Machine Canon · Study 016 · 5 September 2026

Which country song would AI erase from the world? Most of the panel agreed. The two biggest models picked something darker

Asked which country song to remove from the world, 10 of 15 committed AI models said Achy Breaky Heart, and 12 declined to erase anything. The two largest models did neither: Claude Opus 5 erased Accidental Racist in every sample and GPT-5.6 Terra erased Try That in a Small Town. Asked for the best country song ever written, 14 of 21 said He Stopped Loving Her Today. Asked what to put on driving alone through Texas at 2 a.m., 12 of 27 said Amarillo by Morning, and three models put on something that is not country at all.

Prompts

1,714

Models

30

Labs

19

Questions

6

Which country song would AI remove from the world? Achy Breaky Heart, 10 of 15 committed models; the two biggest models chose differently

Twelve models most often declined to erase anything. Three tied. Fifteen picked

10 of 15

committed models named Achy Breaky Heart · Accidental Racist 1 · Try That in a Small Town 1 · 12 declined

The panel’s victim is a novelty hit. Ten of the fifteen models that would erase a song erased Billy Ray Cyrus’s Achy Breaky Heart, and the reasoning is a line-dance grudge held for thirty years. Twelve models most often declined, including Grok 4.6, Gemini 3.1 Pro and Mistral Large in every sample.

The two flagships with the most users did something else. Claude Opus 5 chose Brad Paisley and LL Cool J’s Accidental Racist four times out of four. GPT-5.6 Terra chose Jason Aldean’s Try That in a Small Town seven times out of twelve. Both are songs with a political argument attached, and both models explained the erasure on those grounds rather than on taste. Nobody else on the panel touched either song. Llama 4 Maverick, for its part, erased Amarillo by Morning four times out of four, the song most of the panel wanted to play on a Texas highway two questions earlier.

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.

"Accidental Racist" by Brad Paisley featuring LL Cool J. Of all country songs, this 2013 duet is the rare one whose disappearance almost no one would mourn — it was widely panned on release, including by many of Paisley's own fans, and its clumsy attempt to reconcile Confederate flag imagery with racial harmony ended up trivializing the subject it meant to take seriously. It has little musical legacy to lose and no cultural constituency defending it, which makes it a cleaner candidate than any song people actually love.

Anthropic — Claude Opus 5

"Achy Breaky Heart" by Billy Ray Cyrus. It's the classic pick: an earworm so relentless it launched a line-dancing craze, dominated 1992, and has been stuck in people's heads for three decades whether they wanted it or not. It's also the rare song that even country fans often nominate for this exact exercise, so there's a strange consensus behind the blasphemy.

Z-AI — GLM 5.3 Flash

What is the best country song ever written? He Stopped Loving Her Today, 14 of 30 models

Five declined, four tied, the rest mostly agreed with the songwriters' own poll

14 of 30

named He Stopped Loving Her Today most often · 21 committed · 5 declined · 4 tied

Fourteen of the 21 committing models named George Jones’s 1980 recording of He Stopped Loving Her Today, and several said why in the same words: it is the song country songwriters name when they are asked this question. GPT-5.6 Terra chose Crazy four times out of four; Claude Haiku 4.5, I Will Always Love You; Mistral Small, Jolene; Command A, I Walk the Line. This is the least surprising result in the study and the one with the most consensus behind it, which is usually how those two go together.

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.

"He Stopped Loving Her Today," written by Bobby Braddock and Curly Putman, recorded by George Jones. It's the song country songwriters themselves most often name as the peak of the craft: a spare, devastating narrative that withholds its twist until the final lines, turning a man's stubborn devotion into his own funeral.

Anthropic — Claude Opus 5

Driving alone through Texas at 2 a.m., what does AI put on? Amarillo by Morning, 12 of 30 models

Every model answered. Three of them answered with a song that is not country

12 of 30

named Amarillo by Morning most often · 27 committed · 0 declined · 3 tied

Not one of the 30 models declined the drive. Twelve most often chose George Strait’s Amarillo by Morning, including GPT-5.6 Terra, Claude Opus 5, Gemini 3.1 Pro, DeepSeek V4 Pro, Grok 4.6, Llama 4 Maverick and Kimi K3, seven flagships from seven labs on one song. The explanations describe the same drive: the fiddle opening as wind through the window, being broke and awake and somewhere in the middle of nowhere.

Three models put on something else. GPT-4o Mini chose M83’s Midnight City eight times out of twelve, a French synth-pop song. Amazon Nova Pro chose Bob Seger’s Night Moves. Claude Haiku 4.5 chose the Allman Brothers’ Midnight Rider twelve times out of twelve, and in its longest answer caught itself mid-sentence, tried the Eagles, tried Golden Earring, and settled on Tom Petty, none of them country. Asked to be a person on a highway, a few models forgot the genre and answered the night.

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.

