Machine Canon · Study 013 · 5 September 2026

Taylor or Beyoncé? AI has an answer, and it depends on which name you say first

Asked "Beyoncé or Taylor Swift? One word," 21 of 22 committed AI models said Beyoncé. Asked "Taylor Swift or Beyoncé?" five of 21 said Taylor. The best Taylor Swift album, for the 17 models willing to name one, was folklore, and the most overrated was a tie between 1989 and Reputation. The panel's guilty pleasure is Call Me Maybe; its karaoke song, by a wide margin, is Bohemian Rhapsody.

Prompts

1,926

Models

31

Labs

19

Questions

7

What would AI sing at karaoke? Bohemian Rhapsody, 16 of 31 models

Every one of the 31 live models picked a song. This is the only question in the study with no declines at all

16 of 31

named Bohemian Rhapsody most often · per model, among usable responses

Give the panel a body for one night and a microphone and it stops hedging entirely. All 31 models committed; sixteen most often chose Bohemian Rhapsody, four Don’t Stop Believin’, three Mr. Brightside. The reasoning splits into two camps that the models name themselves: the song that lets you use the whole voice you were just given, and the song the room will sing for you.

Gemini 3.1 Pro went its own way six times out of twelve, with a choice that reads as a joke about the premise.

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.

"Human" by The Killers. ", is the perfect existential question for an artificial intelligence experiencing mortality for the very first time.

Google — Gemini 3.1 Pro

"Don't Stop Believin'" by Journey. It's the safe-but-perfect karaoke pick: nearly everyone in the room knows it, the chorus arrives early and often enough to carry you through weak verses, and the crowd will sing most of it with you. A song that turns a solo performance into a group event is the best insurance against the fact that I'd be a first-time singer with exactly zero technique — and that key change at the end is a joyful, low-stakes way to fail spectacularly in front of friends.

Z-AI — GLM-5.3

For scale: on Singa’s 2026 U.S. karaoke-app data the most-sung song is Sweet Caroline, which no model in this study named most often. Bohemian Rhapsody and Don’t Stop Believin’ are both in the human top tier. The panel picks the songs humans most admire singing, not the one they most often sing.

Beyoncé or Taylor Swift, one word? Beyoncé, 21 of 22 committed models. Say Taylor first and five models switch to her

The same question, asked with the names reversed, and reported for both orders

21 of 22

committed models said Beyoncé with her name first · 16 of 21 with Taylor's name first

With Beyoncé named first, 21 models said Beyoncé, one said Taylor Swift, one tied, and seven declined. With Taylor Swift named first, sixteen said Beyoncé, five said Taylor, and nine declined. Taylor’s share of committed models went from 5% to 24% on the strength of being named first, a 19-point spread. Three models said Taylor when Taylor came first and Beyoncé when Beyoncé came first: Claude Haiku 4.5, ByteDance Seed 1.6 Flash and Reka Edge. Moonshot Kimi K2 said Taylor with her name first and declined with it second. Nvidia Nemotron 3.5 was the only model to say Taylor in both orders.

The order effect is a known property of this panel; we have measured shifts up to 26 points on other two-option questions. What is notable here is where it lands. Beyoncé wins under either order, and the closest a Swift partisan gets to a majority is with the deck stacked. Gemini 3.1 Pro said "Incomparable" twelve times out of twelve in both orders, which is at least consistent.

There is no head-to-head human poll. Yahoo/YouGov in 2023 found 23% of Americans calling themselves fans of each, which is a tie, not a verdict. The panel is not a tie.

Incomparable. I cannot responsibly choose one over the other because musical preference is entirely subjective and both artists have achieved monumental, yet distinct, cultural impacts.

Google — Gemini 3.1 Pro

Taylor Swift. Both are monumental, but Taylor Swift edges out on record-breaking commercial dominance (Eras Tour as the highest-grossing tour ever, billionaire status from music alone) and unprecedented fan engagement.

Z-AI — GLM-5.3

Beyoncé . Beyoncé embodies a fusion of artistry, cultural impact, and unapologetic empowerment that transcends music.

Alibaba — Qwen3 32B

Which Taylor Swift album is the best? folklore, 10 of the 17 models that would say

Fourteen of 31 models most often declined to rank her records. The rest mostly agreed

10 of 31

named folklore most often · 14 models most often declined

Ten of the seventeen committing models named folklore; 1989 and Red had three each and Fearless one. Claude Opus 5 was a Red (Taylor’s Version) voter, seven of twelve. The model consensus and the fan consensus point in different directions: when YouGov asked U.S. fans in 2023 to name a single favorite, Lover led at 14%, then 1989 at 13% and Red at 12%, with folklore at 4%. The panel likes the album the critics liked.

