My first stumbles with AI

When people see I've released 273 singles in about two years, they sometimes imagine it all went smoothly from day one. That I turned up, pressed a button and songs started rolling off the line like hot cakes. That makes me laugh, because my beginnings with these tools looked more like someone trying to assemble Swedish flat-pack furniture without the instructions and ending up with six screws left over.

I want to tell you about some of those stumbles. Not to make you feel sorry for me, but because I think you learn more from what went wrong than from the heroic version.

Thinking the machine knew what I wanted

The first mistake was one of attitude. I thought that if I gave the tool a good lyric, it would hand back the song I had in my head. As if it could read minds.

It doesn't work like that. The machine doesn't know a line is a joke. It doesn't know a certain word needs to land with a knowing edge. It reads the lyric and does something with it, sometimes something very beautiful, but without understanding what I meant. The first few times it frustrated me enormously, because I'd hear the result and think the song was "wrong", when really what was wrong was the way I'd asked for it.

It took me a while to understand that the tool needs direction. That it isn't enough to know what you want: you have to learn to say it in a way that's useful. That's a craft in itself, and at the start I didn't have it.

Falling for the first version

The second stumble was sillier. I'd keep the first thing I liked.

I'd hear a version, get excited, and that was it, that was the one. No comparing. No letting it settle. Over time, going back to some of those decisions, I realised I'd been carried away by novelty. What thrilled me was that the song existed, not that it was good. They're two different feelings that get muddled at first.

These days I'm quite suspicious of my first burst of enthusiasm. Not completely, because sometimes it's right. But I make it wait a little before I agree with it.

Writing for the machine

This was the worst one, and it took me a while to see it.

For a spell I started writing lyrics with an eye on what the tool did best. Simpler lines, more predictable rhymes, structures I knew would work. The songs came out cleaner. And blander. I was adapting the way I write to the machine instead of making the machine adapt to me.

I noticed one day rereading some lyrics. Something didn't sound like me. The twistiness was missing, the line that runs long because it has something to say, the odd title. If I write «Mitad Venezuela, mitad Pakistán» (Half Venezuela, half Pakistan), it's because I fancy a title nobody expects, not because it's convenient to produce.

Since then I've had a rule: the lyric comes first, exactly as I want it, and then we'll see how the rest gets solved. If the tool struggles with a line, I solve the problem by directing better, not by cutting what I wanted to say. I don't always manage. But I no longer write to please a program.

Another stumble, a quieter one. When a tool lets you do a lot, it's easy to believe you're making progress because you're doing a lot. I fell for it. For a while I measured my days in finished songs, like someone counting steps on their phone.

Quantity has its value, I won't deny it. Releasing a lot has taught me more than waiting for the perfect album would have. But there were moments when I realised I was producing on autopilot. Competent songs I didn't much care about. I learnt to stop and ask whether anyone needed this, starting with me.

The warning sign was very simple, and I was slow to heed it. When I finished a song and didn't feel like showing it to anyone, something was off. With the ones I care about, the opposite happens: I can hardly wait for them to come out. That impatience is a better thermometer than any counter of finished songs.

Not knowing how to explain it

One stumble had nothing to do with music. It was the way I talked about it.

At first, when someone asked how I made the songs, I'd get in a tangle. I'd say things by halves, or too technically, or too vaguely. And misunderstandings came out of that: people who thought the vocals were mine, or people who thought the machine did everything and I just pressed the button. Neither was true.

It took me a while to find a simple way. Now I have one: I write every lyric, I produce the music with Suno, directing it myself, I decide what stays and what goes, and I direct the music videos. The vocals are made with AI, not by me. I introduce myself as the first AI songwriter because it's a short label that says how I work, even if it needs explaining.

I don't regret any of these stumbles. Well, maybe writing for the machine, which lasted longer than I'd like to admit. But they all taught me something I wouldn't have learnt from reading.

The funny thing is, I'm still stumbling. Over other, smaller things. The tools change, I change, and every change brings its own new clumsiness. I've made my peace with it. I suppose the day I stop stumbling will be the day I've stopped trying things, and that really would worry me.

If you want to judge the result of all that stumbling, the songs are at mikesaintlouis.com.