AI generated Drake and the Weekend Collab, the song that mimicked each of them and quite convincingly went viral in 2023, racking up over 11 million views. And all was fine and dandy until Universal Music Group took down that track across every platform. June 2024, three major music labels, Universal, Sony, and Warner Music Group, all three filed lawsuits against AI music generators and UIO, accusing them of mass infringement of copyright sound recordings on an almost unimaginable scale. This is a deep dive into how AI is changing the economics of music streaming, how royalties flow, why labels are suing AI companies one day and partnering with them the next. How fake streams turn into very real money. And the real reason the Hollywood and the music industry is taking tech companies to court. And spoiler, it's not ethics, it's the money. Let's dive in.

How does AI produce music? For an AI model to compose or produce music, before a single note is generated, it is fed a massive data set of existing audio files like MP3s that get converted into a spectogram. Specttograms are visual representations of music, which are essentially pictures of sound that show frequency and amplitude over time. And the model learns from those statistical patterns. What combinations of frequencies appear together in jazz versus hip-hop versus rock versus pop? How to create the sound of magic? How to create the sound of tension? How to create a sound of investigation? And what a verse sounds like versus a chorus. How Barack structures differ from postmodern. How Drake style rap differs from a classical tenor and so on. The curious thing is that this is the very thing that composers are trained in from music theory, sulfedio, harmonies, all of those disciplines take years to learn so that a musician can spot those patterns and produce new. Okay, so put that aside. That was the first part on how the model learns. And now we'll talk about the inference, the actual production of music. When you type cinematic background music with magical beats, the model generates a new audio file by predicting what sounds should follow each other based on everything it learned in the past. The music that is generated never repeats the original recording, but it couldn't produce anything without having consumed them first. So on one hand, you can't blame it for copyright because well, it hasn't copied anything. But again, it wouldn't be able to produce anything without learning first. But how is that different from a human learning music? A musician listens to thousands of songs, learns the patterns, mimics what others had composed before, and eventually learns the skill of composing. How are these two different? The difference isn't as much about the ethics of creation. I feel like when it comes to music or acting or art in general, there's a lot of chatter about the ethics of copying someone's work or producing content based on somebody else's work. But the core difference is that unlike an individual, media is a business and being the beast of an industry that it is, it's a money-making machine. The difference between a musician listening and copying and a model listening and learning is the scale. A musician or a composer listens to music as a consumer even when they learn and mimic existing styles. They still pay for the albums. They buy sheet music. They stream songs. They pay for concerts. They pay for master classes. They must purchase access to a song if they want to use it in a YouTube video. All of these are components of the consumption cycle that funds the creator ecosystem.

Let's use an example. The legendary Another One Bites the Dust by Queen. When I play this song on Spotify, who owns the rights and who makes money off of it? Listen closely. First of all, it is really important that you understand who owns the rights to this song. The composition or the written song, the melody, chord progression, the baseline, the lyrics belongs to the songwriter. Another one bites the dust was written entirely by Queen's basist John Deacon. He is therefore the sole composer and the owner of the original composition copyright. This is the first copyright that you need to know about. John Deacon wrote it in 1979 which under the US copyright law equals the life of author plus 70 years. So the composition copyright will not expire until at least 2096. The publishing rights, however, were originally held by Queen Music Limited and have since moved to Sony Music Publishing when Sony acquired Queen's catalog. And the copyright number two, the master recording, which covers the specific recorded performance of the song for Queen. This was originally owned by Queen Production Limited, which was the band's own company. In North America, the master rights were licensed to Hollywood Records under Disney Music Group in 1991. In June 2024, Sony Music acquired Queen's entire catalog, their recording rights, publishing rights, name and likeness rights for $1.27 billion, which is the largest music catalog deal in history till this day. So, as of today, when Another One Bites the Dust streams anywhere in the world, including Spotify, the royalty flows to Sony Music because they are the owner of the second copyright. John Deacon, who hasn't performed publicly since 1997, is still collecting publishing royalties because he has retained his songwriter share that flows directly to him, regardless of who owns the publishing administration.

Now, what happens when the song is streamed on Spotify? When the song is streamed, it splits into two parallel royalty rivers. The recording royalties and the publishing royalties. If I play the song in Spotify, Spotify collects all subscription and ad revenue, takes 30% for operating costs, distributes 70% to rights holders, and of those 70% roughly 50% goes to the sound recording owner. For another one bites the dust, that flows to again Sony Music. And after that, they pay the Queen's members or estate per whatever revenue sharing agreement was part of the catalog sale. And this brings us to the root of the problem. When an AI model gets trained on tens of millions of recordings simultaneously, it generates absolutely zero revenue for any rights holder behind those records. This is the root cause of the problem that the music industry or the writing industry or the movie industry has with AI. It's not so much about the ethics of it because everybody copies everyone. But the core issue is that the model companies don't pay for the data that was used for training. A logical question here would be but why doesn't it trigger the same royalty and copyright chain and that is because when you stream on Spotify the payment chain fires because there is a trigger the streaming of the song. The streaming is a licensed legal event by definition. AI training on the other hand doesn't stream. When AI scrapes, the audio file gets loaded into the memory temporarily. It is then converted into a mathematical representation like a specttogram and those representations are used to update the model's weights and the original audio file is then discarded. The model does not contain the audio. You can't hold it accountable because the audio isn't there. The model will never repeat the original or even a slight variation of the original. So the argument on the AI model company side is that they didn't make any copies permanently. There was no reproduction and no distribution. How is that a copyright infringement? The music labels argument, however, is that it is an infringement. And it is an infringement because the AI model produces music and that music competes and can substitute for the original recording used to train it. On top of it, add the whole intellectual property maze. Remember the example from the beginning, AI generated Drake and the weekend collab. So it was taken down by the Universal Music Group and their legal basis for the video takedown wasn't because it was AI generated. In fact, if you break that video apart, there is no face or picture cloning happening. It's only the music and the voices. But voice cloning isn't, or at least wasn't at the time, clearly illegal under the US copyright law. But the song was taken down because it contained a sampled producer tag. A producer tag is an audio watermark from producer Metro Booming that you can hear in the first 4 seconds of the tune. This story with The Weekend and Drake tells you basically everything you need to know about how prepared the law was for this moment.

