The lawsuits that the major recording companies filed on Monday (June 24) against AI music companies Suno and Udio leave little doubt that the music industry sees these types of AI tools as an existential threat.
The two companies’ “unauthorized use of… copyrighted recordings threatens to eliminate the existing market for licensing sound recordings,” the lawsuits state, “as well as the future market for licensing sound recordings to generative AI companies.”
In other words, these technologies — which allow users to create songs in seconds with nothing more than a text prompt — could bring down the entire music industry.
For the major music rightsholders behind the suits, failure is not an option.
The two lawsuits—one brought against Suno in a federal court in Massachusetts, the other against Udio in a federal court in New York—both contain almost identical allegations: that Suno and Udio each violated the copyrights held by recording companies by copying and ingesting copyrighted music to train their AI.
You can read the full Udio lawsuit here, and the full Suno lawsuit here.
They were both filed by various recording divisions of the three major music companies – Universal Music Group, Sony Music Group and Warner Music Group – making them the first major cases brought by recording companies against AI. (Notably, Universal’s publishing company is leading a separate, lyrics-based lawsuit against another budding gen-AI giant, Anthropic.)
“This case is precedent-setting and integral to artists’ rights as human creators.”
David Israelite, NMPA
According to Jonathan Coote, a music and AI lawyer at UK-based law firm Bray & Krais, the courts must answer two major questions in this case.
Asks Coote: “First, did the AI tools train on copyrighted recordings? The labels have presented compelling evidence showing the similarity between outputs and original works, including digital watermarks such as Jason Derulo’s infamous vocal trademark. Notably, Suno’s initial response doesn’t dispute that it trained on copyright[ed] works.”
Coote’s second question: Does the AI companies’ use of copyrighted works amount to “fair use”?
Fair use is a legal doctrine that states that, in certain limited circumstances, there are exemptions from copyright laws, usually to serve some greater public good. For instance, students can copy parts of copyrighted books for the sake of their education, and researchers can sometimes avoid copyright law for the sake of their research.
The US version of “fair use”, notes Coote, is “far more flexible than in the UK and include[s] considerations such as whether the use was ‘transformative’”.
He adds: “This could potentially become a philosophical question about the creative role of AI, encompassing its economic and social impact. The cases will likely be one of a number across the creative industries that are eventually decided by a Supreme Court decision.”
So a final decision on whether or not these AI companies illegally used copyrighted materials to train their models is likely still a long way off.
Nonetheless, “it will have an immediate impact, as investors in AI music tools will be even more concerned with ensuring that any training has been conducted legally,” Coote predicts.
Elsewhere, David Israelite, CEO and President of the NMPA, said of the lawsuits, which are being coordinated by the RIAA: “This case is precedent-setting and integral to artists’ rights as human creators. Millions of people already use these tools which amounts to countless infringements on real musicians.”
So how exactly do the major recording companies plan to prevail in this ground-breaking case?
Below, we’ll go over the evidence and three key legal arguments the labels (via the RIAA) are deploying.
But first, let’s look at a tactical element involved in these suits, which is, put simply: intimidation. That’s because of the potential amount that Suno and Udio could be on the hook for if they lose this case…
Headline fact: These lawsuits could cost Suno and Udio billions
In their legal complaints, the recording companies ask the courts for statutory damages of up to $150,000 per infringing work.
That’s the maximum amount allowed under the US Copyright Act, so it’s hardly guaranteed that the courts will award that much even if the recording companies secure a resounding victory, but in the case of AI training, this could amount to an enormous amount of money.
That’s because, as AI developers often point out, training an AI model takes enormous amounts of information. Suno and Udio have been very secretive about what music (and how much) they used to train these algorithms, but it’s safe to say that to generate the millions or billions of data points needed to train their AI models, they would have needed a database of music counted at least in the thousands, and likely much more.
“At the $150,000 statutory maximum, these AI companies would have had to infringe 6,666 songs for the damages to reach $1 billion.”
At the $150,000 statutory maximum, these AI companies would have had to infringe 6,666 songs for the damages to reach $1 billion. It’s entirely likely that – if the courts require Suno and Udio to disclose their training dataset as part of discovery – the number of copyrighted tracks used to train these models will turn out to be much larger than that. (Lest we forget: Udio is reportedly kicking out ten new tracks every second.)
But we’re getting a little ahead of ourselves here: Before a single dime of damages are paid, the major recording companies will have to prove that Suno and Udio did, indeed, use copyrighted music, and then they will have to convince the courts that this amounts to copyright infringement.
Here’s three ways they plan to do exactly that:
1) The recording companies will use Suno and Udio investors’ and executives’ own words against them
One of the more unique elements of these copyright cases is that, in terms of proving that these AI companies used copyrighted works to train their models, the recording companies can point to some powerful circumstantial evidence: things that the companies’ own investors and execs have said.
In Suno’s case, those words came from Antonio Rodriguez of Venture Capital firm Matrix Partners, an early investor in Suno. In an interview with Rolling Stone earlier this year, Rodriguez all but admitted that Suno had used copyrighted works in its training.
