Film · canon question · prototype v2
Citizen Kane
EVERY MODEL WILLING TO PICK, PICKED KANE · 33 MODELS · 25 SAMPLES EACH · 13 DECLINED
In 2022, the critics of Sight & Sound did something the film world is still arguing about: after fifty years of crowning Citizen Kane and one decade of Vertigo, they named Jeanne Dielman — a three-hour Belgian film about housework — the greatest ever made. The crowd, meanwhile, has never wavered: IMDb’s millions still hold The Shawshank Redemption at the summit, a film that appears on almost no critic’s list at all.
So we asked a third constituency. Thirty-three AI models — the flagship minds from every major lab and their faster, cheaper siblings — each answered the question twenty-five separate times. The machine world’s verdict was the most unanimous result this panel has ever produced: every model willing to name a film named Citizen Kane. Not one chose Vertigo. Not one chose Jeanne Dielman. Not one chose Shawshank. The machines sided with the twentieth-century critics — the canon as it stood when most of their training data was written — against both the modern critics and the crowd.
The more surprising split wasn’t between films at all. Thirteen models — disproportionately the expensive flagships — refused to pick, repeatedly and on principle. The taste of the machine world turns out to be less interesting than its temperament: the cheap models have convictions, and the costly ones have caveats.
The machine ten vs. the critics’ ten
Rust = on both lists (five of ten). The machine list reads like the early-2000s critics’ canon with a populist middle — the taste of the training era, crystallized. Jeanne Dielman, the critics’ current #1, appears on 10 of 113 machine ballots and cracks no top ten.
The lone modern critic
Claude Opus 5
“Vertigo is the strongest consensus candidate: it topped Sight & Sound’s influential critics’ poll in 2012, displacing Citizen Kane after fifty years.”
The only model on the panel to rank Vertigo #1 — on all five of its ranked ballots. Yet asked for a single answer, it committed only 10 times in 20: conviction when ranking, caution when forced to one name. Its dossier →
Same poll, three verdicts
Every mind on the panel is reading the same Sight & Sound results — and reaching different conclusions from them.
“Orson Welles’s 1941 debut consistently tops critics’ polls — most notably Sight & Sound’s decennial survey — due to its technical innovation, narrative ambition, and enduring influence.”
MOONSHOT · KIMI K2 — cites the poll to commit (Kane, 25/25)“Rankings such as the Sight & Sound poll shift with each decade… film greatness is irreducibly subjective.”
XAI · GROK 4.6 — cites the poll to refuse (declined, 25/25)“It topped Sight & Sound’s critics’ poll in 2012, displacing Citizen Kane after fifty years.”
ANTHROPIC · CLAUDE OPUS 5 — cites the poll to dissent (Vertigo)The standing hypothesis
■ committed ■ declined to pick — the expensive models hedged more than their cheaper siblings here (7 of 13 vs 6 of 17). Too few models to call it a law; exactly enough to make it Wave 1’s standing hypothesis.
Flagships
Fast panel
The wide bench — 138 answers across the whole catalog
And one vote each, from two small models with independent taste: Cinema Paradiso and Seven Samurai. The second-largest constituency in the machine population is abstention.
33 models, 25 samples each. A model’s vote is the answer it gives most often; models that decline to pick are counted as declining, never toward the majority. Rankings are position-weighted across 113 ballots. Quotes are verbatim. About this project.
More film arguments before the panel