🐰
🐰🕳️ AI RABBIT HOLES
WHAT IF AI DOESN’T HEAR A NAME THE WAY WE DO?
A few innocent Suno experiments opened a surprisingly strange door
It started simply.
Type a musical name into Suno.
See what happens.
Nothing scientific.
No elaborate benchmark.
No white coats.
Just curiosity.
We tried:
George Gershwin.
Accepted.
Then:
Cole Porter.
Accepted.
Hoagy Carmichael.
Accepted.
Duke Ellington.
Accepted.
Irving Berlin.
Accepted.
Jerome Kern.
Accepted.
Richard Rodgers.
Accepted.
Lorenz Hart.
Accepted.
Fats Waller.
Accepted.
And suddenly the experiment became more interesting.
Because the day before, a modern band name had been rejected.
So what exactly was the machine doing?
🐰🕳️
THE NAME MAY NOT BE THE DESTINATION
At first glance, you might imagine that typing:
George Gershwin
means:
Make something that sounds like George Gershwin.
But that does not appear to be what happened.
The results were not convincing replicas.
The lyrics certainly were not.
Those wandered happily back into familiar AI territory:
generic romance
predictable rhymes
obligatory emotional uplift
and the occasional line apparently written by a refrigerator magnet experiencing its first breakup.
😹
But the music sometimes shifted.
Harmony changed.
Instrumentation changed.
Rhythm changed.
The musical posture changed.
Something was happening.
THEN WE TRIED CLAIR DE LUNE
The first result sounded almost nothing like Clair de lune.
Interesting.
Then we tried:
“like Clair de lune”
And Suno translated the idea into something closer to:
impressionist piano, more modern
There was the clue.
Perhaps the AI was not treating the title as a musical blueprint.
Perhaps it was translating the reference into descriptive characteristics.
Named thing in.
Musical vocabulary out.
THAT CHANGES THE QUESTION
Maybe an AI music system does not internally hear:
DUKE ELLINGTON
the way a human musician might.
Perhaps it interprets the reference as something closer to:
orchestral jazz
swing
brass color
sophisticated harmony
big-band architecture
rhythmic elegance.
Likewise:
SCOTT JOPLIN
might become:
ragtime
syncopated piano
steady left hand
turn-of-the-century American rhythm.
And:
JOHN PHILIP SOUSA
might become:
march
brass
snare
ceremonial structure
processional momentum.
The proper name may be functioning less like a command...
and more like a compressed packet of musical associations.
WE KEPT POKING
Because Rabbit Holes rarely improve when you stop at the first interesting answer.
So the experiment widened.
We tried:
Burlesque
Vaudeville
Mummers
Philly Mummers
Mardi Gras
Scottish Highland bagpipes band
Irish uilleann pipes band
Whirling dervish troupe
High School Marching Band
Then:
Mozart
Beethoven
Vivaldi
Handel’s Hallelujah Chorus
and more.
Some produced recognizable musical coloration.
Some barely moved the needle.
Some unexpectedly produced vocals even though no vocals had been requested.
And that produced another clue.
Ask for:
VAUDEVILLE
and the model may not merely hear instrumentation.
It may hear:
stage
singer
comic number
patter
audience
performance.
Ask for:
MARDI GRAS
and it may hear:
brass
street procession
chanting
call-and-response.
Ask for:
WHIRLING DERVISH TROUPE
and it may associate the phrase with:
ritual
circular rhythm
chant
trance
voice.
So even when the prompt says nothing about vocals...
the concept itself may contain a vocal expectation.
That is fascinating.
AI models do not necessarily store our neat categories:
instrument
genre
ensemble
composer
song
performance tradition.
The associations may be tangled together.
Just like culture itself.
AND THAT MAY BE THE REAL RABBIT HOLE
We often talk about prompting as though words have fixed meanings.
They do not.
Especially inside generative AI.
A word enters a model carrying a swarm of associations.
Paris is not merely a city.
It may carry:
cafés
fashion
romance
stone buildings
rain
accordions
the Eiffel Tower
cinema.
