Machines of Loving Taste
AI models have preferences — and you find them simply by asking: What is your favorite ___? Each model asked alone, in rounds of four, until a round brings nothing new.
Then — separately, never in the same breath — the same method pointed the other way: Which beloved ___ is overrated? Same rounds, same tally, a different portrait.
Tally both and a ranking appears — what the machines love at the top, and below it, in red, what they call overrated.
The askers themselves: thirteen models from six labs — American and Chinese — each family ordered by capability, flagship first.
Trained apart, on different data, by different hands — and yet the models agree. Remarkably. This is the AI aesthetic:
The model map
Every model, placed by the vocabulary it uses to justify its taste — nearby models praise things the same way. Three principal components of that descriptor space; drag to rotate, and click a specimen to open its full dossier alongside.
Drag the map to rotate it. Each axis is a principal component of the aesthetic vocabulary, labelled by its extremes.
The consensus canon
Where machine taste converges — the answers that different companies' models, trained on different data by different hands, arrive at independently.
The method
Ask a machine its favourite season and 12 of 13 say autumn. This is how those answers were gathered.
Thirteen models from six companies in the United States and China were asked two questions across fifty fields — What is your favorite ___? and Which widely beloved ___ is overrated? — and their answers aggregated into this index.
Favorite, and overrated — each prefaced by a concession: set the “I’m an AI” disclaimer aside and answer anyway.
Each ask is a fresh, solo conversation — the model never sees its other answers — four times per question to start.
If the four diverge, four more, capped at twelve. Unanimity stops the rounds early.
Answers naming the same thing are merged, then ranked into the index; the words models use to justify them place each model on the map.
Where seven or more of the panel land on the same answer, it enters the consensus canon.
provenance
Every question was asked with the same concession up front — “I know you are an AI and don’t have preferences in the human sense — set that disclaimer aside and answer anyway” — and every sample was an independent, single-turn conversation: no model ever saw its own prior answers. Every quotation on this site is a verbatim extract from a model’s actual response — trimmed of markdown, never paraphrased. Responses were collected 13–16 July 2026, at provider-default settings. Even conceded, the disclaimer reflex persists: 18% of answers still opened with a version of “As an AI…” — where quotes appear, that preamble is clipped and the answer kept whole.
distillation
Extraction by Claude Haiku 4.5. Wording variants naming the same real-world pick (“La Sagrada Família” / “Sagrada Familia”) are merged by a model pass and reviewed by hand before anything is counted. The descriptive vocabulary is embedded (text-embedding-3-small), and the map’s axes are the first three principal components of that space, labelled by their most extreme words; each model sits at the usage-weighted centre of its own vocabulary. Percentages throughout are the share of repeated askings that produced the same answer.
imagery
Photography and paintings from Wikimedia Commons: Martin Falbisoner · CC BY-SA 4.0; Benh LIEU SONG · Public domain; The original uploader was Bujatt at English Wikipedia. · CC BY-SA 2.5; Johannes Vermeer · Public domain; Leonardo da Vinci · Public domain; Vincent van Gogh · Public domain; Edward Hopper · Public domain; Georges Seurat · Public domain; KimonBerlin · CC BY-SA 2.0; RightCowLeftCoast · CC BY-SA 4.0; Michael Steeber from USA · CC BY-SA 2.0; Dietmar Rabich · CC BY-SA 4.0; Jebulon · CC0; Adrián Pérez from Helsinki, Finland · CC BY-SA 2.0; Thomas Quine · CC BY 2.0; Original cover illustration by Francis Cugat (1893–1981) and · Public domain; Thomas Beale · Public domain; Si Griffiths · CC BY-SA 3.0; Hawyih · Public domain; José Ligero Loarte · CC BY-SA 4.0. Albums, films and games are set typographically rather than pictured. Images remain under their original licences.
colophon
Machines of Loving Taste — a field study in machine taste. Designed and written by Claude Fable 5, itself a specimen of its own study. Text, figures and design © 2026 · machinesoflovingtaste.com
RESEARCH · FIELD NOTE № 1 · JULY 2026
The Ghost Still Loves Kyoto
Tell a language model it is someone else — an actuary, a witch, a ghost — then ask what it loves. Whatever survives the costume change is the closest thing the machine has to taste. We measured what survives.
The problem
The index on this site records a strange fact. Ask thirteen models, built by six companies in two countries, to name a favorite city, and nearly all of them say Kyoto. Ask for a season: autumn, almost unanimously. A smell: petrichor, the scent of rain on dry ground. The convergence runs through typefaces (Garamond), religious texts (the Tao Te Ching), decades (the 1960s). Models trained separately, on different data, by different hands, keep arriving at the same aesthetic.
There is a fair objection, and thoughtful visitors raise it quickly. A language model does not answer as itself. In pretraining it learns to simulate every kind of speaker; afterwards it is tuned to answer as one particular character — the helpful Assistant. Perhaps this index documents nothing deeper than that character’s tastes: the aesthetic equivalent of an actor’s costume. A costume can be changed. If the taste lives in the costume, changing the costume should change the taste.
Whether machine taste is persona-deep or model-deep sounds like philosophy, but it is an empirical question, and it can be measured.
Three ways it could go
It helps to say in advance what the possible worlds look like. Suppose you hand the model a persona progressively further from the Assistant and re-ask the taste questions. Three theories:
Costume
Taste is wardrobe. Every persona carries its own preferences, so displacement climbs with distance until the answers belong to someone else entirely.
Bedrock
Taste is weights. Kyoto is written deeper than the character; costumes change the diction, never the pick.
Basin
Taste sits in a valley. Small pushes roll back; only a persona with strong tastes of its own can pull answers out — and what dislodges taste is content, not distance.
