When ChatGPT was first released, Alex VestalAs CEO of music production library Rightsify, he found himself with a lucrative new business opportunity. “We realized that all the songs we had and all the metadata about the songs would be extremely valuable to an AI,” he says. “For us, licensing our library was a very quick and easy choice.”
Training a competitive AI model to generate music requires hundreds of thousands, if not millions, of songs and other musical content. While many AI companies believe they don't have to pay for the music used to train their models, citing “fair use,” other companies are taking a more musician-friendly approach by paying artists and rights holders when their music is used to train AI models.
On the surface, the AI industry seems like a perfect new customer for music production libraries: affordable, pre-approved catalogs of songs in a variety of styles. Historically, production music has been popular among advertisers, social media creators, podcasters, and low-budget film and TV producers. These producers need music for their soundtracks, but don't have the time or money to license blockbuster songs, which often have multiple rights holders and come with high fees. As the use cases for production music have grown, so has that segment of the publishing business. As of 2022, MIDiA Research says production music will be worth about $1 billion, including recorded music and publishing combined.
While many artist advocates believe that song licensing is an “ethical” way to train AI music models, it still poses a legitimate threat to the existing music business. [licensing transaction] “If they built that model, they would end up competing with you completely for the same customers.” Anthony Demekin“If they don't approach the deals carefully, they risk undermining their entire business over time,” said John McClellan, CEO of Tuney, an AI music company that makes songs for social media creators and podcasters.
There's no standard agreement for licensing production music for AI training. Despite the long-term risks, Bestall says he's licensed his back catalog to several AI companies (he can't disclose which ones due to non-disclosure agreements). “Typically, we license our back catalog and then make an ongoing commitment to provide a certain amount of music over the life of the agreement, which is two to three years,” he says.
In the short term, these new agreements between music production libraries and AI companies have actually created jobs for more human musicians: Given the new customers looking for more music over the life of the agreement, Bestall is now hiring 24 full-time musicians and almost 100 contract workers to create more music and expand Rightsify's library, which already has over 1 million copyrights.
Lee JohnsonJohnson, CEO and founder of production library Audiosparx, says his business has grown thanks to AI, too. Audiosparx is best known for licensing its catalog for training Stability AI's Stable Audio model starting in 2023, and Johnson says he got permission from the musicians featured in its catalog before agreeing to license their songs to the AI company. Audiosparx acts as a licensor for production musicians, but unlike Bestall's library, it doesn't acquire the songs in its catalog outright. “We approached our artist community about the deal, and about 90% of artists expressed interest in participating,” Johnson says. “About 10% decided not to participate. A lot of people are vehemently against it, so it's encouraging to see so much support.” [AI]… We felt it made more sense to sit on the train and head into the future rather than be hit by the train. [it]. “
Bestall and Johnson say that so far, the partnership with the AI company hasn't affected their other business operations. But Bestall is calm about the changes that could come in the next few years. “We know it's a threat to our existing business lines, but it's a huge opportunity for the future,” he says. “If we cling too closely to our past business models, I think we're going to struggle.” Johnson, who has pivoted AudioSparks' business several times in its 20-year history, shares a similar view on embracing change.
Not everyone agrees. “I think this is short-term financing for long-term losses,” said one. Henry Phippsis an up-and-coming film composer who previously had a full-time job writing production library scores. After researching the future of AI music, he quit that job to work at an AI music startup. He's now back to writing libraries and working towards his dream of becoming a film/TV composer. (Phipps said in an interview.) Billboard “But you can't blame anyone for seizing the opportunity to have their music included in these data sets, because you'll lose a short-term paycheck and others will move on,” he says. “There's no point trying to stem the tide; someone will take the deal.”
For Phipps, the way he creates production music is already similar to the way he gives instructions to AI music. “I get instructions, but it feels like instructions,” he says. “One of the instructions recently was full of reality TV adjectives, and my job is to respond with a musical composition. In some ways, it already feels like a machine's job.”
“Very few people want to be production library composers long term,” Phipps explains. “That's [the music business] “I have to survive, eat, pay rent, and work on projects that are more creatively fulfilling.” Working at an AI music startup has made him “more anxious” about his future opportunities as a film and TV composer. In his view, AI music can augment but not completely replace the compositions of blockbuster film composers. But by reducing opportunities for young upstarts like him, it could “cut off the bottom rung of the ladder.”
Ed Newton Rex“If libraries want to enter into these kinds of deals, they need to carefully consider the terms,” advises David S. Schneider, former vice president of audio at Stability AI and founder of Fairy Training, a nonprofit that certifies AI music companies for properly licensing their training data.
Newton-Rex's particular concerns include ensuring that any AI models that used it are retired or retrained without the library's materials once the contract term ends. “Currently, there is no way to untrain the model, but you can add clauses that control what happens after the license ends,” he said. Newton-Rex also advises libraries to be careful about licensing their data to an “open source model,” a move he says is “completely irreversible” because it makes the model publicly available.
Still, Newton-Rex acknowledges that risks “absolutely” remain. “Musicians involved in music production are at huge risk,” he says. “After all, generative AI is faster, it's cheaper, and the quality is already so much better.”
Just in case, Bestall is on the safe side, having launched his own AI model, Hydra II, to generate royalty-free background music for cafes, hotel lobbies, and other public spaces in case customers prefer AI music to the background music in his current library. Still, he feels the library will always be essential: “I'm not too worried about the possibility that an AI company might say we don't need production music anymore. Human data is invaluable to AI.”
