paul Fakler describes the lawsuit's inclusion of examples and notation comparisons as excessive, suggesting it's more for public relations than legal necessity. He notes that creating a soundalike is legal if you have rights to the original song.
When asked for a comment, Suno's spokesperson provided a statement from CEO Mikey Shulman, asserting that their technology is “transformative” and that they do not allow prompts naming existing artists. Udio did not respond to requests for comment.
Is it fair use?
A significant question remains: Is it fair use even if Udio and Suno used copyrighted music for training?
Fair use allows the use of copyrighted material to create something new and transformative. The RIAA argues that the startups cannot claim fair use, stating that their outputs replace actual recordings, are for commercial purposes, involve extensive copying, and threaten the record labels' business.
Fakler believes the startups have a strong fair use argument if they only temporarily copied the copyrighted works to extract their essential features into the AI model's data.
“How do we create songs that sound like music to listeners? It's about extracting these features, just as musicians learn by playing music,” he said.
However, the final decision lies with the courts. Details uncovered during the discovery process could significantly impact the case. As the RIAA's lawsuits progress, the fair use question looms over the AI industry, so we are likely in for a long journey.