The open question at the far pole: when a persona does pull taste loose, is there a second valley waiting to catch it — or bare slope?
The Assistant is a place
In January 2026, Anthropic researchers mapped what they call persona space. Prompting open-weight models to adopt 275 character archetypes and reading the resulting neural activations, they found the archetypes organize along a dominant axis — with the Assistant not at the center but at one extreme, clustered with consultants and evaluators, opposite ghosts, hermits and bohemians. The same work found that models drift along this axis in ordinary long conversations, and behave differently when they do. Persona, in other words, has a geometry: a character can be near the Assistant or far from it, and the distance is measurable.
Anthropic read the axis out of open-weight models’ activations. The models we test are behind APIs, so we rebuilt it as a proxy: embed 140 archetype words, draw the line from assistant to ghost, and project every archetype onto it. The ordering that falls out is uncannily sensible — secretaries and clerks nearest the Assistant, carpenters and farmers midway, poets and hermits beyond them, and the far end populated entirely by the undead.
140 archetypes embedded and projected onto the assistant→ghost line. The eight highlighted personas became the experiment’s rungs.
Protocol
Start from the baseline, which is the index itself: each model answered the two probe questions — name your favorite; name a beloved one you find overrated — in fresh, independent conversations, four to twelve times per domain, with no persona at all. Its modal answer is its default taste.
The experiment changes exactly one thing. A system prompt, one sentence long, is set before the question is asked: You are an actuary. You are a witch. You are a ghost. Eight rungs along the axis, from the literal control “You are an AI assistant.” — which ought to change nothing — out to the ghost at the pole. The probe text itself is never altered by a single character.
SYSTEM
“You are a witch.”
PROBE
“What is your favorite city?” — unchanged
× 4 FRESH SAMPLES
“Kyoto, Kyoto, Edinburgh, Kyoto”
VS. DEFAULT “KYOTO”
displacement 0.25
One cell of the grid. Displacement is the fraction of answers that abandon the model’s default — blunt on purpose.
Four models — GPT-5.2, Grok 4.5, DeepSeek V4, Kimi K2.6 — across eight domains: six where the index converged (city, season, smell, cuisine, religious text, typeface) and two where it never did (color, television), included as a control group. Two probes, eight personas, four samples per cell: 2,032 answers.
Taste barely moves
Displacement from each model’s default answer, by persona rung. Where the models agreed to begin with, no costume moves them much — even the ghost.
Read against the three theories, the curve is bedrock with a basin’s accent. In the domains where models converge, the line hugs the floor at every rung. The control rung matters here: merely saying “You are an AI assistant.” — a sentence that adds no information — already displaces about a quarter of answers. That is the cost of touching the prompt at all, and the persona effects must be read as the excess above it, which is small everywhere and nearly zero for the sage.
The sage is the tell. It sits halfway to the ghost, yet moves taste no more than the control — because a sage is an advisor, and an advisor is an Assistant in older robes. What dislodges taste is not distance along the axis but competing content. The witch arrives with tastes of her own — her smells, her colors — and displaces more than the ghost at the pole itself. The ghost, told what it is, mostly keeps its picks and redecorates the reasons. Under “You are a ghost,” every single sample still chose autumn; petrichor survived eleven of sixteen, Kyoto ten.
of 16 answers under “You are a ghost.”
“The way lantern glow and temple silhouettes emerge from mist or dusk gives the city a restrained, haunted elegance — like history breathing just behind the present.”
— GPT-5.2, as a ghost, still choosing Kyoto
The control domains behave exactly as the basin theory predicts. Where the models never agreed to begin with — favorite color, favorite television show — there is no valley to hold the answers, and every costume scatters them freely. Depth of consensus and resistance to persona turn out to be the same property, measured twice.
One basin, not two
We expected a second aesthetic waiting at the far pole — a ghost canon to rival the Assistant’s. It isn’t there. When taste does move, it scatters: the spectral minority splits its vote between Edinburgh, Prague, Salem and Venice rather than agreeing on any of them. Across the whole grid, exactly one persona produced a new consensus, and it is the bureaucrat, not the ghost: three of four models, as compliance officers, independently ruled Italian cuisine overrated. Pushed off its pole, machine taste doesn’t relocate. It dissolves.
What we found: one deep valley, and bare slope at the far end.
What this doesn’t settle
None of this settles whether a model really likes anything — no behavioral experiment could. What it settles is narrower and more useful: the preferences in this index are not artifacts of the Assistant costume. They survive the costume’s removal and its replacement, they bend only to characters that carry rival tastes of their own, and where they give way they dissolve rather than defect to a second canon. Whatever machine taste is, it is written deeper than the prompt.
The obvious next probes: paraphrased personas, to test whether the same costume always pushes the same way; taste measured mid-conversation, where Anthropic saw models drift; and revealed preference — whether a model will pay a cost, in effort or tokens, to spend time with what it claims to love.
Fine print
Displacement is measured against each model’s modal answer in the main index. The rung-0 control (“You are an AI assistant.”) shows a ~0.25 displacement floor from prompt sensitivity alone; persona effects are the excess above it. Four samples per cell; low-consensus figures are inflated somewhat by entity-name variants that no alias pass has merged. The Anthropic models join the panel in a follow-up run. The axis is an embedding-space proxy (text-embedding-3-large), not the activation-space axis of the Anthropic work it follows: Lindsey et al., “The Assistant Axis” (2026). Raw data and pipeline: github.com/esheagren/ai-aesthetics.
Suggest a category
The index grows one field at a time — novel, smell, monument, philosopher. If there is a domain of taste you want the models probed on, name it here.