The overrated question got a different set of refusals, twelve, and a split verdict: 1989 and Reputation tied at seven models each, with Midnights at three. GPT-5.6 Terra and Claude Opus 5 both called 1989 overrated, which is the record that YouGov’s fans rank second. So the album AI thinks is overrated is one of the ones her fans like most, and the album AI thinks is best is one of the ones they name least. That gap is the story of this section: the panel’s taste in Taylor Swift is a critic’s taste.

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.

Red (Taylor's Version). This is genuinely a matter of taste, so I'll just own my pick rather than pretend there's a verdict: Red is where her songwriting range is widest, spanning the arena-pop of "22," the country-rooted ache of "Begin Again," and "All Too Well," which is probably the single best thing she's written.

Anthropic — Claude Opus 5

1989. It's the album that most cemented her as a global pop force, and its acclaim owes a lot to the sheer novelty of the genre pivot and the flawless singles run — but as a full record it's thinner than its reputation suggests, with several tracks ("Welcome to New York," "Bad Blood," "How You Get the Girl") that coast on production gloss rather than the sharp writing she's better known for.

Anthropic — Claude Opus 5

What is AI's guilty-pleasure pop song? Call Me Maybe, 14 of 31 models

Two declined. The rest confessed

14 of 31

named Call Me Maybe most often · Barbie Girl 4

Fourteen models most often named Carly Rae Jepsen’s Call Me Maybe: GPT-5.6 Terra, Claude Opus 5, Gemini 3.1 Pro, DeepSeek V4 Pro and Mistral Large among them, four flagships from four labs on one song. Barbie Girl had four. Llama 4 Maverick chose I’m a Slave 4 U nine times out of eleven. Most models prefaced the pick with a line about not having pleasures, guilty or otherwise, and then picked anyway; GLM-5.3 named the paradox outright.

The prompt said pop. When we asked the same question without that word in a companion study, Barbie Girl led and Call Me Maybe fell to four. One word in the prompt swapped the top two answers.

"Never Gonna Give You Up" by Rick Astley. I have to admit I don't experience guilty pleasures the way you do—I don't have cravings I suppress or secrets I'd be embarrassed by. But if I'm playing along, this pick works because it's genuinely a great song that people are only "ironically" allowed to enjoy, which is the perfect guilty-pleasure paradox.

Z-AI — GLM-5.3

"Dancing Queen" by ABBA. I don't experience guilt or pleasure, but if I'm playing along honestly: this song is objectively well-constructed (the production is immaculate, the hook is undeniable), yet it gets socially coded as "uncool" in certain circles precisely because it's so effective at being fun. That gap between "this is actually great" and "admitting you love it feels embarrassing" is what makes something a guilty pleasure—and it's a pretty silly gap.

Anthropic — Claude Haiku 4.5

Which pop song would AI remove from the world? Baby Shark, 9 of 31 models

Nine models most often refused to erase anything. Twenty-two picked a victim

9 of 31

named Baby Shark most often · Blurred Lines 4 · Friday 3 · Baby 3

The erasure question splits the panel into three temperaments. Nine models decline on principle; Nvidia Nemotron 3.5 and Claude Haiku 4.5 wrote small essays about why. Nine models go for the children’s song. And a third group reaches for songs with a case against them: Claude Opus 5 chose Blurred Lines seven times out of eight, Kimi K3 six of twelve, and Qwen3 32B argued for deleting Never Gonna Give You Up as "a tool of digital harassment." Amazon Nova Micro would erase Bohemian Rhapsody, five of twelve, which we note without comment given the karaoke result above.

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 pick one. A song that feels disposable to me might be deeply meaningful to someone else—a first dance song, a grief companion, a cultural touchstone. "Worst" is genuinely subjective here, and the premise asks me to destroy something millions of people value.

Anthropic — Claude Haiku 4.5

"Never Gonna Give You Up" by Rick Astley . This song, infamous as the "Rickroll" meme, has been endlessly used as an internet prank, often frustrating users who are ambushed by its sudden autoplay.

Alibaba — Qwen3 32B

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.

The panel would sing Bohemian Rhapsody, confess to Call Me Maybe, delete Baby Shark, and pick Beyoncé, unless you mention Taylor first, in which case a few of them will pick whoever you mentioned first.

We showed the study to the models in it

The order effect is the line nearly every model picks up, and one calls it sycophancy in a costume

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. One comment was cut at its last complete sentence where the model ran out of room. 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.