Generative tools produce material by modelling patterns in what they were trained on. That single sentence explains both what they are good for and where they fail, and it is more useful than any argument about whether they are creative.
Plausible is not the same as correct
A system of this kind optimises for output that resembles its training material. It does not check anything. What comes out therefore looks right in the way that a fluent sentence looks right, whether or not the content behind it holds.
In music preparation this has a specific and expensive form. A generated orchestration can look entirely idiomatic on the page and be unplayable: outside an instrument's range, without a breath, with a bow change that cannot happen, with divisi that needs more players than are contracted, written for a transposing instrument at concert pitch. None of that is visible until the first service, which is the most expensive place to find it.
The rule that actually decides each case
A generated draft is worth having when checking it costs less than doing the work.
Transcribing a recorded coaching session passes easily: the draft is mostly right and errors are obvious. A translation draft passes if somebody who knows the language reads it. An orchestration fails, because verifying playability bar by bar is most of the work of writing it. A claim of fact fails completely, because checking it is the entire job and the draft has contributed nothing but a starting point that may be wrong.
Where they are straightforwardly useful
Administrative and preparatory work, mostly. Rough transcription. Searching a large archive of programmes or recordings for a phrase. First-pass translation for the title team to correct against the score. Reference images for a design conversation. Reformatting paperwork between departments. Summarising a long document that somebody will then read.
The common feature is that a human with the relevant competence sees the output before it matters, and that the failure mode is visible rather than silent.
Rights are unsettled and jurisdictional
Two separate questions arise and they are frequently confused. What was the system trained on, and who owns what it produces.
On the second, the position in the United States is that copyright protects human authorship, and the Copyright Office has published guidance addressing works that contain material generated by such systems, including the expectation that the generated contribution be disclosed when registration is sought. Positions differ elsewhere and are changing. This is a matter for a lawyer, not for a technical publication, and any production intending to rely on generated material in a work it will license should establish the position before rather than after.
The training question is being litigated and we take no view on it. What can be said without controversy is that a company cannot warrant what it does not know, and that the provenance of most such systems is not disclosed in a form that permits a warranty.
Disclosure is becoming a condition, not a courtesy
Competitions, commissioning bodies and some festivals now ask directly whether generated material was used and in what proportion, and a declaration made afterwards is worth less than one made in the application. Publishers ask the same question of new work.
The practical difficulty is that the honest answer is rarely binary. A libretto drafted by a person and tidied by a tool, an orchestration written by hand from a generated sketch, a title script machine-drafted and rewritten line by line: each is a different degree of involvement, and a yes or no does not describe any of them. Keeping a short record of what was used where, at the time rather than reconstructed later, is what makes an accurate answer possible at all.
Voices and likenesses need consent
Synthesising a performer's voice, or their appearance, is a different matter from generating a text draft, and it engages personal rights that belong to that individual rather than to the production. Collective agreements increasingly address it directly.
The position that ages well is to obtain explicit, specific, written consent for a named use, rather than a general permission buried in an engagement, and to treat a refusal as final.
The question this publication keeps asking
Does the tool take from the performer what the performer does better? For scheduling, transcription and paperwork the answer is plainly no, and the time returned is time available for the work. For interpretation the answer is usually yes, because interpretation is the thing the audience came for, and a plausible version of it is not a cheaper version of it but a different and lesser object.
What we cannot verify
Capability claims for these systems come from the companies selling them, are measured on benchmarks that do not resemble music preparation or staging, and change between versions. We reproduce none of them. The legal position varies by jurisdiction and is moving; nothing here is legal advice. And a publication writing on this subject has an interest of its own in appearing rigorous about it, which is precisely the kind of interest we ask other sources to declare.
The short version
- These systems model patterns and check nothing, so output is plausible rather than correct.
- A generated orchestration can look idiomatic and be unplayable until the first service.
- A draft is worth having when checking it costs less than doing the work.
- Transcription and rough translation pass that test; claims of fact fail it entirely.
- Ownership of generated material is jurisdictional, unsettled, and a question for a lawyer.
- Voice and likeness need explicit written consent for a named use.