Ask the producer of a 99-seat show or a festival short what disappears first when the budget tightens, and the answer is rarely the one audiences would predict. Costumes survive. Lighting survives. Original music is where the money quietly stops.

This is not new. What has changed over the past two years is the arithmetic underneath it.

The old math

A composer scoring a one-act play is quoting real hours: sitting with the script, writing cues that land on specific beats, often coordinating live players. The number that comes back is defensible. It is also, for a great many independent productions, larger than the entire design package.

So the field leans on production music libraries. Subscription catalogs solved the licensing problem and created an aesthetic one in its place. Anyone who has sat through three shorts in a row scored from the same catalog knows the symptom: the music is competent, cleared, and completely anonymous. It does not embarrass the production. It also does not belong to it.

What actually shifted

Generative music tools arrived in exactly this gap, and the first wave earned the skepticism it got. The output was structurally vague, full of loops that never resolved and vocal lines that fell apart by the second verse.

The current generation differs in one way that matters to producers rather than to technologists: these tools now return finished songs, with a verse, a chorus, and an actual ending, rather than atmospheric beds. That distinction is the whole ballgame for anyone who needs a diegetic song — something a character sings onstage, something playing on a radio in act two, a fictional pop single the plot requires to be plausible. A director facing that need used to choose between hiring a songwriter and cutting the idea. A tool like Lacuna’s AI song generator will take a written premise and return a complete track with vocals, and will separate the stems on request, which is the part a sound designer actually cares about: a vocal that cannot be ducked under dialogue is not usable.

The licensing question is the real one

This is where most of these platforms have failed the people who read this publication, and it deserves far more scrutiny than the audio quality does. A track you cannot clear is a track you cannot use, and “generated by software” is not by itself a rights position.

Anyone evaluating one of these services should read the commercial terms before listening to a single demo. Does the license cover festival exhibition? Ticketed performance? Eventual distribution, if the short gets picked up? Does it survive a lapsed subscription, or does music you have already shipped revert to something unlicensed? These are unglamorous questions, and they are the only ones that determine whether the tool is usable at all.

Where adoption is actually happening

Less discussed, and far more common in practice, is promotional work. A company mounting a new musical needs social assets on a schedule no editor is being paid to meet. Automated lyric video tools that take an existing recording and its lyrics and return a timed, captioned video remove a step that used to consume an evening. It is drudge work, which is precisely why software absorbed it first.

What it does not do

It does not do dramaturgy. A generated cue does not know that act two needs to stay unresolved, or that the reprise has to land differently than the first statement of the melody because the character has changed in between. A composer reads a script. A model reads a prompt, and the distance between those two things is the entire craft.

The honest framing is not replacement. It is that productions which were never going to commission original music, because the money was never there, now have something that sits between a library cue and a composer. Whether that raises the floor or lowers the ceiling is a question this field will spend the next several seasons arguing about, probably in these pages.