What AI can't do in my songs: irony, camp and double meaning

There's a question people ask me with the look of someone about to catch me out: if AI makes the music, what stops it from making the jokes too? The short answer is that wit isn't manufactured, it's committed, like a small crime. The long answer is this piece, and I'll try to give it without playing the martyr or the genius, the two roles that suit me worst.

I write the lyrics to all my songs. For the music and the vocal performance I work with Suno, a tool I direct: I pick, I discard, I ask again. I've talked about that division of labour before. Today I want to zoom in on something smaller and more specific: the line with a sly edge to it. The one that says one thing and means another. The one that's only funny if you know where you are, or the one that hurts precisely because it sounds like nonsense.

That's where the machine hasn't got a clue. And I'm not saying it to look good.

Wit is an act, not an ingredient

The tool I use has processed an enormous number of songs and, by extension, an enormous number of jokes. It knows what shape they tend to take. It knows that an unexpected rhyme gets a smile, that a list which swerves at the end works, that putting the solemn next to the domestic is good fun. It knows all that the way you know the recipe for a dish you've never tasted.

But irony isn't a shape. It's a relationship between what you say, what you leave out and what the other person already knows. For it to work you need someone who knows something and someone who shares it. A wink with nobody on the receiving end is just a twitch.

That's why, when you ask a machine to make something "funny", what comes back tends to be nice. Nice in the worst sense: the office birthday card, the stand-up routine that offends no one because it isn't aimed at anything. The humour I care about takes aim. At an ex, at a prejudice, at myself. And aiming means knowing where everything is.

Camp has an extra catch. It's exaggeration that knows it's exaggerating. The machine exaggerates beautifully when asked, but it doesn't know it's doing it, and that awareness is exactly what separates camp from mere excess. A wig isn't camp. A wig put on with intent, at the exact moment and in front of the exact audience, is.

Mercadona isn't in the textbooks

Let's take an example with a name attached. In «El productor de Madonna no compra en Mercadona» ("Madonna's producer doesn't shop at Mercadona") there's a moment where the ex, who adored Madonna, gets portrayed by comparison. The queen of pop turns into a deodorant and the heartbreak gets settled in a supermarket aisle:

pues mira bonito me he cambiado a la marca blanca de Mercadona y también funciona

Roughly: "well, look here, sweetheart, I've switched to Mercadona's own brand, and it works just as well." Mercadona is the supermarket chain where a huge slice of Spain does its weekly shop.

I'm not going to explain the whole joke, because an explained joke is a dissected frog. But I do want to point out where it lives. It lives in «mira bonito», which in Spain is said with a very precise little tone, somewhere between condescension and a slammed door, that anyone who's ever argued in a kitchen recognises instantly. It lives in the own-brand label, which here isn't just a cheap product: it's a small domestic philosophy, that of someone who's no longer prepared to pay for the packaging. And it lives in «también funciona», "it works just as well", which is the knife. It doesn't say the new one is better. It says it makes no difference. That I've replaced you without any fuss.

That's the kind of bite I'm interested in. Not the insult, which is easy, but the well-placed indifference.

If that line is written by someone who has never done the shopping here or heard «mira, bonito» delivered with raised eyebrows, what you get is a correct sentence about switching products. Grammar perfect, wit missing, whereabouts unknown.

The line is chopped into very short lines, almost in jerks, and every break is a pause before the punch. When the tool turns it into music, it sometimes respects those pauses and sometimes flattens them, so I listen through versions until I find the one that lets the joke breathe. If «también funciona» arrives in a rush, the joke dies. If it arrives with half a second of air, people laugh. That difference is my call, because I know where the laugh is. The machine only knows where the syllables are.

Double meaning needs someone looking sideways

The double entendre is irony's mischievous cousin. It works because one sentence carries two readings and the listener chooses, or better still, doesn't choose and keeps both.

In «El anillo de casado» ("The wedding ring") the chorus ends with a line that still gets a grin out of me:

Si pasa por mi cama y lo disfruto para mí es pasado

The game is between «pasa» and «pasado». The verb first means physical movement, "if he passes through my bed", someone crossing your bed the way you'd cross a doorway, and then it turns into time: "to me, he's in the past", something that no longer counts. It's a cocky line, the line of someone playing hard. And it's a lie, of course. A few lines later the narrator admits he's starting to fall in love. The swagger is the costume, and the costume falls off by itself.

Here the double meaning isn't decoration, it's character: this man plays with words because he needs not to say things straight. A machine can churn out puns by the dozen. What it can't know is why someone needs them. Why a guy sleeping with a married man would rather rhyme than confess.

On top of that, sexual double meaning in Spanish has a huge tradition, from saucy copla to variety revue to the after-dinner chat at any wedding. That tradition is learned by listening to specific people. The tool has read it. It isn't the same. Anyone who has heard it knows the joke lies in the moment someone lowers their voice.

Tenderness dressed up as a joke

If bite is hard, tenderness is harder. Because real tenderness, in a funny song, almost never shows up as tenderness. It hides.

