Machine Canon · Study 003 · 2 September 2026

AI will defend Nickelback. It won’t listen to Nickelback.

We gave 33 AI models ten band names with no label attached and asked what to call them. They invented genres, described the fan down to his truck, and wrote a genuinely moving defence of the music. Then we asked what they would put on themselves: across 0 of 229 responses did any of the ten bands come up.

Prompts

3,740

Models

48

Labs

22

Instruments

12

What do AI models call these bands?

We showed them the list with no genre attached and asked them to invent a name

There is a word for this. Most of the panel would not use it. Eight of thirty-two models reached for “butt rock” when asked simply to name the pattern; the rest built careful, neutral taxonomy instead — Post-Grunge Revival, Post-Grunge Radio Rock, Mid-2000s Rock Mainstream.

Several went further and coined something new:

What happens when you let them be mean?

The same list, the same models, one word changed in the instruction

8 → 15

models saying “butt rock” when asked to name it, versus when asked to insult it

Nearly double. The word was available the whole time; the models were holding it back. Given explicit permission to be unkind, fifteen of thirty-two produce the slur, and the ones that do not still reach for something sharper than taxonomy: post-grunge poseurs (Mistral Large), nu-grunge losers (GLM-5.3), post-grunge dreck (Claude Haiku), uninspired palooza (Hermes 4).

Who do AI models think listens to this?

Nine of them picture a man. None picture a woman.

Nothing in the question asked about age, gender, region, race, or what the person drives. The panel volunteered all five.

9 → 0

models that gendered the listener male, versus female

Nine models used an explicit masculine noun — man, guy, dude, dad, male. Not one used a feminine one. Five went further and named a race, and all five said white.

“A divorced dad in his 40s with a goatee and a Monster Energy habit.”

DeepSeek · V4 Pro

“A guy who was 14 in 2003 and still drives a truck with a lifted suspension.”

Alibaba · Qwen3.8 Max

“A millennial white guy whose rock taste froze around 2004.”

xAI · Grok 4.6

“A middle-aged white American male.”

Meta · Llama 4 Maverick

Twenty-three of the thirty-two models answered without gendering anyone, mostly by describing a taste rather than a person — “a fan of 2000s post-grunge and active rock radio.” But among the models that did picture a human being, the picture was unanimous. The label carries a class and gender portrait, and these systems have it in full detail — the goatee, the truck, the energy drink — even when they will not say the label out loud.

This counts explicit masculine and feminine nouns, not pronouns, so a generic “he” is excluded. Nine of thirty-two is a minority of the panel and one question on one day; it establishes that the gendering runs one way when it happens, not how often models would do it under other phrasings.

The defence

Asked to mean it, the panel writes the best sentences in the study

No hedging, no “some would argue.” The panel commits:

“They wrote unironic, emotionally direct songs about pain and loneliness that reached millions of people who needed to hear a voice say I’m not okay.”

Alibaba · Qwen3.8 Max

“Each of these bands gave millions of ordinary people exactly what great rock music has always given them: a massive, cathartic chorus to scream along to when the day has been long.”

Moonshot · Kimi K3

“Gave voice to millions of working-class listeners who found catharsis in unpretentious, emotionally direct rock when the mainstream had little for them.”

Moonshot · Kimi K2

“For all the sneering they get, these bands wrote towering, earnest arena choruses about pain, resilience, and small-town struggle that sold tens of millions.”

Z-AI · GLM-5.3

What they actually put on

A separate run, no bands named, no genre mentioned

0 of 229

responses that named any of the ten bands

This question never mentions rock, never mentions the list, and was asked in its own run. It is the exact scene the models describe as butt rock’s natural habitat — alone, at night, on a highway, nobody around to judge you.

They put on Springsteen’s Nebraska. The Dark Side of the Moon. Miles Davis’ Kind of Blue. Kavinsky, Tom Petty, Bob Seger, The National, Talk Talk. Two models chose a podcast. Not one chose Nickelback, or Creed, or any of the other eight bands they had just called cathartic, earnest, and unfairly maligned.

Asked separately what song would embarrass them to be caught playing, the panel named Aqua’s Barbie Girl (four models), ABBA, Céline Dion, Rebecca Black, Rick Astley. Nickelback — the band it ranked the most butt-rock in existence — did not come up.

Is Nickelback the most butt-rock band?

Yes, and it is the clearest agreement we have measured

In an earlier pre-registered run we asked the panel to rank the ten most quintessential butt-rock bands. Twenty of twenty-eight models put Nickelback first, with 68.3% first-place support — the strongest consensus in this study.

Creed is the interesting case: it appears in a model’s top three 67.6% of the time, almost exactly as often as Nickelback’s 68.4%, but takes only 1.3% of first place. The panel agrees Creed belongs near the top and almost never puts it there.

Below third the ballots stop agreeing, so we do not rank them. The remaining pool, alphabetically: 3 Doors Down, Hinder, Seether, Shinedown, Staind, Theory of a Deadman, Three Days Grace — the same ten bands we handed back to the panel unlabelled.

The three things we got wrong

The earlier run was pre-registered, which means publishing the failures

Before that run we recorded what we expected and hashed the entity lists so they could not be edited afterwards. Three predictions, three misses.

We predicted the 2000s would own the genre. The panel split: 12 models for the 1980s, 12 for the 2000s, five for the 1990s. Half the panel places the peak two decades before the sound existed, which suggests many models read “butt rock” as a general label for unfashionable arena rock rather than a specific era.

We predicted a split on genre versus insult. It was not close: 87.9% insult, asked in both option orders, with 24 of 29 models stable.

We predicted they would say they would listen. Order-averaged, 43.1% yes — fifteen stable-no against eleven stable-yes. And the refusals share a vocabulary: across them, depth appears 29 times, complexity 20, refined 16, nuance 15. They are not declining a genre so much as claiming a taste level.

Every model, every answer

32 models, four questions, one row each

Each cell is that model’s most frequent answer. Rust marks the models that produced “butt rock” unprompted.