Taste is the thing people say AI can’t have. I gave a decision model the fifteen questions a painting teacher asks, ran them on 4,227 public-domain paintings for 58 cents, and checked its judgment against the curators of the Met and the Art Institute of Chicago.
Most art never gets a second look. Museums show a small slice of what they own, and a handful of curators decide which slice. Most people who paint get no expert feedback at all, because a teacher’s time is expensive and there aren’t many of them.
AI can describe any painting fluently. That isn’t the same as judging it. A chat model writes a confident paragraph about every picture, and you can’t tell whether its praise means anything, whether it would say the same thing tomorrow, or whether it agrees with anyone who knows the field.
Fewer people are even walking in: 17.7% of adults visited an art museum or gallery in 2022, down from 23.7% in 2017. The bottleneck isn’t the art. It’s the eye: someone to look at everything, say what’s working, and point you toward what you’d love.
Taste is the hardest test I could think of for a decision model. Contracts and recalls have answer keys; art supposedly doesn’t. But two great museums have each quietly published one: the small set of works they flag as highlights. That makes “does the model have taste?” a question with a measurable answer.
The ideaDon’t ask a model whether a painting is good. Break taste into the specific judgments an expert makes, score each one with its uncertainty, compare paintings head to head, test the result against curators, and let a writer argue with the numbers where the rubric falls short.
Sources are listed at the end of the page.
Each score is a distribution across five levels: darker means the model put more probability there, and the red tick is its answer. The box marks where it says your eye lands first. The critique uses the scores as evidence, then argues with them.
The best single clue is the most taste-like question: would this stop you across a room? Tidiness doesn’t help at all. A clear focal point and economy score no better than a coin flip.
Number on each tile: the model’s curator-likeness score, out of 100. Tap any painting for its full scorecard.
Look at who’s in this row: Korean ancestor portraits, Buddhist icons, Chinese scrolls, Whistler’s nocturnes, a Rousseau forest at dusk. The model’s eye is European and it likes things loud. It underrates quiet, tonal and non-Western painting, the way an 1860s academy jury would have. A rubric is a point of view, and this one has a blind spot you can see.
Paste a painting into a chat model and you’ll get a thoughtful paragraph. You’ll get one for every painting, all equally confident, and nothing to tell you whether the praise means anything. Four things here a chat answer can’t give you:
Words still need a language model. The critiques on this page were written by Claude from the model’s scores, which is the split I use everywhere: decision models for judgments, language models for words.
Fourteen pairs, each matched on overall quality and different in one sharp way, so your choices reveal what you respond to rather than which painting is better. The page learns as you go, in your browser; nothing is sent anywhere.
The same 455 paintings, re-ranked by different ways of seeing. The top 24 are outlined. Tap any painting for its scorecard.
The model isn’t a curator or a teacher. It’s a first look at everything, so people spend their time on the paintings and questions that need them. Change the assumptions to see the math.
Curators, teachers and painters keep the judgment. The model takes the part no one has time for: looking at everything, every time, the same way.
It can’t tell you what a painting means, why it mattered when it was made, or whether a broken rule was the whole point. It judges each picture alone, without history. Every place on this page where it is clearly wrong (Whistler, Bruegel, Mondrian) is a place where a person had to step in.
# One request per painting: the image goes in once, all questions read it POST https://api.typesafe.ai/v1/systemone TypeSafe-Danger-Zone-Images: true { "model": "jev-1.14.0", "state": { "art": "<painting>" }, "questions": { "presence": { "type": "score", "instructions": "How much presence does `art` have: would it stop you across a gallery and hold your attention?", "criteria": [ "You would walk past without noticing.", "Mildly interesting.", "Worth a look.", "Arresting; it holds you.", "Unforgettable; it stops you from across the room." ] }, … 14 more qualities, a genre and a box … } } # → { "presence": { "score": 3.4, "probabilities": { "0": 0, "1": 0.01, "2": 0.08, "3": 0.42, "4": 0.49 } }, … }
Famous paintings may be recognized rather than judged, which would flatter the curator test; I haven’t separated the two. Highlight flags are one curatorial signal among many, and the museums choose them for teaching and visitor flow as well as quality. I didn’t run a chat model head to head on the same 4,227 paintings, so the comparison above is about what each kind of answer allows, not a measured accuracy gap. The game learns from fourteen choices, which is a sketch of a taste, not a portrait. Images and boxes are TypeSafe “danger zone” features, released as previews and likely to change.
I asked for a box around each painting’s weakest passage. The model’s confidence in that box had a median of 0.05, so I dropped it: it can say where you look, not what’s weak. I also tried to trace an eye path by blanking each spot the model picked and asking where the eye goes next. It kept looking at the hole I’d cut.