Despite their names, Udio and Suno aren’t the hottest new restaurants on the Lower East Side. They’re an artificial intelligence startup that lets people follow prompts to generate impressive-sounding real songs, complete with instrumental and vocal performances. A group of major record labels sued them on Monday, alleging copyright infringement on an “almost unimaginable scale” and claiming the companies could only do so because they illegally ingested vast amounts of copyrighted music to train their artificial intelligence models.
The two lawsuits add to a growing list of legal headaches for the artificial intelligence industry. Some of the most successful companies in the field have trained their models using data obtained from unsanctioned scrapes of vast swaths of the internet. ChatGPT, for example, was initially trained on millions of documents collected from links posted on Reddit.
The lawsuits, led by the Recording Industry Association of America (RIAA), target music rather than the written word. but like New York TimesIn their lawsuit against OpenAI, they raise a question that could reshape the tech landscape as we know it: whether artificial intelligence companies can simply take whatever they want and turn it into a multibillion-dollar product , and claim it’s fair use?
“This is a critical issue that must be addressed because it spans a variety of different industries,” said Paul Fakler, a partner at Mayer Brown, a law firm that specializes in intellectual property cases.
What are Udio and Suno?
Udio and Suno are both fairly new, but they’re already making quite a splash. Suno was launched in December by a Cambridge team that previously worked at another AI company, Kensho. It quickly formed a partnership with Microsoft to integrate Suno with Microsoft’s artificial intelligence chatbot Copilot.
Udio launched just this year and has raised millions of dollars from heavyweights in the tech investing world (Andreessen Horowitz) and the music world (like Will.i.am and Common). Comedian King Willonius used Udio’s platform to produce “BBL Drizzy,” a Drake diss track that went viral after producer Metro Boomin remixed it and released it to the public for anyone to rap on.
Why is the music industry suing Udio and Suno?
The RIAA’s lawsuit uses lofty language, saying the lawsuit is about “ensuring that copyright continues to inspire human invention and imagination, as it has for centuries.” That sounds great, but ultimately, the incentive it says is money .
The RIAA claims that generative AI poses a risk to record labels’ business models. “Instead of licensing copyrighted recordings, potential licensees interested in licensing such recordings for their own purposes can generate AI-like voices for virtually free,” the lawsuit states, adding that such services can “[flood] The market is flooded with ‘copycats’ and ‘sound-alikes’, disrupting the established sample licensing business.
The RIAA is also seeking $150,000 in damages for each infringed work, which could be an astronomical sum given the vast amounts of data typically used to train artificial intelligence systems.
Does it matter whether songs generated by artificial intelligence are similar to real songs?
The RIAA’s lawsuit includes examples of music generated using Suno and Udio and a comparison of their scores to existing copyrighted works. In some cases, the generated songs have similar little phrases – for example, one song begins with “Jason Derulo” sung at the exact same tempo with which real-life Jason Derulo begins many of his songs. Others expanded on sequences of similar symbols, such as a track inspired by Green Day’s “American Idiot.”
One of the songs begins with the song “Jason Derulo” with the exact same tempo that real life Jason Derulo starts many of his songs with
this seem Pretty damning, but the RIAA is not claiming that these specific sound-alike tracks infringe copyright, but rather that the artificial intelligence company is using copyrighted music as part of its training materials.
Neither Suno nor Udio have made their training data sets public. Both companies are coy about the origins of their training materials – although this is common in the artificial intelligence industry. (OpenAI, for example, sidestepped the question of whether YouTube videos were used to train its Sora video model.)
The RIAA lawsuit notes that Udio CEO David Ding said the company conducts training with the “best quality” music that is “publicly available,” and Suno co-founders wrote in Suno’s official Discord that the company conducts training with “proprietary and mixed music.” way of training”. Public data. “
Fackler said it was “outlandish” to include examples and symbolic comparisons in the lawsuit, saying it went “well beyond” what was needed to make a legal case for the lawsuit. For one, the record company may not own the composition rights to the songs that Udio and Suno allegedly ingested for training. Instead, they own the copyright to the sound recordings, so showing similarity in musical scores doesn’t necessarily help resolve copyright disputes. “I think it was actually designed for PR purposes,” Fackler said.
On top of that, Fackler points out that it’s legal to create a sound-alike recording if you own the copyright to the underlying song.
When reached for comment, a Suno spokesperson shared a statement from CEO Mikey Shulman calling its technology “transformative” and that the company does not allow prompts mentioning the names of existing artists. Udio did not respond to a request for comment.
Fair use?
But even if Udio and Suno used record labels’ copyrighted works to train their models, there’s one very big question that overrides everything else: Is this fair use?
Fair use is a legal defense that allows the use of copyrighted material in the creation of a meaningful new or transformative work. The RIAA argued that startups could not claim fair use, saying that Udio and Suno’s outputs were intended to replace actual recordings, that they were generated for commercial purposes, that the copying was extensive rather than selective, and, finally, that the resulting The resulting product poses a direct threat to the label business.
In Fakler’s view, as long as the copyrighted work is only temporarily copied and its defining features are extracted and abstracted into weights for an artificial intelligence model, the startup has a strong fair use argument.
“It extracts all these things, just like musicians learn these things by playing music.”
“This is how the computer works – it has to make these copies, and then the computer analyzes all this data in order to extract the content that is not protected by copyright,” he said. “How do we create songs that listeners can understand as music and have all kinds of characteristics that are common in pop music? It extracts all of those things, just like musicians learn these things by playing music.
“In my opinion, that’s a very strong fair use argument,” Fackler said.
Of course, a judge or jury may disagree. And what is unearthed during the discovery process — if these lawsuits come to fruition — could have a significant impact on the case. Which musical tracks were filmed and how they ended up in the training set may matter, and specific details about the training regime may weaken a fair use defense.
We face a long journey as the RIAA lawsuit and others like it work their way through the courts. From words and photos to now audio recordings, the question of fair use hangs over all of these cases and the AI industry as a whole.
