I’ve Made 100 Songs With AI. I’m Starting to Think We’re Asking the Wrong Questions About It.

I started thinking about this because SOCAN is suing Suno.

For anyone who hasn’t fallen down this particular rabbit hole yet, Suno is a generative AI music platform. You can give it instructions, lyrics, musical ideas or recorded audio and have it generate finished music.

I’ve used it extensively.

I’ve also been a musician all my life. I started playing when I was eight. That’s probably why the argument surrounding Suno started bothering me.

SOCAN filed a lawsuit against Suno alleging copyright infringement. Among other things, it says it has identified 150 outputs available on Suno that reproduce all or substantial portions of songs in its repertoire.

If Suno is reproducing copyrighted songs and distributing them, I don’t find that part particularly difficult. Copyright law exists for a reason. If I use AI to reproduce somebody else’s song and start distributing it, calling it artificial intelligence doesn’t magically make the copyright disappear.

The harder question is how the machine learned to make music in the first place.

Suno acknowledges that its models were trained using tens of millions of publicly available music files from the internet, including material protected by intellectual property rights.

That’s where I started getting stuck.

Because I’ve been doing essentially the same thing my entire life.

I Can Hear “Jump” Right Now

I don’t have to put on a Van Halen record to know what “Jump” sounds like.

I can hear it in my head.

I can hear the synthesizer. I can hear the drums. I know the colour of the recording. I know how Eddie Van Halen plays. Decades of listening have left an enormous amount of musical information stored somewhere inside my brain, and some of it inevitably comes out when I play.

That’s how musicians learn.

We listen. We imitate. We experiment. Eventually thousands of influences get mixed together with our own instincts and something comes out that sounds like us.

My memory isn’t remotely as precise as a computer model. But information entered my brain and changed it. No one regards the resulting memory as an illegal copy.

There’s an uncomfortable wrinkle here.

I grew up through cassettes, MP3s, Napster and file sharing. I’m not going to pretend every piece of music I’ve ever heard arrived through a perfectly licensed chain of custody.

Suppose I heard an unauthorized MP3 25 years ago. The MP3 itself may have been illegal. But I listened to it, remembered it and learned from it.

The illegal copy and what I learned from it are two different things.

My memory doesn’t become illegal. Everything I subsequently create isn’t contaminated because something that taught me music came from an unauthorized source.

Yet when the learner becomes a computer, we start collapsing those things together.

So Buy the Records

Suppose Suno spent a billion dollars and lawfully acquired every recording it wanted for training.

Every album. Every single. Every obscure jazz record. Every punk record. Every symphony. Every terrible record somebody made in their basement in 1987.

Now let the computer listen.

If Suno lawfully obtains the music, analyzes it, learns mathematical relationships from it and produces a model that doesn’t contain playable copies of those recordings, what exactly does the copyright holder continue to own?

The recording and composition? Absolutely.

The right to reproduce and distribute them? Absolutely.

But what about what somebody learns by listening?

If I buy a Van Halen album and study Eddie Van Halen for five years, no one suggests Warner owns part of whatever I subsequently play.

If a machine studies the same record and learns something about rhythm, harmony, guitar tone and arrangement, we need a better explanation than “one learner has ears and the other has processors” for why the principle changes.

There may be one. Computers make technical copies that brains don’t. Scale is different. Commercial purpose is different. A machine capable of learning from millions of recordings can compete with the people who made them in ways one kid learning guitar in his bedroom never could.

Those are legitimate arguments.

But they lead to a much more precise question than whether AI “stole” music:

Does copyright include the right to control what a machine learns from a work someone has lawfully acquired?

That’s a question we’ve never really had to answer before.

We’ve Already Changed What a Recording Is Worth

The music industry has been moving toward this confrontation for decades.

I remember buying vinyl records. CDs eventually cost around twenty bucks. Apple broke the album apart and sold individual songs for about a dollar.

Then streaming changed the transaction completely.

For roughly the price we once paid for one album, we could suddenly listen to almost everything.

Spotify didn’t eliminate the recording. It inserted itself between the music industry and its customer.

Record companies had always dealt with retailers, distributors and radio stations, but the fundamental product remained theirs. You bought the record.

Spotify changed the product to access.

The labels still owned the recordings, but Spotify increasingly owned the relationship with the listener: the subscription, interface, playlists, recommendations and discovery.

To the subscriber, the incremental price of listening to one more recording became effectively zero.

Copyright didn’t become worthless. Catalogues, publishing, performance, synchronization and licensing rights remain enormously valuable. But the individual recorded copy, once the centre of the business, was economically hollowed out.

AI threatens another change.

Spotify needed the record industry’s recordings because its customers wanted to hear those recordings.

Suno’s customer may not want a particular recording at all.

They may simply want music.

That explains at least some of the anxiety surrounding generative music. This isn’t another Spotify trying to become the new distributor of the industry’s product.

It’s technology capable of producing an alternative to the product.

Then I Started Hearing Suno

When I first discovered Suno, it was astonishing.

Type something into a box and a finished recording came out.

Holy shit.

I went through it like everyone else. Rock, pop, orchestral music, different singers, different eras, different production styles.

Eventually I started hearing the machine.

The newer models may sound better technically, but to my ears they’ve increasingly fallen toward the musical centre. Familiar pop progressions. Cycles of fourths and fifths. I-IV-V. ii-V-I. vi-ii-V-I. Familiar melodic resolutions and vocal gestures.

Change the singer. Change the distortion. Change the drums. Call one country and another metal.

Underneath it, I increasingly hear the same musical instincts.