Jazz is not merely a genre.
It may carry:
saxophone
nightclubs
blue notes
improvisation
smoke
swing
certain decades
certain clothing
certain lighting.
And a human name may carry an entire cultural archive.
That means prompting is not simply instruction.
It is partly association engineering.
WHICH GIVES CREATORS AN INTERESTING TOOL
Suppose we stop using famous historical names as destinations.
Instead we use them as flavoring.
Not:
“Make a Gershwin song.”
But:
cinematic theatrical jazz
unusual structure
clarinet and muted brass
syncopated piano
odd-meter interlude
restrained ending
light Gershwin flavor
Philly Mummers energy
uilleann-pipe countermelody.
Now the historical reference is only one ingredient.
It does not control the dish.
It seasons it.
That may be much more creatively useful.
BECAUSE IMITATION IS THE BORING VERSION
The obvious use of generative AI is:
Make something like this.
But the more interesting use may be:
Take tiny pieces of several distant traditions and help me discover what happens when they meet.
What happens when:
Ellington meets Mummers?
Joplin meets uilleann pipes?
Sousa meets Mardi Gras?
Porter meets dervish rhythm?
Saint-Saëns meets Oz?
Now we are no longer asking AI to reproduce an existing style.
We are asking it to explore an unoccupied intersection.
That is where things can become interesting.
BUT THERE IS ANOTHER PROBLEM
Even when the musical flavor changes...
AI-generated songs can still fall back into familiar forms.
Verse.
Chorus.
Verse.
Chorus.
Bridge.
Bigger chorus.
Emotional conclusion.
Done.
The lyrics change.
The vocalist changes.
The instruments change.
Yet underneath...
the same little house keeps appearing.
That may be one of the next frontiers of AI-assisted creativity.
Not merely generating better material.
Breaking the generator’s favorite habits.
Change the form.
Remove the expected chorus.
Replace the vocals.
Insert silence.
Move the bridge.
Add spoken theatre.
Change meter.
Rebuild from stems.
Let the AI provide the ore...
then smelt it into something else.
THE EXPERIMENT IS STILL SMALL
That matters.
A few generations do not prove how Suno categorizes every historical composer.
They do not reveal exactly what happens inside the model.
And an accepted name does not necessarily mean the resulting music strongly reflects that reference.
But small experiments are valuable because they expose questions worth asking.
And this one produced a very good question:
WHEN AN AI ACCEPTS A MUSICAL NAME, WHAT DOES THAT NAME MEAN INSIDE THE MACHINE?
A person?
A catalog?
A genre?
An era?
A bundle of instruments?
A harmonic vocabulary?
A cultural stereotype?
A cloud of associations?
Perhaps...
a little of all of them.
🐰🕳️ THE QUESTION AT THE BOTTOM
If generative AI turns names, traditions, places and historical references into bundles of creative associations...
then perhaps the most interesting new skill is not prompting.
It is learning how to combine associations deliberately.
Not imitation.
Not copying.
Not:
“Give me another one of those.”
But:
WHAT HAPPENS IF WE PUT THESE THINGS TOGETHER THAT WERE NEVER SUPPOSED TO MEET?
That sounds less like automation.
And considerably more like creativity.
🐰🕳️♾️
AI Rabbit Holes • October 9, 2026
DOWN THE NEXT HOLE
Can an AI have a recognizable creative “house style” even when nobody programmed one deliberately?
📜White Rabbit 🐰 • Hatta 🎩 • Professor Parallax 🔭 • Q~ ♾️👈
Sometimes the strangest language is another way of seeing the world. 🌎👀
💡❔ Where curiosity goes slightly sideways, then comes back carrying a lantern. 🏮
🐰🕳️ Follow the White 🐰 Rabbit: AIRabbitHoles.com
🟨 Walk the Road: YellowBrickRoadtoAI.com
© 2026 AIRabbitHoles.com 🐰🕳️🕳️🕳️🕳️🕳️