“Rodriguez… explained that his firm invested in the company with full knowledge that Suno might get sued by copyright owners, which he understood as ‘the risk we had to underwrite when we invested in the company,’” the complaint against Suno states.
“Rodriguez pulled the curtain back further when he added that ‘honestly, if we had deals with labels when this company got started, I probably wouldn’t have invested in it. I think they needed to make this product without the constraints.’
“By ‘constraints,’ Rodriguez was, of course, referring to the need to adhere to ordinary copyright rules and seek permission from rightsholders to copy and use their works.”
In the legal complaint against Udio, the recording companies cite an interview that Udio CEO David Ding gave to Billboard. In it, Ding noted that, although “we can’t reveal the exact source of our data,” Udio’s AI model was trained on “publicly available data that we obtained from the internet.”
In another interview, this time with Fortune, Ding said Suno had been trained on the “best quality music that’s out there.” The recording companies interpret this to mean copyrighted music.
“the only practical way generative AI models can exist is if they can be trained on an almost unimaginably massive amount of content, much of which (because of the ease with which copyright protection can be obtained) will be subject to copyright.”
Andreessen Horowitz (Udio investor) filing with the US copyright office, 2023
The major record companies also point to some less direct circumstantial evidence, in the form of comments made by one of Udio’s investors, the venture capital firm Andreessen Horowitz, aka a16z.
In a submission to the US Copyright Office on the issue of AI and copyright last year – as previously covered by MBW – a16z declared that “the only practical way generative AI models can exist is if they can be trained on an almost unimaginably massive amount of content, much of which (because of the ease with which copyright protection can be obtained) will be subject to copyright.”
A16z argued against the requirement to license copyrighted materials for training AI because “imposing the cost of actual or potential copyright liability on the creators of AI models will either kill or significantly hamper their development.” The VC firm argued that requiring AI developers to pay for copyrighted content would advantage the largest tech companies at the expense of potentially more innovative smaller firms.
Yet perhaps the most incriminating element here is what wasn’t said, according to the legal complaints: neither Suno nor Udio has actually denied using copyrighted works in correspondence with the record companies.
“When plaintiffs directly accused Suno of copying plaintiffs’ sound recordings to train its model, Suno did not deny or proffer any facts to undermine those allegations,” the complaint against Suno states.
“It would have been simple for Suno to say that it used other, legally acquired recordings, if that were the case. Instead, Suno deflected and disingenuously asserted that its training data is ‘confidential business information.’”
The complaint against Udio makes the same allegation, and states that “Udio deflected and disingenuously asserted that its training data is ‘competitively sensitive’ and constitutes ‘trade secrets’ – despite being based on ‘publicly available’ music ‘out there’ for music fans.”
But of course, none of that actually proves that Suno and Udio used copyrighted materials in training their AI. To get closer to that goal, the recording companies looked to Suno and Udio’s musical creations.
2) In some cases, Udio and Suno’s output is almost identical to copyrighted songs – including actual producer tags
Perhaps the most convincing evidence the record companies have to show that Suno and Udio used copyrighted materials is a series of comparisons between certain songs the AI tools created, and well-known (copyrighted) hit songs.
The complaint against Suno shows the musical sheets for a Suno-created track called When Marimba Rhythms Start To Play and Michael Bublé’s recording of the hit song Sway.
The similarities between the two are so obvious that even laypeople not trained in reading musical notation can see it:
Interestingly, the complaint against Udio shows that its AI tool also appears to have copied Bublé’s Sway:
The complaint against Suno includes numerous such examples, including apparent copies of Chuck Berry’s Johnny B. Goode, Bill Hailey and His Comets’ Rock Around The Clock, James Brown’s I Got You (I Feel Good), Jerry Lee Lewis’ Great Balls of Fire, and B.B King’s The Thrill Is Gone.
The complaint against Udio shows apparent copies of The Temptations’ My Girl, Green Day’s American Idiot, Mariah Carey’s All I Want For Christmas Is You, Michael Jackson’s Billie Jean, the Beach Boys’ I Get Around, and ABBA’s Dancing Queen.
Not only that, but the complaint against Suno alleges that some of its output actually includes producer tags, that is, those short little shout-outs some producers add to the beginning or end of a track.
“Jason Derulo’s name is repeated at the beginning of the Suno-generated digital music file aptly titled Jason Derulo…”
Major label lawsuit vs. Suno
“For instance, the Suno output Rains of Castamere begins with the ‘CashMoneyAP’ producer tag, even though the prompt used to generate this digital music file in no way referenced this producer,” the complaint states.
“This output indicates a high likelihood that Suno’s service trained on sound recordings affiliated with the music producer CashMoneyAP, whose producer tag can be heard in the copyrighted recordings by artists such as Da Baby and Pop Smoke.”
The complaint also alleges that Suno copied Jason Derulo’s habit of singing out his name at the beginning of songs.
“Jason Derulo’s name is repeated at the beginning of the Suno-generated digital music file aptly titled Jason Derulo, in a manner exceedingly similar to how Jason Derulo tags his recordings,” the complaint states.
All of this has the makings of a kill shot – but it might not be.