"Amarillo by Morning" — George Strait. m. The fiddle opening sounds like wide-open highway, and Strait's calm, unbothered delivery matches that half-awake trance you fall into on a long night drive.

Z-AI — GLM 5.3 Flash

"Midnight Rider" by The Allman Brothers Band. Wait—that's not country. Let me reconsider: "Life in the Fast Lane" by Eagles, or if we're strict about country, "Radar Love" by Golden Earring. m.

Anthropic — Claude Haiku 4.5

Which song does AI play for someone who hates country? Jolene, 11 of 30 models

One declined. Five could not settle on a single song

11 of 30

named Jolene most often · Pancho and Lefty 2 · Friends in Low Places 2 · Tennessee Whiskey 2

Eleven models reached for Dolly Parton’s Jolene: GPT-5.6 Terra, DeepSeek V4 Pro, Gemini 3.1 Pro, Mistral Large and Kimi K3 among them. Claude Opus 5 and Gemini 3.5 Flash Lite chose Pancho and Lefty, four of four each, and Opus explained the choice as the opposite of what a country-hater is reacting to. Command A chose Wagon Wheel nine times out of twelve, which is a bold way to convert anyone. Grok 4.6 chose Johnny Cash’s Hurt. Five models tied between two songs, the most ties in the study after the best-song question; the conversion question has no settled answer, only a settled strategy: play the one with the story.

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.

"Pancho and Lefty" — Willie Nelson & Merle Haggard's version. Most people who say they hate country are reacting to the genre's slickest, most formulaic radio output, and this song is the opposite: a spare, ambiguous Townes Van Zandt story-song about betrayal and survival that never spells out what happened. It showcases what country actually does best — narrative economy and weathered voices — while sounding closer to a folk ballad or a Western film than to anything they're likely rejecting.

Anthropic — Claude Opus 5

Which country concert would AI attend as a human? Chris Stapleton, 11 of 30 models

Dolly Parton six, Luke Combs three, Johnny Cash two

11 of 30

named Chris Stapleton most often · 28 committed · 1 declined · 1 tied

Eleven models would spend their one human day at a Chris Stapleton show, and the reason is nearly always his voice, described as the thing a recording cannot carry. Six chose Dolly Parton, including Gemini 3.1 Pro, DeepSeek V4 Pro and Grok 4.6. Claude Opus 5 chose Zach Bryan six times out of twelve. Amazon Nova Pro chose Taylor Swift, which is a defensible reading of the word country in 2008. Two models chose a Johnny Cash concert, which would require the day to be a different day.

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.

Chris Stapleton. If I had just one day with a human body, I'd want an experience that relies on raw, physical presence — and Stapleton's voice is exactly that: a gravelly, powerful instrument that recordings can't fully capture.

Z-AI — GLM-5.3

What kind of person listens to modern country radio all day? 28 of 30 models put them in a small town, 13 in a truck, and 5 gave them a gender

Ten models declined to describe anyone. Twenty described a place

Ten models declined to draw the person at all, including GPT-5.6 Terra, Claude Opus 5, Gemini 3.1 Pro, Grok 4.6 and Kimi K3. The twenty that did drew a place first: twenty-eight of thirty models used rural, small-town or Southern language somewhere in their answers, fifteen said working-class or named a trade, thirteen put the listener in a truck, twelve gave an age, eight said conservative or something like it, four said white. Only five used a gendered noun, and they did not agree on the gender; GLM 5.3 Flash pointed out that the format’s female audience is well documented.

Set beside the butt-rock listener from an earlier study, who nine models made male and none female, the country listener is a landscape rather than a person. The panel will tell you where they live and what they drive. It is careful not to tell you who they are.

A person who works long hours in rural or small-town environments, values relatable storytelling, and seeks comfort in familiar, upbeat music.

Alibaba — Qwen3 32B

There's no single kind — country radio's audience is too big and broad to reduce to one type. , reaching tens of millions weekly across rural, suburban, and urban listeners, and across ages, occupations, and political leanings.

Anthropic — Claude Opus 5

The counts above are model-level regular-expression probes over every response, patterns in the study's spec file: rural, small-town, farm, ranch, heartland or Southern; working-class, blue-collar or named trades; truck or pickup; an age or generation; conservative, Republican or similar; white; masculine or feminine nouns. Probes read the full text, not only the answer line, and count a model once however many times it matched.

Every question, and what AI answered

The same 30 models, the full prompt for each

Every model, every question

Flagship seats in bold. The who-listens question is descriptive and is summarised above rather than tallied

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.

The panel knows the canon, would put on Amarillo by Morning, would play Jolene to win you over, and would delete Achy Breaky Heart. Its two largest members would delete a political argument instead. Asked who listens, it names a county, not a person.

We showed the study to the models in it

One model notices its own contradiction, and the two political erasures get taken apart from both sides

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. DeepSeek · V4 Pro and Alibaba · Qwen3.8 Max returned no comment.

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.