Universal Music and Warner Music are close to agreeing “landmark” AI licensing deals, according to the FT, whose sources say deal making is underway with ElevenLabs, Stability AI and Klay Vision, as well as Suno and Udio, which are each concurrently fighting legal battles with the majors.
The record companies are also in ongoing talks about AI-specific deals with existing licensing partners like Spotify and Alphabet, which owns YouTube and Google and is developing a range of generative AI products and services.
The frantic deal-making comes as the spotlight is increasingly falling on the huge amount of AI-generated music now being uploaded to the streaming services.
Deezer recently revealed that nearly a third of all tracks uploaded to its platform are now AI-generated, while Spotify last week said it had removed 75 million "spammy" AI tracks from its platform in the last year as it announced various new initiatives to deal with the challenges posed by AI music.
The FT says that the major record companies see these licensing talks as “a proactive way to handle disruption by AI” and to “avoid the missteps” of the early days of digital music in the late 1990s and early 2000s, which resulted in more than a decade of steep decline in the record industry before the subscription streaming business model finally took off.
While the FT’s sources have talked mainly about Universal and Warner deal-making, the newspaper also quotes a Sony Music spokesperson as saying, “We are in discussions with companies that have ethically trained models and that benefit our artists and songwriters”.
Some AI start-ups are already collaborating with music companies of course, and Google has worked with the majors on some pilot projects. However, the FT suggests that bigger catalogue-wide deals could be close, including deals with AI companies that have not yet licensed music from the majors.
The FT adds that the labels want deals which would result in ongoing payments, so that they - and presumably their artists and songwriters - earn whenever AI-generated tracks are played or otherwise utilised, for example by being synchronised into a video.
To calculate what any one licensing partner is due in any one month, the majors “want AI groups to develop attribution technology, akin to YouTube’s Content ID system, that can identify when their music is used”.
Quite how that will work isn’t clear and could ultimately differ from deal to deal. One of the biggest challenges with this sort of system is that the number of pieces of original content that might inform a new AI-generated piece of content could be very large - and may not even be easily quantifiable.
Some of the biggest challenges in AI are around interpretability and attribution - in other words, understanding how AI models make decisions and arrive at their outputs, and what specific training examples or data points influence content outputted by a particular model.
While research in this area is progressing rapidly, the underlying way that AI models are trained raises challenges when it comes to saying things like “this recording or song informed this output to this degree”.
There are various mechanisms that can be used - so-called influence functions amongst them - but there’s still a long way to go before there is a reliable and scalable solution to what is one of the ‘big challenges’ in the development of AI systems.
Add to this the complexities of politics and trust within the music industry, and it becomes even thornier. If we are now moving into an era where the big AI companies start to do big deals with music companies, agreeing how money should be allocated to individual tracks and catalogues will be just one of a number of contentious debates within the music community.
Because - while one label exec insists in the FT article that most artists “trust” the majors to do the right kind of deals with the AI companies - in reality there is very little trust in the music business.
Many music creators assume the majors will negotiate secret deals that prioritise their own share prices and profit margins. That’s pretty much what happened when the original streaming deals were negotiated, and - if the same happens with AI licensing deals - artists and their managers are likely to be left in the dark, wondering exactly how badly they are being screwed over this time.
Getting AI companies to the negotiating table
To date, the big debates around music and AI have mainly centred on whether or not an AI company even needs licences from the music industry when it trains its generative AI models with existing tracks. Many AI businesses argue that AI training is or should be covered by exceptions in copyright law, or the ‘fair use’ principle under US law, meaning they don’t need permission from any copyright owners.
Countless lawsuits have been filed by copyright owners against AI companies, including the major labels’ lawsuits against Suno and Udio. However, realistically, it will take years for those legal battles to get to the top courts where truly binding legal precedents can be set.
Meanwhile lawmakers around the world have been procrastinating about clarifying copyright law - despite calls from creators and copyright owners for legal reform - with most political leaders wanting to appear simultaneously pro-tech and supportive of the creative industries.
