What Skills Should Artists Value in the Age of AI Music?

AI music is often described as a shortcut. For Brian Myk of Rocket Horse, that description misses the more interesting change.

Rocket Horse existed before generative AI entered the production workflow. Brian has made music through traditional studio processes, where trying another arrangement could mean more performers, more studio time, more revisions and more money. Today, he uses AI-assisted production while keeping the songwriting, lyrics, melodies, hooks, concepts and overall creative direction his own.

So when we asked him what had actually changed, his answer was not that making music had suddenly become effortless.

“The biggest change is that the bottleneck moved from logistics to judgment.”

That distinction runs through almost everything Brian told us. AI has made it much easier for him to hear different possible versions of a song. What it has not done is decide which of those versions deserves to become the record.

In Lunar Boom’s earlier analysis of AI music and the new burden of choice, we argued that abundance can create its own creative pressure. Brian’s workflow offers a practical example of what that looks like.

“The bottleneck moved from logistics to judgment”

We asked Brian to compare the older Rocket Horse production process with the way he works now.

Traditionally, hearing a song in a substantially different form could be expensive. A heavier rock arrangement, an acoustic interpretation or a polished pop version might require new recording sessions, new performers, new arrangements and another round of revisions.

That friction naturally limited experimentation.

With AI-assisted production, Brian can test those possibilities much faster.

That does not mean every possibility deserves to survive.

For Brian, one of the biggest changes is that experimentation has become cheap enough that he can afford to explore ideas he might previously have abandoned. But he is also clear that most experiments should remain experiments.

That changes the creative problem.

Previously, an artist might ask:

Is this alternative arrangement worth the time and money required to hear it?

Now the question can become:

I have already heard ten versions. Which one is actually worth releasing?

Imagine being offered 100 ice cream flavours. The difficulty is no longer finding one that tastes good. Many of them might taste good. The difficulty is knowing which one you actually prefer, which one fits the moment and, for a commercial artist, which one is most likely to connect with the audience you are trying to reach.

That is close to the pressure Brian is describing.

If an artist can produce ten, fifty or one hundred plausible interpretations of the same composition, generating one more version becomes less valuable. Knowing when to stop becomes more valuable.

How do you decide which version wins?

That led to another question: when several versions work, what makes Brian choose one over another?

His answer comes back to the song itself.

A version is stronger when it strengthens the hook, clarifies the emotion and makes the song more memorable. Different is not automatically better.

A technically impressive production can still lose.

A larger arrangement might bury the chorus. A glossy mix might make a lyric feel less intimate. An unexpected genre transformation can be interesting for thirty seconds while taking the song further away from what originally made it compelling.

This is where AI can expose taste unusually clearly.

When high production quality becomes easier to approximate, an artist still has to decide whether the bigger drums actually improve the track, whether the new guitar arrangement serves the vocal and whether the surprising version is meaningful or merely novel.

Brian’s approach is therefore highly selective. He does not treat an output as the answer simply because the model produced something polished.

His traditional studio experience appears important here. He has spent years making decisions when arrangement changes carried real practical consequences. AI removes much of that friction, but it does not remove the need to recognise when an arrangement works.

The outputs are, in his words, “material to judge.”

That may be one of the most useful ways of understanding the role of AI in this kind of workflow.

The generation is not necessarily the finished creative act. It creates material. The artist still has to interrogate it.

Has AI made musical knowledge less important?

One assumption around generative music is that easier production should reduce the importance of traditional musical experience.

Brian’s workflow suggests almost the opposite.

When there were fewer practical options, the production process itself filtered what was possible. Once those restrictions weaken, the artist has to provide more of the filtering personally.

Experience can help identify when a chorus has become buried, when an arrangement is distracting from the lyric, when a vocal delivery feels emotionally false or when a technically excellent production simply does not sound like Rocket Horse.

The number of possible roads has increased.

That makes the ability to choose a road more important.

Brian sees production increasingly as “a field of possible roads.” The artist can explore further before committing to one.

But that raised another question for us: if production can change so dramatically, what stays constant?

“What did not change is the part I care about most”

For Brian, the answer is songwriting.

“What did not change is the part I care about most: the song itself.”

He continued:

“The lyrics, melodies, hooks, structure, concept and overall creative direction still have to be there. AI can give me more ways to dress a song, but it cannot make me care about a song that was not strong to begin with.”

This is an important distinction.

AI has changed how Brian thinks about production much more than how he thinks about writing songs.

