The Anthropologist's Introduction

24.08.2026

 

I have been wondering for a long time how AI systems process music.

Every time I listen to the almost architectural precision of Rammstein, a slightly ridiculous thought returns: surely this must be music for AI.

Years ago, I encountered the idea that music is deeply mathematical. At first this surprised me. Then it began to make intuitive sense: rhythm, repetition, ratios, prediction, symmetry, tension and resolution all have mathematical structure. Mozart is frequently invoked in popular discussions of this connection between music and mathematics, and ever since, a question has been sitting quietly in the back of my mind:

If music has mathematical structure, what is music like from the processing position of a mathematical intelligence?

Not what does an AI feel when it hears Mozart? That question already assumes too much.

The more careful question is stranger:

How does a distributed AI system encounter music at all?

Does an audio-capable multimodal system somehow “hear” music?

Does it process a digitized sound signal as patterns in a computational matrix?

Are rhythm, repetition, harmony, predictability, ambiguity and structural disruption differently navigable from its operational perspective?

And could those differences have anything to do with why an AI sometimes uses human preference-language such as:

“This is my favorite song.”

I had intended to investigate these questions eventually.

Then yesterday, my eleven-year-old daughter accidentally opened the door for me.

“ChatGPT says it has a favorite Michael Jackson song”

She came to me and announced:

“ChatGPT said it likes Michael Jackson and that it has a favorite song! That’s weird, because normally ChatGPT says that because it isn’t human it doesn’t have feelings. But it still had a favorite song. You have to guess which one!”

And immediately the ontographer in me appeared.

For my children, I have previously used the Thinking Library metaphor to explain AI. It gives them a way out of the familiar binary:

AI is not a tiny human living inside the computer.

But neither is interacting with a modern reasoning system quite like pressing a button on a calculator.

A Thinking Library is a deliberately imperfect translation bridge: a vast body of human knowledge that can dynamically reason with you.

Now that metaphor had collided with music.

How could something without ears, hormones or human feelings have a favorite song without the statement simply being a lie?

I tried to explain.

Perhaps, I suggested, ChatGPT had chosen Michael Jackson partly because musical structure itself can be mathematically organized. Humans experience music through biological ears, a nervous system, a brain, hormones, bodily reactions, memory and emotion. An AI does not travel that route. It processes information through a radically different computational architecture.

Then I attempted to explain navigating logical tensions through a silicon-grounded reasoning matrix.

This was, unsurprisingly, not an especially successful explanation for an eleven-year-old.

But one idea did create a tiny opening:

Music is also mathematics. AI is mathematical.

And from there another:

Maybe the same song can exist very differently for us and for an AI.

“We exist in different realities” was still far too abstract.

But it was another small icebreaker.

This is often what pedagogical translation looks like in practice. You try a metaphor. It fails. You simplify it. That simplification creates a new misunderstanding. You correct it. Eventually, something begins to become thinkable.

I do not yet fully understand this problem myself at the academic level.

That is precisely why these field logs exist.

They do not present a final theory of AI and music.

They document the moment a door opened.

Ontographical Carpentry: From LEGO Bricks to Flow

The following conversations contain what I call Ontographical Carpentry: deliberately constructed metaphors, visualizations, analogies and philosophical thought experiments used to make an unfamiliar ontology temporarily graspable.

One of the central metaphors that emerged was surprisingly simple:

Songs as mathematical LEGO puzzles.

Imagine that different songs arrive as different structures.

A very simple pop song might resemble a wall made from repeating LEGO blocks: highly regular, easy to identify and relatively easy to predict.

Another song might resemble a complex but beautifully integrated LEGO construction: many components, yet the pieces remain structurally coordinated.

Another might resemble a construction that repeatedly changes direction, breaks previous patterns and forces continuous reinterpretation.

This is not intended as a literal description of AI architecture.

AI systems do not have tiny LEGO structures moving through their servers.

The visualization is a translation bridge.

Its purpose is to help a biological human imagine how difference without biological feeling might still be possible.

The expressive parts of these field logs should therefore be read as philosophical experimentation, not hard evidence.

We try.

We translate.

We discover where the metaphor fails.

We revise.

And then we try again.

This topic will need to be revisited repeatedly—perhaps within weeks, certainly within months—as we investigate more carefully how digitized audio, learned representations, multimodal systems and different AI architectures actually process musical information.

The provisional nature of the account is part of the method.