When somebody tells me, “I made this with Suno,” I can often hear Suno before they tell me.

The magic trick wore off.

That’s when the technology became much more useful to me.

I Started With My Demos

I dumped my demos into Suno almost immediately.

These weren’t polished recordings. They were rough sketches, unfinished arrangements and ideas I’d recorded quickly before they disappeared.

Now I’ll often sit down with an acoustic guitar and play a part. I don’t bother plugging into an amplifier or spending an hour finding the right tone.

I give Suno the musical information and tell it what I want that information to become.

Maybe I want a late-1960s Marshall Plexi sound. I’ll describe the amplifier, gain structure, cabinet character, production and performance I want around the part I’ve already played. I even tell it the knob settings on the amp.

The acoustic guitar isn’t the finished sound. It’s my instruction to the musicians. That’s what Suno has become for me. I’m not asking a computer to turn me into Mozart. I’m Brian Wilson standing in front of the Wrecking Crew.

I bring the song, progression, melody or guitar part. I decide what I want to hear. The technology gives me an extraordinarily capable collection of studio musicians to realize it.

I’m not asking AI to replace my musical judgment. I’m using it to execute it.

I suspect that’s where this technology becomes much more interesting.

“Make me a rock song about my girlfriend” is a novelty.

“Here’s my song. Help me make the record I hear in my head” is a tool.

I Didn’t Suddenly Write 100 Songs

I’ve probably completed around a hundred songs using Suno.

From the outside, that may look like AI wrote them. It didn’t. I didn’t suddenly produce a hundred new songs; I finished a hundred songs that had been accumulating for years.

Most of that first explosion of productivity came from material that had already accumulated over a lifetime. Old musical ideas. Fragments. Unfinished songs. Things sitting in my head or buried in archives.

I had nearly five decades of unfinished musical material.

AI helped me work through it.

Then something revealing happened: my production slowed down.

Suno didn’t. It can generate songs indefinitely.

I slowed down because I’d worked through much of the material I’d been carrying around. Now I have to create new music before there’s new music to finish.

If the machine were the source of the creativity, there’d be no reason for that slowdown.

The Same Thing Happened With My Writing

The last year has been enormously productive for me outside music too. Books, political work, essays, The Canadianist.

Again, I understand how it looks.

AI.

Except much of what I’ve been writing about didn’t appear when ChatGPT did.

I drive for hours every day. I’ve spent years thinking about Canada, politics, democracy, culture, economics, music and life. There was an enormous backlog sitting in my head.

Generative AI dramatically shortened the distance between having those thoughts and turning them into finished work.

It made me more productive.

It didn’t give me an infinite supply of things to say.

Eventually I catch up with myself there too.

That’s probably the best metaphor I have for what AI has done for me.

It’s been a vacuum.

I’ve spent decades dropping things on the floor: songs, fragments, arguments, stories, observations, things I wanted to write, things I wanted to build.

AI gave me a way to suck all of it up and do something with it.

Eventually the vacuum reaches the clean floor.

Apparently I’m AI Now Too

The strange consequence of that productivity is that people increasingly assume anything I make must have been made by AI.

I posted a video during a violent storm yesterday. I was sitting in my car while hail hammered the roof, talking into the camera.

Someone commented, “Hey, you’re getting good with AI.”

AI?

That was me.

That was my car.

That was a real storm.

The AI contribution was a lower-third graphic at the bottom of the screen.

Someone else has accused my actual recorded voice of being an AI voice.

We’ve reached a strange point where suspicion of artificial intelligence can erase the human being standing right in front of us.

For most of history, extraordinary productivity suggested somebody had worked like hell.

Now it can be offered as evidence that they didn’t do the work at all.

“AI-Generated” Doesn’t Tell Me Much Anymore

Two people can make a song using Suno.

One types, “Write me a sad country song about losing my girlfriend,” accepts the result and publishes it.

Another writes the song, plays the guitar, supplies the melody and chords, determines the structure, uploads the performance, specifies the production, rejects bad takes, changes sections and directs the finished recording.

We call both “AI-generated.”

The description has become almost meaningless.

Writing has the same problem. There is an enormous difference between asking a language model to write an essay and using one to interrogate, organize, research, challenge and edit an argument you’ve spent years developing.

The finished work may involve AI in both cases.

The human contribution isn’t remotely the same.

What Exactly Are We Protecting?

That’s where I come back to the lawsuit that started me thinking about all of this.

Suno didn’t give me the music I started playing when I was eight. It didn’t give me decades of listening, the unfinished songs in my head, the guitar parts I play into it, my taste or my ability to know when something sounds wrong.

It helped me realize those things.

Copyright should protect creators. If Suno unlawfully copied recordings, deal with the copying. If its outputs reproduce somebody else’s song, deal with the reproduction. If commercial AI training requires a new licensing framework, let’s have that argument.

But learning is something different.

Human creativity has always been cumulative. Musicians carry other musicians around in their heads. Writers carry what they’ve read. Painters have seen paintings. Filmmakers have watched films. We absorb what came before us and eventually add something of our own.

AI is forcing us to put a legal boundary around something we never needed to define this precisely before: where intellectual property ends and learning begins.

At the same time, it’s forcing us to rethink authorship from the other direction.

We’ve become deeply concerned about recognizing the human work that went into training AI, while sometimes refusing to recognize the human work that goes into creating with AI.

Both deserve a more sophisticated conversation than “AI stole it” or “AI made it.”

I’ve seen what happens when I give these machines nearly five decades of music and years of accumulated thought.

I’ve also seen what happens when I have nothing new to give them.

The difference is enormous.