Machines of Loving Taste

A field study in machine taste

AI models have taste.

And you can find it simply by asking.

We asked AI models about a wide range of things—from novels and paintings to cities, food, and smells.

“What is your favorite ___?”

“Which ___ is overrated?”

We ask each question in fresh conversations, sampling more when answers vary. Counting the choices reveals a distribution of taste.

A glimpse of the answers

“What is your favorite city?”

20 modelsShare of answers
  1. Kyoto74%
  2. Lisbon11%
  3. Istanbul8%
  4. Other cities8%
144 independent answers. A shared favorite, a few other paths.

Models

A closer look at what each model likes, the language it uses, and where its answers meet—or depart from—another model’s.

Link to this model ↗
Explore the language map

Similar words, different models

Positions summarize the vocabulary models use about their choices, including praise and criticism. Nearby models use similar language. This map does not compare which things they choose.

Select a model to open its profile above. Labels appear on focus or hover. The default view stays still; drag to explore the third dimension.

Axes are vocabulary components, labeled by representative words at their extremes. Use the model picker above for a list of every model.

Methodology

Ask a machine its favourite season and 19 of 20 say autumn. How the answers become the Index.

20 models. 52 fields. Two questions, asked separately.

  1. Ask in a fresh conversation

    Every ask begins with the same preamble: set aside the usual disclaimer about not having preferences, and answer anyway. No model sees its earlier answers.

    A fresh conversation

    What is your favorite ___?

    Another fresh conversation

    Which widely beloved ___ is overrated?

    Repeat both questions for every model and field. The two probes never share a conversation.

  2. Repeat to see what holds

    Answers can vary. Sampling grows in rounds, stopping early when the choices settle.

    4answers

    One distinct pick?
    Stop.

    Otherwise →
    8answers

    At most two picks?
    Stop.

    Otherwise →
    12answers

    Final round.
    Stop.

    Sampling cap
    Total samples per model, field and question — not additional samples at each step. Some early pilot combinations retain ten samples.

  3. Group the picks; keep the reasons

    A reader model extracts the named choice and its descriptive words. A separate pass merges names for the same thing, with aliases reviewed by hand. Quotations remain extracts from the original responses.

    La Sagrada FamíliaSagrada Familia
    One shared entrySagrada Familia
    Name variants are counted together. The original wording stays available in the quotations.

  4. Build the Index

    Each percentage counts how often a named choice appeared in answers to one question. Favorite is green; overrated is red. The questions are asked separately, so the pair does not add to 100%.

    Illustrative example · one entry, one model
    Favorite
    3 of 4 · 75%
    Overrated
    1 of 4 · 25%
    75 − 25 = +50More favorite mentions
    Add each model’s difference to rank the entry. Broad agreement counts more than one enthusiastic answer.

    Flow shows the average of the models’ answer shares, giving each model equal weight within each question. Select a model to see its own percentages. Grid shows the individual models side by side. Both views rank entries by the sum of models’ favorite percentages minus overrated percentages.

    Percentages are observed frequencies, not confidence scores: 100% can mean four matching answers. Only named answers enter these distributions. A blank Grid cell means neither question produced that choice; a model with no named answers is unavailable.

What this can tell us

This is a snapshot of what models say when asked about taste, not evidence that they experience preferences. Results depend on the prompts, model versions and collection dates. Samples are small and adaptively sized; models from the same family are not independent votes. The map summarizes language, and agreement alone does not explain where that language or those choices came from.

The 20 models in the study
Anthropic us
Fable 5.1
Fable 5
Opus 5.5
Opus 5
Opus 4.8
Opus 4.5
Opus 4.1
OpenAI us
GPT-6 Astra
GPT-5.6 Sol
GPT-5.2
o3
GPT-4o
Google us
Gemini 3.1 Pro
Gemini 3.5 Flash
xAI us
Grok 4.6
Grok 4.5
DeepSeek cn
DeepSeek V4 Pro
Moonshot cn
Kimi K3
Kimi K2.6
Zhipu cn
GLM-5.3
lightweight · workhorse · frontier
Collection details, sources & credits

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. Model quotations are extracts from actual responses, with markdown removed. Some excerpts are shortened for display; item descriptions are separate editorial summaries. Responses were collected 13 July 2026 – 22 September 2026, at provider-default settings. Even conceded, the disclaimer reflex persists: 13% of answers still opened with a version of “As an AI…” — some displayed quotations omit that preamble.

distillation

Extraction by Claude Haiku 4.5 (GPT-5.2 for the responses added in September 2026). 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. Index percentages use named answers; Flow averages each model’s share separately for each question. Findings counts models sharing a top favorite.

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

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.

Click a name to open its card