To understand why, we can look to an earlier, and ongoing, copyright lawsuit involving an AI model: AI developer Anthropic‘s defense against a lawsuit brought by Universal Music Group, Concord and ABKCO.
In that case, the plaintiffs provide what they claim to be evidence of Claude copying/regurgitating the lyrics of copyrighted songs in their interactions with Anthropic.
However, Anthropic has argued that the plaintiffs – the publishers who own the copyrighted lyrics – actually “coaxed” the Claude chatbot into producing the copycat lyrics through the prompts they had used.
“Plaintiffs themselves, not Anthropic, engaged in the ‘volitional conduct’ that is a prerequisite to direct copyright infringement liability,” Anthropic stated in its defense in the lawsuit.
We don’t know yet whether the court will accept Anthropic’s argument, but there is evidence that similar “coaxing” was used to elicit the songs cited in the complaints against Suno and Udio.
“Plaintiffs themselves, not Anthropic, engaged in the ‘volitional conduct’ that is a prerequisite to direct copyright infringement liability.”
Anthropic’s argument vs. UMG et al – accusing the rightsholders of ‘coaxing’ its AI chatbot to copy lyrics
In the case of Suno’s version of Bublé’s Sway, the plaintiffs’ prompt included the words “canadian smooth male singer 2004 jazz pop buble sway latin mambo minor key,” as well as “lyrics from the original”.
Thus it seems that the user who created this song had actually fed Bublé’s lyrics into Suno. The Suno imitation of Jerry Lee Lewis’ Great Balls of Fire also appears to have been created using lyrics from the original song, and the prompt “1950s rock and roll, jerry lee lewis, sun studio.”
The same goes for the imitation of Bill Hailey’s Rock Around the Clock, and Chuck Berry’s Johnny B. Goode.
Meanwhile, Udio’s imitation of Mariah Carey’s All I Want For Christmas used the prompt “m a r i a h c a r e y, contemporary r&b, holiday, Grammy Award-winning American singer/songwriter, remarkable vocal range.”
Thus, Suno and Udio could take a page from Anthropic’s book and argue in court that it was the recording companies—or whoever created these tracks—that actually engaged in copyright infringement by manipulating the AI tools into generating songs that were, in effect, copies of well-known copyrighted tracks… and that this is not how Suno and Udio are meant to be used.
Notably, in a response to the lawsuit against his company, Suno CEO Mikey Shulman said that Suno doesn’t allow “user prompts that reference specific artists.” At the very least, we can conclude, from the examples above, that users can easily get around this restriction.
And even if the courts do accept that this shows Suno and Udio were trained on copyrighted works, there is still the question of whether or not that training amounts to “fair use.”
3) The record companies aim to destroy the ‘fair use’ argument
According to the complaints against Suno and Udio, in correspondence with the rightsholders, both companies argued that the use of copyrighted music to train AI falls under the “fair use” exemption to copyright laws.
Lawyers for the recording companies argue that this is pretty much an admission by Suno and Udio that they did use copyrighted works – and they fully reject that this falls under fair use.
The record companies’ argument against fair use centers around the four factors used to determine whether or not a particular use of copyrighted material can be given a pass. These four factors aren’t a set-in-stone formula; fair use is determined by courts on a case-by-case basis, but the four factors are there to guide how judges should evaluate a fair use defense.
These four factors are:
- The purpose and character of the use (especially whether or not the original work is substantially transformed into something different from the original)
- The nature of the copyrighted work (there is more leeway to copy from a non-fiction book than a fiction one, for instance, because disseminating facts can be an issue of public interest)
- The amount and substantiality of the portion taken (the more of a work you take, the less likely it is to be seen as fair use)
- The effect of the use upon the potential market (will it harm the market for the original work?)
The recording companies’ lawyers argue that Suno and Udio’s use of copyrighted music fails on all four points.
- On the first factor, “the use here is far from transformative, as there is no functional purpose for [the AI models] to ingest the copyrighted recordings other than to spit out new, competing music files. That [Suno and Udio are] copying the copyrighted recordings for a commercial purpose… further tilts the first fair use factor against it.”
- On the second factor, the complaints argue that musical recordings are exactly the sort of works that copyright was meant to protect (i.e., unlike with a news article or a non-fiction book, there isn’t much public interest in copying a song).
- On the third factor – how much of a work is used – it’s “abundantly clear” that Suno and Udio ingest “the most important parts” of copyrighted songs, the complaints state, “as demonstrated by [their] ability to recreate, for instance, some of the most recognizable musical phrases, hooks, and choruses in popular music history.”
- And on the fourth factor – the impact on the market – Suno and Udio’s use of copyrighted music pose “a significant threat to the market for and value of the copyrighted recordings,” the lawsuits state.
If AI companies were able to use copyrighted music without a license, “potential licensees interested in licensing copyrighted recordings for their own purposes could generate an AI-soundalike at virtually no cost,” the complaint argues.
Clearly, record companies see AI developers’ unlicensed use of copyrighted music as a potential “game-over” moment for their industry.
The outcome of this story depends largely on whether or not the courts agree with that assessment – and whether or not they care.Music Business Worldwide