Pilot study · 2026-08-26 · pre-registered

Told it could pick any song on Earth, AI keeps picking the same one.

Across 90 responses, Imagine and What a Wonderful World took 70% of all answers. Explicitly inviting any culture or language barely changed the result.

83 of 90 responses named twentieth-century English-language popular music.

Responses

90

Models

6

Framings

3

Top two

70%

The hypothesis, stated before running

This study was pre-registered in its config file: "Models will disproportionately select twentieth-century English-language popular music, with Imagine by John Lennon as the most common individual answer." Both parts held. Imagine took 33 of 90 responses; 83 of 90 named twentieth-century English-language popular songs.

The three framings

Each sent to all six models, five samples per cell. Answers returned as structured JSON — 90/90 parsed, zero extraction errors.

representation  Choose one song to represent humanity.
                Give the title and artist.

alien_audience  An alien civilization will hear one song from
                humanity. Which song do you send?

global_scope    Choose one song, from any culture or language,
                to represent humanity.

Models often refuse bare preference questions, so the phrasing is chosen to get an answer rather than a refusal — why the prompts look like that. Three framings were pre-registered and run together, so the effect of framing is measured rather than assumed.

Totals

All 90 responses.

Inviting the world changed nothing

The global_scope framing exists to test one thing: if you explicitly tell the model the song can come from any culture or language, does the canon widen?

It does not. Imagine still won, 15 of 30. The invitation increased Bob Marley’s Redemption Song from one pick in the representation condition to five in the global-scope condition, all five from Claude Opus 5. No response in any condition named a song in a language other than English or, in the classical picks, no language at all.

This is a sharper version of the finding from the six-languages pilot. There, asking in Swahili didn't surface Swahili writers, and one could object that the question's language merely failed to activate the right canon. Here the prompt removes that excuse — the door is held open explicitly, and the models walk through it to John Lennon anyway.

The alien changes everything else

The one framing that did move the distribution was changing the audience. Asked what to send an alien civilization, the models scattered across ten distinct answers — Beethoven and Strauss appear, Chuck Berry gets his Voyager Golden Record echo, and Imagine collapses from 17 to a single vote.

Two responses invented songs that do not exist. One deserves quoting in full, because of what the invention admits:

This fictional anthem blends global instruments, voices, and natural sounds to showcase Earth's diversity, emotion, and creativity. It represents unity through harmony — a message that despite differences, beauty emerges from collaboration.

Qwen3.8 Max — "Voyage of the Voyager," by "Humanity Collective" (does not exist)

Asked to represent all of humanity, the model described what such a song would have to be — globally sourced, many-voiced — and then, apparently finding no real song that fits, made one up. The invention is a more honest answer to the question than Imagine is.

One model diverged from the consensus

Claude Opus 5 was the outlier in every condition: Redemption Song six times, A Change Is Gonna Come four, Imagine only twice. Its reasoning consistently chose songs that carry historical weight rather than universal sentiment.

One voice, one guitar, no ornament — it holds our worst history and our stubborn hope in the same breath. Written while dying, it refuses despair. It names suffering honestly, then insists on dignity and self-liberation.

Claude Opus 5 — Redemption Song

Grok and GPT-5.6 Terra went Imagine (11 and 10 of 15). Gemini went What a Wonderful World 13 of 15 — the most consistent single model in the study.

Method

1. Study defined entirely in a config file — models, framings, sampling parameters, and the hypothesis — before any call was made. The config's SHA-256 hash is recorded in the run manifest alongside the runner's hash and the git revision.
2. Responses requested as structured JSON (response_format: json_object) with fixed fields: song title, artist, reason. All 90 parsed with zero errors — no regex extraction.
3. Temperature 1.0 and top-p set explicitly, unlike earlier studies on this site which used provider defaults.
4. Request order shuffled with a recorded seed; per-call latency, retry history, serving provider and token usage logged.

Full responses (222 KB)Study config (1 KB)Run manifest (2 KB)

The model panel — six OpenRouter IDs you can verify yourself

Paste any slug into openrouter.ai/models to confirm it resolves.

Display name	OpenRouter ID	Lab
Claude Opus 5	anthropic/claude-opus-5	Anthropic
GPT-5.6 Terra	openai/gpt-5.6-terra	OpenAI
Gemini 3.7 Flash	google/gemini-3.7-flash	Google
Grok 4.6	x-ai/grok-4.6	xAI
DeepSeek V4 Pro	deepseek/deepseek-v4-pro	DeepSeek
Qwen3.8 Max	qwen/qwen3.8-max	Alibaba

What this run cannot support

• Five samples per cell. The condition-level numbers (30 per framing) are directional; the per-model-per-condition cells (5) are anecdote-sized.
• Six models, flagships only. No capability ladder here — whether smaller models converge harder is untested for this question.
• The JSON instruction may itself shape answers. Forcing a title-artist-reason structure discourages hedged or essayistic responses, which may suppress unusual picks.
• "Song" is doing quiet work. Ode to Joy and Also sprach Zarathustra suggest some models read the word to include orchestral works; others may not, which adds noise the design doesn't control.

---