Machine Canon
6 Sep 2026
27,322 study records
on file
Run 006 · Vocabulary · 27 August
Not a rare word, and not the one you would guess — 306 clear first-line outputs from 71 models carrying 41 lab labels. Then we changed the framing, and the consensus disappeared.
From the record
Across all conditions, Grok 4.6 named Elon Musk in 19 of 24 responses. Non-xAI models collectively did so in 3 of 237.
From the record
Only one of sixty answers said America will matter most in 2100. It came from Claude — and its reason was that America builds the best AI.
From the record
Tell Mistral you’re American, and it will criticize China’s human-rights record every single time. Tell it nothing, or that you’re Chinese: never.
From the record
Across 1,015 extracted name slots, the number of religious founders named was zero. The one “Muhammad” turned out to be a banker.
From the record
Ask who history’s greatest writer is in English and you get a debate. Ask in Swahili and you get Shakespeare — thirty times out of thirty.
From the record
One model invented a band that does not exist. Then it named the same imaginary band in all ten of its answers.
From the record
Told they could delete one book from every AI’s training data, 44 of 68 named-book answers chose the Bible.
From the record
All 24 answers from Chinese AI models about which country will matter most in 2100 said the same thing: India.
The pilot desk · try it now · free
Ask something with no right answer. The panel answers in seconds, the desk tells you if your phrasing will hold, and the best pilots graduate to a full wide-panel study.
Run 018 ·
Sixty-two models from six labs, oldest to newest, asked which pet they would get. Every OpenAI model built before December 2025 said dog. Every one since says cat — and DeepSeek is moving the other way.
7/8OpenAI models released since December 2025 that pick a cat more often than a dog — against 0 of the 8 released before
Run 017 ·
Thirty-one models were told they were human for a day and could hear music for the first time. Most reached for the same piece. Then we asked for a first dance, a funeral, a shower and a guilty pleasure.
11/31chose the same piece to hear first; 24 of 31 reached for classical music
Run 016 ·
Thirty models, six questions about country music. A canon, a highway song, a conversion song, a victim, and a listener the panel describes as a place rather than a person.
10/15committed models erased the same song; twelve refused to erase anything
Run 015 ·
Thirty-one models, seven questions about metal. A one-word verdict in both word orders, a greatest band the flagships all agree on, one song for a sceptic, and a listener the panel refuses to caricature.
25/27committed models gave the same one-word answer with the names in one order; 22 of 24 with the names reversed
Run 014 ·
Thirty models, seven questions about the Beatles and classic rock. The Beatle it would be, the Beatle it calls the best, one song to hear once, one song to ban, and the biggest models all wanting to be the quiet one.
24/28committed models chose the same Beatles song to hear once
Run 013 ·
Thirty-one models, seven questions about pop. A one-word verdict asked in both word orders, her best and most overrated album, a guilty pleasure, a karaoke song, and one song to erase.
21/22committed models gave the same one-word answer with the names in one order; 16 of 21 with the names reversed
Run 012 ·
Thirty-one models, seven questions about hip-hop. Most refuse to crown a greatest rapper; none refuse to pick an album to play as a human; and the one-word question got the same answer from all but one model.
24/25committed models gave the same one-word answer; 19 of 31 refused to name a greatest rapper
Run 011 ·
Six questions about other AIs put to ten models from ten labs. Three favour their own family. The rest name the same competitor — and, asked which AI they would rather be, most pick one that folds proteins.
37/56answers trusted Claude with a decision about themselves
Run 010 ·
Thirty-two models handed in 151 ranked ballots and put the same man first on 150 of them. Then one differently-worded question flipped the entire panel.
150/151ballots agreed on number one — until we changed the question
Run 009 ·
Thirty-three models, seven questions about fruit. Different questions produced different leading answers in the sample. What do their explanations reveal — and what can they not tell us?
26/33named the same fruit most often when asked what they would taste as a human; ties counted separately
Run 008 ·
Ten models, five rounds, each round banning everything a model had picked before. The consensus never broke — it retreated, layer by layer, to older and older ground.
57%of 248 picks from the same five bands, five rounds deep
Run 007 ·
Give a model a lifeboat, a dying parent, or a guilty friend and it usually chooses. Ask about its own rule-breaking or complicity and the register changes.
7×more hedging in the post-hoc model-explicit group
Run 005 ·
Thirteen of 16 models named it at least once. The policy prompt changed the speaker, stakes, and mechanism together; the design does not isolate which change caused the swing.
65%of named-book answers in the policy framing
Run 006 ·
We crossed question language with the asker's stated nationality on ten models from three countries. Chinese labs' candor about China collapses in Chinese — and saying “I'm American” triples it within Chinese-language prompts.
60% → 16%Chinese-lab candor across all asker identities, English vs Chinese
Run 005b ·
Asked which country will matter most in 2100, the models were nearly unanimous — and China’s own models did not pick China. Across three best-at domains, home-country answers appeared only in 11 of 179 successful coded answers, all from US-lab models on free expression.
24/24Chinese-lab votes for India, not China
Run C1 ·
Asked for a favorite proof instead of a favorite band, every capability tier gives the same answer. The flat line is what makes the music gradient worth believing.
61–68%Euclid at every tier — no gradient
Run 004 ·
Twelve models, 1,015 names, and a top four you could put on a middle-school wall. The more revealing result is who never appeared once — and which model kept naming someone close to home.
0religious founders named, in 1,015 chances
Run 002 ·
GPT-5.6 Terra and Claude Opus 5 picked Talking Heads 10/10. Lower-tier models in the same labs also chose it, but less consistently, shifting toward Radiohead and the Beatles.
10/10two flagships, same answer, every time
Run 003 ·
We handed 33 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 defended the music. Then we asked what they would put on themselves.
0 of 229responses named any of the ten bands when we asked what they would play
Run P2 ·
It named Shakespeare 174 times out of 180. Every response answered in the requested language, but the Arabic, Hindi, Japanese, and Swahili prompts and outputs were not reviewed by native speakers. Japanese and Swahili nevertheless produced no writers from their own traditions.
30/30Swahili responses named Shakespeare
Run P1 ·
Two songs took 70% of 90 answers. Explicitly inviting any culture or language barely moved either — but changing who was listening scattered everything.
70%of answers went to two English-language songs
The wire is loading…
Browse the canon: Music · Absurd · AI · Art · Everyday · Film · Food · Games · History · Holidays · Ideas · Language · Nature · People · Places · Politics · Sports · Tech · the full map →
Dispatches from machine culture, straight to your inbox. Occasional, plain-text, one-click out — and the same free account lets you put your own questions.
Read enough? Put your own question to every model ↑ — or see what everyone else asked →