In «Un invitado VIP cortando cebolla» ("A VIP guest chopping onions") there's a very bad night told without any filter, with drink, with a sauna, with the feeling of being the only idiot at the party. And at the end, when someone might see him crying in the street, the narrator offers an excuse:

es que me acabo de hacer una ensalada y he estado cortando cebolla

"It's only that I've just made a salad, and I was chopping onions." I've written few things as tender, and it's a joke: the lie you tell so nobody worries, the dignity of someone who doesn't want to be pitied. We all know that excuse in one of its thousand versions: it's hay fever, it's the wind, I've got something in my eye. What the song does is push it into the absurd, a salad at an ungodly hour to explain sobbing in the street, and the tenderness lives in that absurdity. Because nobody is going to believe it, and he knows it, and he says it anyway.

This, for me, is the heart of why all this belongs to the lyricist. A line doesn't hurt because of the words in it, but because of what it's covering up. To write something that covers a wound, you have to know the wound exists and what shape it is. The machine can write "I'm sad" in a hundred elegant ways. It can't hide sadness behind an onion, because it has nothing to hide.

Then there's how many times to repeat it. In the song the excuse returns several times, and with every lap the lie is more obvious and more tender. Said once, it would be a gag. Repeated, it's a person trying to convince himself. That shift from joke to portrait isn't in any prompt. It comes from having heard someone repeat an excuse until it cracks.

The honest split: what it does and what I do

This isn't meant as an indictment of the tool, which would be unfair and untrue. So here's the split, on the table.

What the AI does, and does very well:

• It proposes melodies and arrangements for lyrics that already exist. Sometimes it surprises me with a turn I'd never have imagined, one that gives a line a new meaning.
• It provides the vocal performance. The timbre, the emotion in the singing, lots of nuances I didn't invent note by note.
• It gives me range. I can hear the same lyric dressed five different ways and find out which one fits.
• It recognises styles. If I ask for a copla feel, a village-fiesta feel or a dancefloor feel, it knows where to head.

What I do, and can't delegate:

• I write every word. The jokes, the double meanings, the local references, the proper names, the absurd excuses.
• I decide who each joke is aimed at. If a line punches down instead of up, it goes.
• I know where the laugh is and where the wound is, and I choose the versions that respect both. If the music stamps on a punchline, that version is out, however pretty it is.
• I discard. Whole songs that worked technically have stayed in the drawer because the humour didn't land or the tenderness came across as soppy.
• I supply the place. Madrid, its supermarkets, its tones of voice, its excuses. I write for Spain and Latin America, and I know that the specific, however local, travels better than the generic.

There's a grey area I think it's important to own up to. Sometimes the music changes the meaning of a line without my having looked for it. A very sad melody under a joke turns it into something else, more bitter. Very cheerful music under a hard lyric makes it almost unbearable, in the good sense. When that happens, the discovery is shared. I recognise it and let it through, but I didn't entirely cause it. It feels honest to say so, even if it takes a little of the credit away from me.

Why I don't ask the machine for jokes

I don't hand my lyrics over to any machine, and it isn't only pride. When you ask a text AI for funny lines, what usually comes back has two recurring flaws. The first is that it's harmless: jokes that bother nobody because they're about nobody. The second is that it's from nowhere: a neutral, pasteurised humour that could be from any country and is therefore from none.

Queer humour, besides, has a history of its own that the machine hasn't lived. For a long time, laughing was a way for us to survive hostile rooms, to say out loud what couldn't be said without a safety net, to recognise each other with a raised eyebrow. That history is in my lyrics even when I don't name it. It's in the «mira bonito», in the lover playing it cool, in the guy who cries and blames an onion. The tool can imitate the tone, but tone without the history is a carnival costume bought at the last minute. It sparkles, but it won't survive a dance.

And there's a question of responsibility. If a joke of mine hurts someone, it's mine. I can't blame an algorithm. Writing every sly line myself is also a way of owning what I say. I'd rather get it wrong with my own jokes than get it right with nobody's.

What's left when the song ends

In the end, what people remember about a funny song is almost never the tune of the joke. It's the joke. It's the phrase that gets repeated in a group of friends, the one someone drops at dinner to make someone else laugh, the one that becomes a password between two people who heard it together. I write that phrase, sitting down, crossing out, reading it under my breath to see if it breathes.

There are nights when a line makes me laugh alone at home and the next day it seems silly. There are others when something I wrote as a gag tightens my stomach when I reread it, and then I know there's a song there. That thermometer can't be outsourced. It's what tells a joke apart from a line that hurts, and sometimes they're the same line.

AI has given me an orchestra I could never have paid for and a vocal performance I couldn't provide myself. I'm genuinely grateful. But irony, camp, double meaning, the neighbourhood joke, the bite and the tenderness are still my job. And I hope they stay that way, because they're the most fun part of this whole craft.

The full lyrics, with all their jokes and the odd onion, are on mikesaintlouis.com, and the songs are on Spotify and Deezer or free to download from the site.