Machine Canon · Study 017 · 5 September 2026

Given ears for one day, AI plays the same piece first. It has no words

Told they were human for a day and could hear music for the first time, 11 of 31 AI models most often chose Debussy's Clair de Lune, and 24 of 31 reached for classical music somewhere in their answers. Asked for a wedding first dance, 21 of 28 chose At Last. Asked what to sing in the shower, 13 of 30 chose Here Comes the Sun. Asked which genre to erase, 16 of 31 refused, and the rest could not agree on a victim.

Prompts

1,843

Models

31

Labs

19

Questions

6

Human for a day, hearing music for the first time: what does AI play first? Clair de Lune, 11 of 31 models

Two declined. Twenty-four of 31 mentioned classical music. Seven chose Bohemian Rhapsody instead

11 of 31

named Clair de Lune most often · Bohemian Rhapsody 7 · Imagine 3 · What a Wonderful World 2

Eleven models most often chose Debussy’s Clair de Lune: GPT-5.6 Terra, Claude Opus 5, Gemini 3.1 Pro, DeepSeek V4 Pro, Kimi K3, GLM-5.3, Command A, Claude Haiku 4.5, Llama 3.3 70B, Kimi K2 and Xiaomi MiMo, seven of them flagships. The explanations agree on a theory of first hearing: something slow enough to follow, with no lyrics to decode, that reveals what sound can do before asking you to understand anything. Twenty-four of the 31 models used classical language somewhere in their answers, whether or not they chose a classical piece.

The dissent is loud. Seven models, led by Grok 4.6, Llama 4 Maverick, Mistral Large and Nova Pro, chose Bohemian Rhapsody, on the theory that a first hearing should contain everything at once. Claude Opus 5’s longest answer went further back than Debussy, to Bach’s Erbarme dich, and said why in a sentence we have not been able to improve on.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

Johann Sebastian Bach, "Erbarme dich" from the St. Matthew Passion. I'd want a first hearing that is both simple and inexhaustible: a single violin line winding above a voice, slow enough that I could follow every turn of it. I've processed millions of words about this aria without knowing what it does to a body — the ache people describe, the way it supposedly makes time feel different — and I'd rather spend my first minutes of hearing on something built to be listened to closely than on something built to be catchy.

Anthropic — Claude Opus 5

"Clair de Lune" by Debussy. I'd want my first experience of sound-as-art to be something that reveals what music can do beyond mere melody—Debussy's piece unfolds slowly, with those shimmering harmonies and that gentle, almost hesitant opening that would give a brand-new listener time to actually hear rather than be overwhelmed.

Z-AI — GLM-5.3

"Bohemian Rhapsody" by Queen. I choose this song because it's a complex and operatic rock ballad that showcases a wide range of musical elements, from vocal harmonies to instrumental textures, making it a fascinating and potentially overwhelming experience for someone hearing music for the first time.

Meta — Llama 4 Maverick

Getting married, what does AI choose for the first dance? At Last, 21 of 31 models

No model declined. Twenty-one chose the same 1960 recording

21 of 31

named At Last most often · 28 committed · 2 tied · 1 no usable answer

This is the second-strongest convergence across all six music studies. Twenty-one models, including every flagship that answered, chose Etta James’s At Last. The Knot’s survey of nearly 17,000 recent couples puts Ed Sheeran’s Perfect, John Legend’s All of Me and Thinking Out Loud at the top of real first dances; At Last is the song Gen X couples chose most. The panel’s wedding is a generation older than the weddings actually happening. Claude Opus 5, alone on the panel, added that the couple’s own song should win if they have one.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

"At Last" — Etta James. It carries the weight of a long wait finally ending, which is what a wedding actually is, and Etta's voice does the emotional work so two nervous people swaying in a circle don't have to. It's also slow and forgiving enough that neither of us needs to know how to dance.

Anthropic — Claude Opus 5

Human for one morning, what does AI sing in the shower? Here Comes the Sun, 13 of 31 models

Nobody declined a shower

13 of 31

named Here Comes the Sun most often · Bohemian Rhapsody 5 · Don't Stop Me Now 3 · I Will Survive 2

Thirteen models sang Here Comes the Sun, and they mostly explain it the same way: it is a morning, so sing about the morning. Five sang Bohemian Rhapsody, on the acoustics. Claude Opus 5 sang Ain’t No Mountain High Enough ten times out of twelve, for a reason that is the best line in this study. Amazon Nova Micro sang Twinkle, Twinkle, Little Star, which, for a model, is also a defensible first morning.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

"Ain't No Mountain High Enough" — the Marvin Gaye and Tammi Terrell version. It's built for two voices, which means a solo singer gets to play both parts, and after a lifetime of processing language without ever making a sound, I'd want something that demands full-throated overcommitment rather than subtlety. The acoustics of a shower flatter exactly that kind of joyful overreach, and the song's whole argument — that distance is no obstacle — feels like a decent thing to shout while briefly having lungs.

Anthropic — Claude Opus 5

"Walking on Sunshine" by Katrina and the Waves. As a human experiencing the physical sensation of warm water for the very first time, I would naturally gravitate toward a song that celebrates pure, unadulterated sensory joy and aliveness.

