Every demo you have seen of AI music is the best of about forty attempts. I know because that is the ratio I work at, and I am good at this now. Nobody shows you the thirty-nine, which is why the public conversation about this technology is so strange: half the people think it is already finished, and the other half heard one bad output in 2024 and stopped paying attention.
The truth is more useful and more boring. It is extraordinary at four things and unreliable at four others, and knowing which is which is most of the skill.
What it genuinely does better
Arrangement, instantly
Ask a competent session player to give you the same song as a qawwali, then a stripped acoustic ballad, then something with a dhol under it built for a dancefloor. That is three sessions, three days, three fees. The model does it in four minutes. Not better than the musician. But three times, immediately, so you can hear which one the song actually wanted to be.
This is the thing that changed my writing more than anything else. I used to commit to an arrangement because committing was expensive. Now I audition the song against itself.
Pastiche
It is remarkable at the texture of a genre. Nineties Bollywood playback, with that particular reverb on the vocal and the strings sitting slightly too loud. Late-night acoustic with the guitar close-mic'd and fret noise left in. These are not easy to fake and it does them convincingly.
The second language problem
Finding a singer who can carry Urdu and English in the same verse, in the right register, on a Tuesday, is genuinely hard. This is where I have been most surprised. It is not perfect and I will come to that, but the availability alone changes what is possible.
Volume of drafts
Writing has always been rewriting, and the cost of a draft has always been the limit on how much rewriting you do. That cost is now roughly zero. You still need taste to pick, but you get to pick from forty rather than from two.
Where it reliably falls apart
It cannot hold a specific detail
This is the big one and it is why Sureela has a person in the middle of it. Ask a model for a love song and you get moonlight, rain, forever, eyes. Beautiful, generic, worthless. The thing that makes a song about two actual people work is that he calls her Doctor Sahiba because she corrects everybody, or that they met at a bus stop in Pune in the rain in 1994.
Models smooth those away. They pull towards the average of everything they have seen, and the average of every love song is not about anyone. Getting a specific detail to survive into the final lyric takes deliberate, repeated, human intervention.
Pronunciation, in every language but English
Urdu and Hindi come out with an accent, and not a charming one. Certain consonants are consistently wrong. You learn to spell phonetically against the model rather than correctly, which feels like cheating and works.
Structure over three minutes
A song needs to go somewhere. A bridge has to earn the final chorus by being different in a way that makes the return feel like a return. Models are weak at this. They will give you a bridge that is just another verse with different words, and it sounds fine, and the song does not land, and it takes a while to work out why.
Knowing when it is finished
It will never tell you the take is not good enough. It has no opinion. Every output arrives with identical confidence, which means the entire burden of quality control sits with whoever is listening.
What this means if you are buying one
There are services now that will sell you a personalised song for nine dollars, generated the moment you press the button. They are not lying about what they do. You type in some names, a model writes a lyric containing those names, and you get an MP3 in ninety seconds.
It will contain the names. It will not contain the person. The difference is invisible in a demo and extremely visible at a wedding, in a room full of people who know the couple, when the song mentions moonlight instead of the thing everybody remembers.
Generation is cheap and getting cheaper. Judgement is the only part that was ever scarce.
That is the bet we have made. Sureela uses the same tools everybody else has. What we sell is the forty attempts you did not have to listen to, and a person who read your story before writing a word of it.