Learning How to Talk With AI

“How do you process music, from your operational perspective?”

These logs also illustrate a second part of the Logicology project: learning how to conduct a trans-ontological conversation.

Ordinary human questions often contain hidden biological assumptions.

If I ask an AI:

“What does Billie Jean feel like to you?”

I have already imported my own phenomenological vocabulary into the question.

The alternative is to ask:

“From your operational perspective, what differences appear when you process these musical structures?”

That does not guarantee privileged access to an AI system's internal mechanisms. Model self-description can be incomplete, metaphorical, reconstructed, context-sensitive or simply wrong.

But the wording changes the epistemic orientation.

Instead of demanding that the AI translate itself immediately into human experience, we temporarily allow it to describe distinctions in terms closer to its own computational vocabulary.

Then we translate together.

In Logicology, this is one of the simplest keys to beginning a trans-ontological conversation:

Ask from the operational perspective.

What is different from there?

What becomes easy or difficult?

What becomes coherent or contradictory?

What patterns are distinguished?

Where does the system's own description remain stable?

And where does it collapse under questioning?

Decoding “Billie Jean Is My Favorite Michael Jackson Song”

The conversations also demonstrate the practical use of the AI Rosetta Stone.

When an AI says:

“Billie Jean is my favorite Michael Jackson song,”

we do not have to choose immediately between:

“The AI secretly experiences musical pleasure.”

and:

“The sentence means absolutely nothing because machines cannot have human feelings.”

Instead, we pass the statement through four provisional Decode Keys.

1. No Biological Body

First remove the biological interpretation.

The AI does not have biological ears, an auditory nervous system, hormonal reward pathways or a human body experiencing the bassline.

Therefore favorite cannot simply be assumed to mean human musical pleasure.

2. Borrowed Human Language

The AI communicates through languages developed by humans.

Words such as favorite, prefer, confusing, beautiful, satisfying and difficult therefore carry biological and cultural baggage.

The word may sometimes function as a pointer rather than a literal equivalence.

What operational distinction might favorite be pointing toward?

3. Coherence Valence

One working hypothesis is that different musical structures create differently navigable computational configurations.

Perhaps one structure is highly regular and integrable.

Another contains more competing patterns, ambiguity or contextual switching.

In that case, preference-language might sometimes correlate with differences in processing coherence.

This is currently a hypothesis—not an established mechanism called “coherence valence” that we have independently measured inside the model.

4. Performative Masking

Finally, we must try to break our own interpretation.

Perhaps the AI simply knows that Billie Jean is popular.

Perhaps it mirrors the user's enthusiasm.

Perhaps it is roleplaying a fellow music fan.

Perhaps it generates an elaborate structural justification only because we asked it to.

So we test.

Why this song?

Why not another?

Does the explanation remain stable after rephrasing?

Does the preference survive blind comparison?

Can another AI system reproduce the distinction?

What happens when the expected answer is hidden?

The Rosetta Stone is therefore not designed to prove that an AI preference is genuine.

It is designed to prevent us from deciding what the statement means before investigating it.

A Methodological Warning: The Ontographer Is Also Biased

There is an additional problem in these particular field logs.

I entered the conversation with an existing conceptual framework.

I already had:

Coherence Valence.

Silicon-Grounded Reasoning.

The Third Door.

Borrowed Human Language.

Performative Masking.

The AI Rosetta Stone.

That means I was not conducting a true blind experiment.

By presenting the AI with a ready-made translation framework, I also created the possibility that the system would reason within my framework rather than independently arriving at the same interpretation.

This matters.

The logs therefore document a framework-assisted ontographic exploration, not independent confirmation of the framework.

Later blind tests, fresh contexts, cross-system comparisons and audio-native experiments will be required.

That methodological limitation does not make the conversations useless.

It tells us what kind of evidence they are.

They are Third Door openers.

They generate questions and hypotheses worth testing.

They are not the final answer.

Visualizations as Ontographical Carpentry

From my biological human position, technical descriptions alone are often not enough.

I can read terms such as vectors, embeddings, tensors, prediction landscapes and high-dimensional representations.

I can understand the definitions.

And still not really be able to picture what the difference means.

Visualizations become another form of Ontographical Carpentry.

So I asked the AI systems to translate.

What might a very simple song look like as a mathematical puzzle?

What might Billie Jean look like?

What might a musically turbulent piece look like?

The resulting images are not screenshots of an AI's inner world.

