The death of stock music
How AI is replacing human music, and what we should do about it
Music is all around you, even when you may not know it. Behind every advertisement, theme park waiting line, movie trailers, video game announcements — catchy music has been composed to keep your brain engaged and entertained.
This type of short-form “stock” music (jingles, royalty-free tracks, incidental music, others) has historically been an important revenue stream for instrumental musicians. Playing in bands or as a studio musician doesn’t often pay the full bills, and being able to soundtrack various parts of everyday life has been vital. An entire industry of music licensing has been built around this, giving lesser known artists the ability to live their dream jobs while keeping a steady paycheck.
Sadly — this is a segment of the music industry that AI can, and likely will, almost entirely replace.
Don’t believe me? Listen to the following generations from the current top performing AI music generation platforms Udio and Suno:
Platform: Udio
Prompt: Orchestral score, cinematic chords, emotional, violin solo, swells
Output 1:
Output 2:
Platform: Suno
Prompt: Inspirational song for the background of a presentation
Output 1:
Output 2:
These songs aren’t perfect, but they are good enough to replace a large number of existing stock music sources.
The (big) problem
So why can’t human-made and AI-made stock music thrive in the marketplace together? It can for a short period of time, but as the music industry continues in its endless pursuit to shrink the bottom line, it may come to an upsetting conclusion:
AI gives corporate rightsholders the ability to cut artists out of the equation.
For a studio or label looking to find a suitable track for advertising, the AI-generated method offers an immediate and highly customizable solution. To compound the problem, these companies typically retain the consumer demand data. They know exactly what their consumers want at any given time.
Stock music is merely the current low-hanging fruit. It follows trendy music structure and doesn’t need to be high quality to sell. But this is simply the market that is currently most vulnerable to AI generations. The problem will soon extend beyond this.
Listen to this Suno generated track (both lyrics and music AI-generated):
Music prompt: Pop country, strong vocals, energetic
Lyric prompt: Country song about a man who finds out he's actually AI
Output:
This is a track that the average Spotify listener might easily expect popping up on their casual summer playlist. The AI music generation model gives a label the ability to analyze their market data, generate the above track based on current consumer demands, distribute the song globally, and retain all rights and revenue. And not a single artist would be involved.
The (necessary) solution
Legal protections for existing music artists are needed and necessary. There are two immediate areas where the oncoming AI music storm causes the biggest problems:
Music used in training data without permission. Most (not all) mainstream AI music generation models use non-licensed music. This is a standard practice for AI companies, who typically will ingest large amounts of publicly available internet data For music, its likely that these models have scraped the entirety of YouTube’s music library, which is “public” in that its easily accessible by automated scraping software.
The decimation of the human stock music market. As discussed, the ease of AI music generation (coupled with its existing high quality that will only get better) will lead to the demand segment of the market skipping over human artists together and becoming the source themselves. Eventually, this will extend outside of stock music and put the entire longtail music industry at risk.
With the above in consideration, a two-fold solution is necessary:
Require permission for training data. It’s a no brainer, but something that is constantly circumvented out of model development. Get permission (and pay!) for the copyrighted content in your training data. Companies should also consider a retraction and re-tooling of their current models, which contain copious amounts of copyrighted music.
Pay the artists for every use. Every time an artist’s song is use within the context of a song generation, the platforms need to pay the artist. The unfortunate truth here is that generative AI models currently function in a manner where it’s difficult to tell which parameters are being actively used — an immense amount of training data is used for each song generated. So platforms like Udio and Suno need to both add a level of transparency around this, and rework their models to better trace back to songs used in each generation. If that’s not possible, don’t use any non-licensed music in the model at all (back to #1).
It’s a simple and straightforward solution, but requires an immense amount of pressure on the generative AI platforms to concede that their technology is improperly developed and needs to be retrained.
There are parties interested in applying this pressure, but will it be enough? Or will the platforms gain popularity, pay some fines, and watch as the musicians are laid to rest?