The production can now branch outward into many possible arrangements. The song remains the anchor against which those arrangements are judged.

That also creates a warning for AI musicians.

Unlimited production options cannot necessarily rescue a weak lyric, melody or hook. In some cases, abundance may make those weaknesses easier to expose. The same underwritten chorus can be rendered as rock, synth-pop, acoustic folk or orchestral music and still fail to stay in the listener’s memory.

The ability to regenerate is therefore not the same thing as the ability to improve.

Does easier experimentation actually make creativity easier?

Only partly.

Brian can now hear ideas that would previously have been too inconvenient or expensive to test. That is a genuine expansion of creative freedom.

But experimentation without a standard can become its own trap.

If every generation suggests another possible direction, an artist can spend indefinitely moving sideways.

Version 12 is different from version 11.

Version 13 has a better guitar tone.

Version 14 has a more dramatic chorus.

Version 15 suddenly turns the song into something else entirely.

At some point, someone has to decide whether any of those changes actually make the song better.

That is why rejection becomes important.

Traditional production created its own sunk-cost problem. Once musicians, studio time and money had already been committed to an arrangement, abandoning it could be painful.

AI creates almost the reverse problem.

When versions are inexpensive to produce, artists can become reluctant to discard them because there is always something interesting in each one.

Brian’s process depends on being willing to say no.

An output can be technically good and still not belong on the record.

What about the audience?

Another part of that judgment is understanding who the song is actually for.

Rocket Horse does not need every possible version of a track to appeal to every possible listener. Brian is making decisions within an existing artist identity.

That matters because AI makes genre switching unusually easy.

A song can suddenly become country, hard rock, electronic pop or acoustic folk without requiring the practical effort those transformations once demanded.

But the fact that a transformation is possible does not mean it is appropriate.

A version can be impressive in isolation and still be wrong for the artist releasing it.

That makes audience understanding another part of the creative filter. The artist has to consider whether the production works for the song, whether it fits the identity surrounding it and whether it speaks to the listeners who are most likely to care.

Again, the scarce resource becomes judgment.

What happens when that judgment disappears?

Brian’s workflow is also useful because it contrasts with one of the largest concerns surrounding generative music: what happens when generation scales but curation does not?

The wider music industry is already wrestling with systems where creation and distribution can expand faster than quality control.

The IFPI’s Streaming Integrity Initiative asks distributors to adopt anti-fraud practices around rights verification and streaming manipulation, according to Music Business Worldwide. The same report noted that DistroKid, which says it distributes roughly 40 percent of all new music, had not signed the initiative at the time of publication.

A separate Music Business Worldwide report covered Universal Music Group’s lawsuit against DistroKid, including allegations concerning AI-generated content and unlawful practices. Those remain contested legal claims rather than settled facts.

Still, the dispute reflects the broader concern.

If generation, automation and distribution are paired with weak judgment, weak rights checks or weak curation, the result can be huge quantities of material that listeners, artists and rights holders struggle to trust.

That is not the workflow Brian describes.

His use of AI sits inside an artist-led process where outputs are tested, rejected, reshaped and ultimately judged against a song that existed before the production option appeared.

The distinction matters.

The future of AI music will not be determined only by what models can generate. It will also depend on who is choosing between those generations and what standards they use when they choose.

Consent and artist control are becoming part of that conversation as well. The Verge reported that Pippa, an AI platform built around licensed artist participation, had around 800 paying subscribers and agreements with four artists, with more planned, while allowing participating artists to use pseudonyms if they feared backlash The Verge.

That points toward another form of judgment: not just deciding which version of a song sounds best, but deciding how AI should be incorporated into an artist’s work at all.

The difficult part did not disappear

Perhaps the most interesting thing about Brian’s answers is that they complicate the idea of AI as simply a labour-saving device.

Some labour clearly disappears.

It is easier to test ideas. It is cheaper to hear alternatives. Production directions that once required substantial coordination can now be explored quickly.

But another form of work expands in its place.

The artist has to compare.

Reject.

Recognise.

Experiment without becoming distracted.

Understand the audience.

Protect the identity of the song.

And eventually stop generating and release something.

For Brian, the song remains the stable point underneath all of those possibilities.

AI can offer another arrangement.

Then another.

Then another.

But it cannot remove the moment when the artist has to listen to them and decide:

This is the one.

If everyone can generate 100 good options, the valuable skill may no longer be creating option 101.

It may be knowing which one people will actually remember.