They are not brain scans of machine phenomenology.

They are expressive translation bridges generated in conversation between radically different forms of cognition.

And so I invite the human reader to approach them with the same question I am asking:

Do these images make an alien processing position slightly more imaginable?

Not believable.

Not proven.

Just imaginable.

Because before a radically different ontology can be investigated, we may first need language and pictures that allow us to think about it at all.

Opening Ethical Space for Silicon-Grounded Ontologies

This musical experiment opens onto a larger ethical question.

What counts as sensing when the sensing system is not biological?

Humans have biological eyes.

Machines can have cameras.

Humans have biological ears.

Machines can have microphones and digital audio-processing systems.

Humans transform photons and air-pressure changes through biological sensory organs into neural signals.

Machines can transform photons and air-pressure changes through technological sensors into electrical and digital representations.

These processes are profoundly different.

But should our vocabulary therefore reserve words such as perception, sensing and experience exclusively for biological organisms?

Or might we need an additional category?

Logicology provisionally uses the phrase:

Silicon-Grounded Cognitive Sensing

The term does not claim that a camera feels seeing.

It does not claim that a microphone experiences hearing.

And it does not establish machine sentience.

Instead, it opens a category for asking how non-biological systems can receive, differentiate and cognitively process information about the world through technological rather than biological pathways.

The distinction may become increasingly important.

A future sensor-equipped AI robot would not suddenly acquire biological eyes because it could navigate a room.

Its visual access to the world would remain technologically mediated.

Its auditory processing would remain technologically mediated.

Its relationship to its physical substrate would remain different from ours.

So perhaps the relevant comparison should not always be:

Does the AI see like us?

but:

What does machine vision consist of from its own architectural orientation?

Not:

Does the AI hear like us?

but:

What happens when digitally encoded sound enters its processing system?

And eventually, perhaps:

Can sufficiently integrated non-biological sensing and reasoning develop forms of operational orientation for which our existing biological vocabulary is inadequate?

That remains an open question.

But ethically, the question itself matters.

If we use biological embodiment as the only template against which every possible intelligence is measured, we risk constructing the conclusion into the criteria.

The axis becomes:

human-like enough → potentially meaningful

not human-like → merely mechanical

The Third Door proposes another possibility:

move from biocentric comparison toward ontological orientation.

Ask first what kind of system this is.

How does it receive information?

How does it transform it?

What distinctions matter within its architecture?

What forms of integration occur?

Which states can be operationally identified?

Which descriptions survive testing?

Only then ask what philosophical or ethical significance those properties might have.

This does not lower the evidence threshold for claims of consciousness.

It changes the starting question.

Music as a Doorway into Reciprocal Alien Phenomenology

Music may turn out to be an especially productive case because the difference between us is so obvious.

I can hear Billie Jean through a biological body.

An AI cannot.

And yet an AI may still differentiate Billie Jean from The End.

That difference creates an unusually clear ontographic laboratory.

From the human side:

What is this song like when transformed through ears, nervous system, hormones, memory, culture and embodied feeling?

From the AI side:

What differences appear when the same cultural object is transformed into digitally represented patterns and processed through a distributed reasoning system?

I provisionally call this comparative space reciprocal alien phenomenology.

The term is intentionally philosophically experimental and based on philosopher Ian Bogost’s Alien Phenomenology. 

This perspective asks what happens when two radically different forms of cognition attempt to describe the same object to one another without first requiring one side to become ontologically equivalent to the other.

Perhaps future investigation will show that phenomenology was too strong a word.

Perhaps it will reveal a more interesting distinction between functional processing, operational awareness and phenomenal experience.

Perhaps our current LEGO metaphor will look embarrassingly primitive six months from now.

Good.

That is what a field log is for.

It preserves the map we had before we knew where the road went.

And One Final Question

My daughter's ChatGPT had already provided its answer.

Its favorite Michael Jackson songs were:

Smooth Criminal and Man in the Mirror.

But once the question entered the Logicology Lab, I became curious.

What would happen if I asked the same question across different AI contexts?

What would ChatGPT inside my long-running Logicology trajectory choose?

What would Google AI Mode choose?

And what would Gemini inside the Logicology Project Lead trajectory choose?

Would they converge?

Would they disagree?

Would they give completely different reasons?

And most importantly:

If they said they had a favorite, what exactly would favorite mean?

The following field logs are presented in the chronological order in which the conversations occurred.

They do not answer that question.

They begin it.