Interlude. Why AI Sometimes Writes a Beautiful Chorus and an Utterly Idiotic Verse

Anyone who has listened to a dozen songs out of Suno has noticed the same thing. The chorus can be very good: a catchy melody, the right lift, a word you want to sing along to by the second time. And the verse before it sounds as if it were written by a different, tired person: lines about nothing, a melody treading water, a forced rhyme. This unevenness has clear reasons.

The model composes by predicting the next small piece from the previous ones, again and again. Each of its decisions is plausible over the nearest bars, and over a stretch of a few seconds it hardly makes mistakes. A good song, though, needs a plan three minutes ahead: the verse has to prepare the chorus, the second verse has to say something new after the first, the bridge has to lead off to the side so that the last chorus sounds like a homecoming. The model builds no such plan. It is like a storyteller who has a perfect feel for every sentence but doesn’t know how the story ends.