- The RIAA has sued AI music platforms Suno and Udio alleging that they trained their models on copyrighted content without permission
- It amounts to theft “on an almost unimaginable scale”, says the trade body
- The platforms can now output content at such a “speed and scale” that the impact on human creativity will be “rapid and devastating” without intervention
- Both companies have claimed ‘fair use’, but the RIAA insists that the US copyright rule does not apply when training AI models
The RIAA is suing AI music platforms Suno and Udio, alleging that both have trained their generative AI models with existing music and, in doing so, have infringed an awful lot of copyright.
The litigation has been coordinated by the Recording Industry Association Of America on behalf of its members, which include the three major record companies. RIAA CEO Mitch Glazier says that the music community has already “embraced AI” and forged partnerships with various “responsible developers”, adding “there is both promise and peril in AI”.
“Devastating impacts” on human creativity
Suno and Udio, say the lawsuits, have been “deliberately evasive” about how they trained their AI models. The RIAA alleges that both services have stolen copyright protected recordings “on an almost unimaginable scale” to develop technologies that generate “synthetic musical outputs”.
As a result, the outputs of the two platforms will have “rapid and devastating impacts”, because the “speed and scale” at which music can be generated “could saturate the market with machine-generated sound-a-likes and knockoffs”, which will “directly compete with, cheapen, and ultimately drown out the genuine sound recordings on which they were built”.
“There is nothing that exempts AI technology from copyright law or that excuses AI companies from playing by the rules”, the lawsuits adds. “AI companies, like all other enterprises, must abide by the laws that protect human creativity and ingenuity”.
The first big test of music industry’s instance that ‘fair use’ does not apply to AI training
These are the first big record industry lawsuits filed against AI companies that are generating music. The other big music case - against Anthropic - was filed by a group of music publishers and relates to the alleged infringement of the copyright in lyrics.
In the early days of digital music, the RIAA coordinated many of the key lawsuits that were filed against file-sharing platforms, starting with the Napster litigation in 2000.
Although the record industry was usually successful in court during the file-sharing days, the lawsuits did little to stop online music piracy in the short term. However, the cases did clarify the legal obligations of the digital platforms and the RIAA might argue that that helped provide an environment where legit digital music services could prosper.
It is certainly seeking to position these lawsuits as having that objective, arguing that going after the unlicensed AI platforms will help the AI companies that are collaborating with the industry. The RIAA states that “unlicensed services like Suno and Udio that claim it’s ‘fair’ to copy an artist’s life’s work and exploit it for their own profit without consent or pay set back the promise of genuinely innovative AI for us all”.
The RIAA “believes strongly in the human-to-human, artist-to-fan connection that comes from lived experience”, it adds. “That is worth fighting for. We embrace artificial intelligence tools that are responsible, pro-artist, and keep humans at the centre of creativity. That is what these cases are about”.
Suno initially released its generative AI product to create music in December 2023, with Udio following in April 2024. Both have come under significant scrutiny from the music industry, with lots of debate about the abilities of the two technologies, but also the training data employed. Both already offer subscription and credit pricing options, meaning they are already monetising the models which, the industry claims, were built on rampant infringement.
What do we know about how Suno and Udio trained their models?
When it comes to training data, both Suno and Udio have been vague about what existing music has been used. That isn’t necessarily surprising; a lack of transparency around training data is a common complaint about generative AI companies. However, some other businesses developing music AI, such as Meta and Stability, have licensed production music to train their generative AI models, and have clearly stated that fact.
Despite the lack of clarity regarding training data, it’s widely believed that large quantities of commercially released music was in both Suno and Udio’s datasets. Some users of the services have noted that they have generated music with striking similarities to well-known sound recordings and vocals that are pretty much indistinguishable from famous recording artists.
