MBW Views is a series of op-eds from eminent music industry people… with something to say. The following MBW op/ed comes from Deborah Mannis-Gardner, Owner/President of DMG Clearances, Inc., who knows everything there is to know about clearing samples for popular music, TV shows, and more.
Below, she argues the industry’s existing sample-clearance framework could now be sensibly applied to AI licensing…
AI is sampling. There, I said it. And as someone who’s been handling sample clearances for over three decades, I should know.
When generative AI first hit the music industry, many didn’t know what to make of it. Here was a new, rapidly growing technology that was capable of creating a serviceable ‘Fake Drake’ track, and the usual suspects were claiming ‘fair use’. Were we in the midst of another Napster-style industry disruption?
The answer is no, but many would sure like it to be yes. Of course, technology companies are going to claim ‘fair use’ whenever they need copyrighted material to make their new products work correctly. So when Suno and Udio were sued for the age-old “ask forgiveness, not permission” approach and claimed ‘fair use’, it wasn’t a big surprise.
Now, deals have begun to be cut, and many companies are starting to pay to license tracks for their generative AI models. Yet still, there remains debate around what this licensing system should look like.
Which brings me back to my original point: AI is sampling. The key thing to understand here is that AI can never generate something new; it can only recontextualize existing data that is fed into it.
When it comes to music, that data is the 1s and 0s of digital music files. So when you ask an AI platform to generate a song, the result is inherently tied to the music that was used to train the model.
Take the example from this year’s Winter Olympics in which an ice dancing duo used an AI-generated track as part of their performance with lyrics suspiciously similar to Bon Jovi.
This track replaced another AI-generated song they had used earlier in the season, which pulled several lyrics wholesale from You Get What You Give by New Radicals. It goes to show just how much AI relies on its inputs to generate outputs – even when a new track was created to replace the clearly infringing one, it still ended up pretty close to the original lyrics used in its training.
Now let’s take a look at sampling. In this case, artists are taking pieces of pre-existing songs and using them to build new tracks. The resulting songs might sound different, but they are still inherently tied to the original, just like AI. Imagine removing the Annie sample from Jay-Z’s Hard Knock Life (Ghetto Anthem) or The Clash sample from M.I.A.’s Paper Planes. The former would have no chorus at all, and the latter would have no beat.
“Sampling is now considered an art form. AI-generated music can get to the same place, but its proponents must let go of the ‘fair use’ argument.”
When I first entered the sampling world in the early 1990s, the prevailing wisdom was that it was theft. Over time, and after we worked out a viable licensing framework, sampling is now considered an art form, and rightfully so.
AI-generated music can get to the same place, but its proponents must let go of the ‘fair use’ argument and embrace licensing. Here’s the good news: because AI is sampling, there’s already a framework in place.
It sounds complicated at first: to clear a sample, all rightsholders in both the recording and the publishing must sign off on a license and be paid accordingly. It can sometimes be hard to find them all, and it’s certainly not cheap.
But we’ve figured out ways to simplify the process. Look at services like Tracklib, which offers pre-cleared samples for a standard price. We can easily apply this to generative-AI licensing. In fact, companies like SourceAudio already are!
The main difference between sample licensing and AI music licensing lies in the output. Sampling deals clearly lay out how much of the copyright in the new work goes to the original artists, and we have infrastructure in place to make sure they get their royalties. This is possible because we know exactly what is being sampled and how much of it is used in the final product.
It’s more difficult with AI, as it’s often unclear how much of any particular input track was used to create any specific output track. Once we resolve this issue, we’ll have a tried-and-true system that we can apply to AI licensing – so why aren’t we fully focused on this?
The answer is that tech companies are determined to pay as little to artists as possible. That’s why we’re still talking about ‘fair use’ and fighting legal battles over uncleared usages rather than fixing the one hiccup that will bring AI licensing into the legal sphere. Once we accomplish that, the possibilities become truly exciting.
We’ve seen talented musicians build an entire genre from a crate of vinyl and a dream. Imagine what they could do with properly cleared AI music models at their fingertips!
All MBW+ subscribers get digital access to our Music Business Worldwide magazine, with six issues released each year. Music Business Worldwide
