Once AI Music Makes Money, Who Gets It?

Over the past few weeks, I’ve been looking at how music will be licensed by AI companies — why I think it’s inevitable, how it might happen and what it might look like. Conceptually, this is a fascinating subject, because dividing money fairly among rights holders may depend on having after-the-fact control of training data created by the initial license. But which rights holders get paid, and how much?

Right now, the popular way to look at this issue is to think about what’s fair: how conceptually important songs are compared to recordings, what kind of popularity can be used as a proxy for AI training — that kind of thing. I am sorry to say, although by now it should be obvious, that none of this will matter much. Music and technology companies tend to operate by litigating, lobbying and leveraging their influence. When certain interests align — labels and artists both want to make money, for example — the result can be fair. But the way the industry works is shaped more by money and legal leverage, which is why my newsletter is called “Follow the Money,” not “Follow Your Arrow.”

The most relevant example is the revenue split between recordings and songs. Whatever different industry figures think about their relative worth, the reason that recordings generate more revenue is because labels sell and license them in a free market, while, in the U.S., the music publishing business is constrained by a rate court for mechanical royalties and antitrust consent decrees on the public performance side. This won’t apply in AI, just as it doesn’t apply for synch licenses, where the split between the two sides is 50-50.

In an ideal world, settling this would involve three-way negotiations between AI companies and recording and publishing rights holders. In this one, it’s more complicated. Often, rights holders that sue a technology company for infringement can reach a settlement — negotiation through litigation — that involves both compensation for past liability and a deal going forward. In this case, that could mean that the major labels, which sued Suno and Udio and are now said to be in settlement negotiations, reach a deal that gives them, and the recorded music business in general, an advantage over the publishing side of the business. It could also mean that executives on that side of the business will fight that much harder to get a deal that they see as fair. I’d bet on the latter, and I’d expect to see a split that’s either 50-50 or close to it.

There are also questions about what entities will represent the publishing side. In the U.S., most people react to this as a trick question: Publishers! In Europe, it’s far more complicated, because AI training involves the mechanical reproduction right, which is controlled by the international mechanical collecting society BIEM in most of Continental Europe — often exclusively. Some publishers would prefer to license this themselves, but it could be hard for them to do without undermining their case that AI companies need a mechanical license in the first place. This could also cause a rift between major publishers and indies, which is something both sides prefer to avoid when they get into contentious negotiations with outside industries.

In the age of streaming, the music industry has become accustomed to zero-sum negotiations, where a win for one set of rights holders often leads to less money for others. After all, no one can offer more than 100% of incoming revenue (except for Dean Martin, who according to legend once owed various entities more than 100% of his income). But that’s streaming logic, where piracy and a dysfunctional market have put something of a ceiling on price. Who’s to say whether music AI services will operate that way?

In fact, I think there’s a good chance that AI licensing will compensate rights holders from a pool of money — half to labels and half to publishers, through various channels — and then offer the biggest artists more money for additional rights. Imagine that big AI companies make big deals based on settled lawsuits and mass negotiations. From a legal perspective, that would only give them the right to train their algorithms on copyrighted material — not use specific styles or “likeness rights.” What if a user wants an algorithm to create a song with a vocal in the style of Taylor Swift? (Or even write a breakup song in her style?) Those rights might not be included and the biggest stars might want to be paid for them.

A negotiation that seems to involve two parts of the music business would then involve a third, as well as another set of rights. Those “likeness rights” are not as well defined as copyright, but they could become a value-added product that gives some companies an advantage over others. Right now, only one thing is certain: creators and executives will want more than they end up with.