The RIAA says that users have been able to create “sound-a-likes of numerous sound recordings using Suno and Udio, including The Temptations’ ‘My Girl’, Green Day’s ‘American Idiot’, Mariah Carey’s ‘All I Want For Christmas’, along with recordings by Chuck Berry, James Brown, and others”.
It also claims that “they left producer tags on their outputs”, with tags from CashMoneyAP and Jason Derulo “present in Suno’s outputs”. These tags could not be there if the model hadn't been trained on copyright protected material.
The trade group also notes that Udio executives have talked about their model being trained on “a large amount of publicly-available and high quality music” that is “obtained from the internet”, which basically sounds like they scraped all the pop music available online. One of Suno’s co-founders, meanwhile, said his company used a “mix of proprietary and public data”.
Suno investor Antonio Rodriguez of Matrix Partners - which also invested in The Echo Nest - even seemed to accept that a legal battle with the music industry was inevitable,
Speaking in an interview with Rolling Stone earlier this year, he said that being sued by the music industry was “the risk we had to underwrite when we invested in the company”, adding that - if Suno had already done deals with labels at the point he invested - “Honestly… I probably wouldn’t have invested in it”. That’s because he feels the start-up had to begin developing its AI without the labels so that it could make its product without “constraints”.
Assuming it can be easily proven that both Suno and Udio trained their models with existing music owned by RIAA members, both these lawsuits may then fairly quickly turn into a debate about fair use and AI. Many technology companies argue that training AI constitutes fair use under US copyright law, meaning they can use existing content without getting permission.
The RIAA’s argument
However, the RIAA’s filing makes it clear that it believes any claims of fair use by Suno and Udio are invalid, pointing out that, under US law, fair use requires four factors to be considered. Possibly the most important is that - in the RIAA’s words - “a fair use defence is not available for uses that could commercially compete with or act as substitutes in the marketplace for the original works that were copied - surely the case here”.
The RIAA reveals that both AI companies have already raised fair use defences in pre-litigation correspondence, and says that this fact is “tantamount to an admission” that the record companies’ copyrights have been infringed.
Fair use “only comes up as a defence to unlicensed copying”, the RIAA says. “If Suno and Udio did not copy recorded music, why would they need to invoke this defence?”
It then adds, “In a bizarre and unbelievable ‘I didn’t do it, but if I did, it was legal’-type contention, both companies refused to come clean about their copying of sound recording while simultaneously claiming that if they did copy, it was fair use”.
As with all the other copyright lawsuits filed against AI companies, only a court of law can declare whether or not AI training is or is not fair use. If the record companies prevail, however, the damages could be massive.
The RIAA says it is seeking statutory damages of up to $150,000 for every infringed work. Without any transparency on Suno and Udio’s dataset it will be hard to identify the specific works that have been infringed. But assuming it can be demonstrated they used a huge number of tracks, the damages could be astronomical. For example, if it could be demonstrated that training had involved 70,000 recordings, damages could exceed $10 billion.
Similar claims in cases involving the infringement of thousands of tracks have resulted in mega-damages in the past, including the billion dollar judgement against US internet service provider Cox Communications. In that judgement, the damages were calculated at $99,830.29 for each of the 10,017 illegally downloaded and shared music files identified by the labels.
During the file-sharing era, the record industry’s litigation - and subsequent damages - ultimately forced a number of companies out of business, including Napster and Limewire.
Aware of the potential mega-damages, many AI companies - while remaining bullish that the fair use defence applies - are also trying to negotiate licensing deals with rights owners to future proof themselves should the first landmark ruling in this domain favour the copyright industries.
This means that both Suno and Udio might seek to agree licensing deals with the record industry while the litigation goes through the motions. In the early days of digital music there were some services that launched without licences and then negotiated deals with the industry, with forgiveness for past infringement factored in.
Though with Suno’s investor implying a deliberate strategy of “infringe now and ask for forgiveness later if we really have to”, some in the industry might want to respond with a “sue them out of business whatever they offer” approach.