The Authorial Architecture

One author. Multiple identities. Synthetic embodiment. Orchestrated production.

For a long time, I resisted calling what I make “AI music”.

I still do.

The phrase tells you something about the machinery of production and remarkably little about the authorship. It names a class of tools and then encourages the tools to become the explanation for the work.

What has changed is not my view of the label. It is my view of what should be visible.

Generative systems are materially involved in some of my music. They can provide voices, instrumentation, performances and production elements that I do not literally execute myself. That is not incidental to the process, and I no longer think it should sit somewhere offstage in the account of how the work is made.

But nor does making that contribution visible require pretending that the machinery is the author.

The first three essays in this series have been approaching this problem from different directions. Authorship has moved. The human signal persists through judgement. And, once several musical identities have developed enough coherence, the ecosystem connecting them can itself become something I compose with.

That leaves a more basic question.

Who, exactly, is making this work?

The answer begins with the aliases.

I do not experience MIRA KAI, Sable Rue or the other musical identities as fictional artists that I happen to manage from somewhere behind the curtain. They are closer to bounded regions of my own musical imagination.

Each permits something different.

One may hold a particular emotional temperature, another a different relationship to rhythm, another a vocal attitude or sonic density that would feel false elsewhere. Over time those conditions become recognisable enough that I can hear when something belongs.

More importantly, I can hear when it does not.

A model can generate possibilities. I recognise identity.

That distinction has become increasingly important to me.

There are plenty of outputs that are technically competent, sometimes even impressive, but are simply not MIRA. Or not Sable. Or not right for a particular album, regardless of how good they might be in isolation.

The identity therefore does not begin with the successful generation. It already exists as a set of constraints, permissions, memories and expectations against which the generation is judged.

The aliases do not conceal the author. They reveal parts of him.

Sable Rue makes this particularly clear.

The project did not begin with a lucky generated track followed by the invention of a name to explain it. Before an individual recording could earn its place, there was already a differentiated world around the project: a world bible, production laws, musical doctrine, an album blueprint and working systems for deciding what Sable Rue could and could not become.

The documents are not the art, obviously. Their value is that they make an internal creative model explicit enough to work against.

They establish limits.

They help answer questions such as: is this too polished? Too theatrical? Too dense? Is the voice pushing when it should be withholding? Has darkness become generic rather than specific? Does this belong in this world, or is the technology merely showing me something it can do?

Generation happens inside those conditions.

Rejection does too.

The audible voice or instrumentation that finally emerges may be substantially generated. I am not claiming to have stood at a microphone and sung the released vocal, or to have physically performed each instrumental part. But neither did the generative system arrive in an empty room and invent the project for me.

It entered a world that was already taking shape.

I have come to think of this as synthetic embodiment.

An internal musical identity can acquire an external voice, instrumentation and performance without my having to embody those things literally myself. The embodiment may be synthetic. The identity it is being asked to embody is not therefore reducible to the system producing the sound.

That does not make every output equally authored.

Quite the opposite.

Generative output is abundant. Admission to the canon is scarce.

The work sits increasingly in the difference between those two conditions.

MIRA KAI offers another version of the same problem, spread across a longer period.

Across three releases, MIRA has developed enough continuity that the identity is no longer recreated from scratch each time. There is a persistent sonic and emotional register, recurring material, an evolving relationship between body, voice and motion, and a growing history against which new work is judged.

Again, persistence is not automatic.

A model does not preserve an artistic identity simply because I type the same name into it. Continuity has to be recognised and maintained through repeated decisions about what stays, what changes, what is rejected, how one record relates to the previous one, and what kind of departure still feels like growth rather than drift.

That is why judgement matters so much to this way of working.

A generated performance becomes part of the body of work only after surviving a sequence of questions.

Does it belong to this identity?

Does it belong to this record?

Does it strengthen what already exists or merely repeat it?

Is the variation meaningful?

What should disappear?

What needs to remain unresolved?

And, sometimes, what needs to be withheld now so that something arriving later has more force?

Those decisions accumulate.

Eventually they create history.

The previous essay, The Ecosystem as Instrument, explored what happens when these identities become relational as well as distinct. On Midnight Congregation, MIRA KAI and Sable Rue enter a THISISGRAEME record rather than remaining sealed inside their own catalogues. Their appearance, reappearance and timing begin to operate at a level above the individual track.

That is the orchestration layer.

The creative object is no longer only the generated performance, or even the finished song. It also includes the relation between works: which identity carries the frame, which crosses into another project, which signal appears before it has a standalone home, how an album is sequenced, how its sonic field is shaped after generation, and what ultimately becomes part of the released catalogue.

None of this requires me to pretend that authorship and performance are the same thing.

They are not.

I can take responsibility for the conception, constraints, direction, iteration, rejection, selection, sequencing and final form of a project without claiming that I personally sang or played a generated performance. Ownership and copyright introduce different questions again, and I do not need to collapse them into authorship in order to describe the creative process accurately.

That distinction is useful because the conversation around generative music has often been forced towards two equally unhelpful extremes.

At one pole: the machine made it.

At the other: the human maker feels obliged to imply that every audible element passed directly through their own hands.

Neither describes what I am doing.

The making is distributed.

I establish worlds and identities. I set conditions. I direct systems towards particular territory. I generate and regenerate. I compare. I reject. I recognise. I sequence. I master where the project requires it. I decide what becomes public and what disappears.

Generative systems materially realise possibilities inside that process.

Conventional production tools do other parts of the work.

And the finished recordings emerge from the interaction between them.

For some time I was cautious about explaining this publicly because “AI music” was such a crude category. It was too easy for the production method to swallow the work itself: suddenly the music became evidence in a culture-war argument, a novelty demonstration or a debate about whether pressing a button counted as creativity.

None of those frames was particularly useful for making better records.

But opacity has its own limit.

The surrounding environment is moving towards finer distinctions about provenance and contribution. That seems reasonable to me. If a voice or performance is materially generated, there is no need to imply otherwise.

Transparency, though, should work in both directions.

The synthetic contribution should be visible.

So should the human source behind the identities, the constraints, the selections and the body of work.

That is a much more interesting proposition than attaching a single label to an entire piece of music.

It means being able to say: this performance was generated; this identity is an authored musical project; these decisions belong to me; this relationship between works was deliberate; this is what I kept; this is what I rejected; this is the record that finally emerged.

Not confession.

Not camouflage.

Provenance.

I still do not think “AI music” adequately describes what I make. But I increasingly think you should be able to see how I make it.

As production becomes more distributed, authorship does not disappear. Its location simply becomes easier to misunderstand.

In this work, it resides increasingly in the architecture: the worlds established, the identities held, the constraints chosen, the directions given, the possibilities rejected, the relationships composed, and the decisions about what finally belongs.

One author. Multiple identities. Synthetic embodiment. Orchestrated production.

The making is distributed.

The authorship is architectural.