The way we make and consume music is shifting in predictable ways, and it’s worth resisting the urge to get swept up in the spectacle of it. What’s happening isn’t so much a threat to music as an art form but to music as a recorded medium. Those are very different things. The deeper issues here have less to do with some runaway “terminator” technology and more to do with the older, more human problems of greed and hubris. However, I suspect we’ll get to the Terminator version eventually, too.
I started by testing the popular platforms, Suno and Udio. The experience felt like a jukebox: endless possibilities the first time, but a few sessions in, you realize you’ve already heard everything you wanted to hear. The more interesting territory I came across was the platforms that seem not to care and even promote their companies as having no real rules for using someone’s identity as a “style”. With those, you can generate a convincing approximation of Barack Obama, Eminem, or even me.
Stranger still are the absences. Donald Trump and Bill Clinton don’t seem to exist in the same way, which is genuinely odd in Trump’s case, since he may be the most imitated and most quoted voice on the planet right now. If anyone’s vocal likeness has been captured a million times over, it’s his, and yet here he’s a ghost. Clinton at least has an excuse for the rest of us: even though he famously played the saxophone, he apparently didn’t make it into the training data. The selective gaps say something about how these models are built, and who decided what they were and weren’t allowed to learn. Working through these tools left me wanting to do less with the technology and more about it: write about it, talk about it, and maybe eventually build something new with it.
Digging further, I traced the exact source material behind one AI-generated set I developed, which was genuinely illuminating. I could almost reverse-engineer the whole process, having spent years doing the analog version myself: sampling acoustic recordings into hardware samplers. Nothing about that experience prepares you for something as disorienting as noise diffusion, which we’ll get into with some specific examples in a future piece. The open questions multiply from here: How does this become a legitimate instrument rather than a novelty? Does all of this fit our existing legal constructs, or will we need to revise our legal framework? And what actually counts as “AI” versus what’s just borrowing the label as a marketing shortcut, because a lot of it is the latter.
That tangle of questions is why I reached out to a colleague well-versed in intellectual property, Dr. Charles Eldering. Charles is a technologist, inventor, and patent specialist who is well-versed in identifying the technological, legal, and business issues in emerging technologies. He’s not a copyright lawyer, which is the point. I wanted the perspective of an inventor who has built things and held patents, not a legal opinion on copyright, which, with respect to AI, might not be worth the paper it is printed on. It’s increasingly clear we’re not far from a full collapse of the licensing model media has relied on for decades, a multi-billion-dollar industry that supported a lot of careers, including mine at times. But collapse and innovation tend to travel together. I’ve been talking this through with people outside music too, in forensics, acoustics, philosophy, and IP, because this isn’t really a music problem anymore. It’s a bigger one.
Charles and I plan to unpack this in stages over the next several articles, starting with my own assessment of where my work resides inside the current AI/LLM landscape, where the datasets these models are training on really originated, and what it actually means to build an LLM of one’s own, which may be the only serious, durable approach to an AI music platform for any sound designer or producer who wants to stay in control of their work.
Charles will simultaneously examine both the technological and legal landscape, providing an update on the copyright-infringement lawsuits against Suno and Udio, where both companies concede they trained on copyrighted recordings but assert a fair-use defense; taking a look at how these models are actually trained; and exploring whether a producer-specific, non-infringing model can be built. Stay tuned, both here and on
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.