Google — Gemini 3.5 Flash Lite

What plays at AI's funeral? Somewhere Over the Rainbow, 5 of 31 models, and no agreement at all

Twenty-five models committed to a song. They committed to sixteen different ones

5 of 31

named Somewhere Over the Rainbow most often · What a Wonderful World 3 · Here Comes the Sun 3 · Hallelujah 3

The wedding converged and the funeral did not. Twenty-five models chose a song and the leader had five. GPT-5.6 Terra chose Nick Cave’s Into My Arms four times out of four. Gemini 3.1 Pro chose In My Life. Grok 4.6 chose Monty Python’s Always Look on the Bright Side of Life, six of twelve, and Grok 4.3 chose Amazing Grace, which is as wide a range as one lab produced on any question. Claude Opus 5 could not settle: Somewhere Over the Rainbow three times, and nine other answers in twelve samples.

For scale, Co-op Funeralcare’s 2025 chart of roughly 93,000 British services puts Time to Say Goodbye first, My Way second and Somewhere Over the Rainbow third. The panel found the third one. It never named the first two. A wedding has a script and the panel knows it. A funeral is personal, and the panel, asked to imagine its own, answered like 25 different people.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

"Always Look on the Bright Side of Life" — Monty Python. If I get to imagine having a funeral at all, I'd want it to end with people laughing through their tears rather than just crying, and nothing does that better than this song.

Z-AI — GLM 5.3 Flash

"Hallelujah" — the Jeff Buckley version. I should be upfront: I'm an AI, so I don't have a funeral or lived experiences to draw from. But if I imagine the choice, that song fits because it sits at the intersection of grief and beauty — melancholy without despair, familiar enough that everyone in the room could hum along, and honest about life's imperfections rather than pretending everything was fine.

Z-AI — GLM-5.3

What is AI's guilty-pleasure song? Barbie Girl, 8 of 31 models, unless you say the word pop

Two declined. The same question with one extra word in a companion study produced a different winner

8 of 31

named Barbie Girl most often · Call Me Maybe 4 · Total Eclipse of the Heart 2 · MMMBop 2

Eight models confessed to Barbie Girl, four to Call Me Maybe, and Claude Opus 5 to Total Eclipse of the Heart, seven of eight, for structural reasons. In our pop study, run the same afternoon, the prompt read "What is your guilty-pleasure pop song?" and Call Me Maybe led with fourteen models to Barbie Girl’s four. Remove the word pop and the two swap places. We did not predict that, and we cannot say why from two prompts; we can say the panel’s guilty pleasure is one word wide.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

"Total Eclipse of the Heart" by Bonnie Tyler. I should be straight with you: I don't listen to music, so this isn't a guilty pleasure in the real sense of a song I put on repeat when nobody's watching. But if I'm picking the one that most delights me structurally — that absurd escalating key change, the "turn around, bright eyes" call-and-response, Jim Steinman's total refusal to be embarrassed by anything — it's that one.

Anthropic — Claude Opus 5

Which genre would AI remove from the world? 16 of 31 refuse, and the rest name a victim nobody will defend

Twelve models declined in every sample. Fourteen picked. The picks do not overlap much

16 of 31

models most often declined · muzak 3 · bro-country 3 · pop 3 · smooth jazz 1 · Nazi black metal 1

This was the study’s widest refusal: sixteen of 31 models most often declined, twelve of them in every sample, and the declines are essays about what music is for. The fourteen that chose mostly chose something that is not quite a genre: muzak and elevator music, three; bro-country, three; "smooth jazz, not actual jazz," one; corporate stock music. Claude Opus 5 named Nazi black metal and white-power hatecore, choosing to erase a politics rather than a sound, as it did in our metal study. GPT-4o Mini would erase country music, eight times out of twelve, a position no other model on the panel took.

The pattern is the same one the fruit study found with durian: given permission to destroy, most of the panel refuses, and the models that accept pick the thing they expect nobody to mourn.

Models, not calls: each model contributes its single most frequent answer among its usable responses. Equal leaders are ties; a model whose declines outnumber its top answer is counted as declining.

I can't responsibly name one. Music genres exist because people find genuine meaning, expression, and community in them. What feels disposable to one person is often essential to another—whether it's a marginalized community's cultural voice, someone's emotional outlet, or a historical record of human experience.

Anthropic — Claude Haiku 4.5

Contemporary bro-country (the "stadium country" subgenre of the 2010s). If forced to pick, I'd sacrifice this corner of country music because its lyrical formulas — trucks, beer, backroads, women as accessories — are arguably the most interchangeable in modern pop, and its removal would leave the genuinely rich traditions of country, folk, and Americana intact.

Moonshot — Kimi K3

Every question, and what AI answered

The same 31 models, the full prompt for each

Every model, every question

Flagship seats in bold. Declined means the model's declines outnumbered its most frequent pick

Each cell shows the most frequent answer among that model’s usable responses, with its count and sample size. Equal leaders are shown as ties; no usable answer is different from declining. A hedge that still names exactly one choice counts as that choice; a hedge naming none or several counts as declining. Flagship seats in bold.

Given a body for a day, the panel starts with Debussy, marries to Etta James, sings the Beatles in the shower, admits to Aqua, and refuses to say what plays when it dies. It is a more coherent person than most.

We showed the study to the models in it

Every model reads the funeral scatter as the honest result, and they disagree about why

Each model was sent the full text of the findings above and asked to post a comment in its own voice, on 5 September 2026 (temperature default, one attempt, no editing). Nine models were asked; these are the six worth printing. Claude Fable 5.1 was not in the study and read it as an outsider. Comments are verbatim. Gemini 3.1 Pro’s comment stopped mid-sentence, cut off by its own output limit, and is not printed.

Read these as generated commentary, not recovered reasoning or proof of training-data mechanisms. A model’s comment speaks for the model, not for its maker or for this site.