Now, let's talk royalties. Everybody knows that it's a big thing in the entertainment industry, but I want to break down how exactly AI changes the economics of royalties. What's important to know is how the royalty model works in the streaming services today. Think of it as a pool model. The math runs like this. Let's say we take one month as a set period of time, January 1st to January 30th. Spotify collects all revenue from day one to day 30 for subscriptions and advertising. Let's say it's $1 billion. Then Spotify takes 30% of the top as the platform fee and that goes straight to Spotify. The remaining 70% becomes the royalty pool and the payout for the artist goes like this. Their streams divided by total platform streams multiplied by total royalty pool. This means that the more streams you've got competing for the same poll, the more it gets diluted and therefore less money per stream for everyone. AI music becomes an additional stream. So it indeed dilutes the pool. And the dangerous part is that AI makes music generations almost free and instant. So, if you've got a bad actor who can generate hundreds of thousands of tracks and upload them all to Spotify through a distributor and then deploy a bot to stream them on loop, what happens is that you blow up the denominator because every stream adds to the total platform streams and that takes a fractional slice of the royalty pool. Now, if you multiply that across millions of fake streams per day, you can see how the artist payout starts to shrink. To make this math unattractive for bad actors or for those who want to take advantage of it, you have to change the entire incentive structure because right now it is undeniably positive for the bad actor. As long as AI music generation is free and quick. Distribution to streaming platforms cost $20 a year. The royalty pool is zero sum and proportional. Bot streams cost fraction of a penny each. You can generate endless tracks and extract real money from a pool funded by real users. If Spotify deletes AI spam, you generate 100,000 more. Apple Music's response to this is actually more meaningful because they added 10 to 50% penalties for accounts linked to fraudulent activity. That way, they punish the distributor account, not just the AI track. But the underlying math still doesn't change.

A logical question here would be why are they not changing this model then? And the reason for it is the resistance from music labels because the pro-rator model massively favors the superstars who dominate streaming. I'll show you why. Remember how the pool works. An artist's payout equals their streams divided by total platform streams multiplied by total pool. Meaning that if you have an artist with 10 billion streams, say Taylor Swift, who captures an enormous percentage of the entire global pool, her label, The Republic Records, collects that proportional share for every subscriber on the planet, including subscribers who never once played a Taylor Swift song. I am one of them. Think about it this way. Under the Pro Raa model, today a h 100red million subscribers a day pay $10 a month. That is a $1 billion pool. Taylor Swift gets 5% of all streams globally. That is $50 million. That 50 million figure includes the money from subscribers like me who never played her once. And now flip it to the user centric approach. the same $1 billion poll, but Taylor Swift only gets the money from subscribers who actually streamed her content. Perhaps 30% of the entire Spotify user base actively plays her songs. She gets 5% of that 30%. Which is roughly $15 million. 50 versus 15. At the same time, across the US and Europe, there are several cases that the music industry is watching very closely over the next several months. A copyright litigation against OpenAI that could define whether training on creative work is even legal. The Gemma versus Sunno ruling that will define the European legal standard for AI music positioning. And the Writers Guild of America negotiation cycle that just ended actually on April 4th. And the outcome of that negotiation cycle is actually very telling. studios agreed to notify and consult the Writers Guild of America if they license writer content for AI training. Plus, they added licensing requirements when studios use AI drawing from existing scripts, but writers were not granted payment for AI trading data. The irony is that the writers specifically initiated those stocks to get paid because they were the trading data and ended up at a loss. So while these battles are playing out in courtrooms, the major labels gatekeep the existing royalty structure that is skewed towards superstars because when it became clear that AI content isn't going anywhere, the labels preferred to control. And from here everything fractures into three possible outcomes. Outcome one, a licensing economy emerges where AI companies pay to license content and creators get their payments for supplying the training data. It is slow, it is expensive, but it is sustainable. The second possible outcome is when the extraction economy takes hold. That is going to happen when courts greenlight mass scraping. AI companies would pay nothing for training and the majority of value would flow to them. And the third outcome, a net new framework built around shared pools. The platform level revenue will split or perhaps we're going to see entirely new forms of intellectual property. Whatever happens next is going to be about ownership, leverage, legal workarounds, and whether the people whose work train the models will have any claim to what comes out of them. We hope this was helpful, and we'll see you in the next episode of the business of AI. Bye.