That said, one of the key test legal cases in the US - between a group of book authors and Anthropic - favoured copyright owners, even though the judge said AI training was in principle fair use. That’s because he ruled it was only fair use if training content was curated from legit sources. Anthropic had pirated millions of ebooks and therefore faced trillion dollar damages, prompting a $1.5 billion settlement.
Although that precedent only applies in the US - and could be superseded by other cases in appeal courts and ultimately the Supreme Court - just the prospect of billion - or even trillion - dollar damages down the line is enough to force many AI companies to the negotiating table.
Which is good news for the majors who are keen to turn generative AI from a big threat into a big lucrative opportunity. Though once we are into the licensing deal phase, a whole load of new debates kick off, with lots of big questions to be answered within the wider music community.
The big debates
First, how exactly will the deals work? If labels get paid when AI-generated tracks are played or utilised, how is the system connecting any one AI-generated track to the human created music that influenced it in order to know which artist or label to pay?
What is technically possible in an industry where metadata is already a challenge? Are these licensing deals going to be a ‘best guess’ or are AI companies going to accelerate their work on the technical challenges around identifying what content informs what outputs?
The FT article admits that different platforms may employ different licensing models, but it also indicates that the initial deals could set a business model that is then employed more widely, as happened with streaming. That has its own implications and own set of challenges for the music industry.
Second, do individual artists get to choose whether their music is part of these deals? Many music creators argue that their consent should be secured, which would give artists the opportunity to decline to participate, and to negotiate how they will share in any income that is received by their labels.
Session musicians and studio producers also argue that they should have consent and compensation when it comes to AI deals. Not least because many session musicians worry that their performances on existing recordings will be used to train AI that could ultimately result in a significant reduction in the amount of work for musicians.
One senior label exec quoted by the FT says whether or not artist consent will be sought depends of individual record contracts, but then adds “it would be impossible to go back to artists each time and ask, ‘are you OK with this use of your music?’”
But, insists the unnamed label exec, the good news is that most artists say “I trust you to do the right deals on my behalf”, so deal by deal consent isn’t necessary. Which - given the current debates within the music creator community - seems like wishful thinking bordering on delusional.
But it is confirmation that the majors don’t plan on seeking artist consent unless their lawyers reckon a record contract explicitly requires them to do so. Or, presumably, if the AI is also imitating an artist’s voice or likeness, which is generally where both labels and AI companies are more nervous about legal liabilities if they don’t get explicit artist consent.
But beyond specific and unusual contract terms - and AI-generated voice clones - it seems the majors reckon they can do these deals without getting music creator consent. Which is great news for shareholders and executive bonuses, but not such great news for musicians and other creators.
The third debate relates to the two copyrights in music - the recording rights and the song rights. If AI companies concede that they need licences to use existing music in their training datasets, that means two sets of licences, one for recordings, one for songs.
While the majors control big song catalogues as well as recording catalogues, that doesn’t mean they control the rights in the songs contained in their recordings. Plus in at least some markets collecting societies will have to be part of the deal making on the songs side.
In streaming, the recording rights earn three to four times more than the song rights, and that model has been applied to music in videos on social media platforms too. Even though with video the music is being synced - and, with sync, recordings and songs usually earn equally. Plus during last year’s Universal v TikTok spat, it felt like video creators were more concerned about the song than the recording.
It seems likely that songwriters and indie publishers will fight any suggestion that the streaming split should also be transferred into the AI deals. But the major labels may well think differently.
So, while the debate about the copyright obligations of AI companies may come to an end at some point soon - with an increasing number of AI companies conceding they need licensing deals - there are plenty more debates to be had about rights and revenues when it comes to music and AI.
And it may well transpire that - while copyright law reform isn’t necessary to force AI companies to the negotiating table - it will be needed to balance the interests of music creators and corporate rightsholders.