Two datasets, one answer
We used datasets that keep every individual rating, so each person could be analysed alone: 555 people who rated artworks (281,174 ratings in all) and 1,205 professional Eurovision jurors who ranked songs from 2016 to 2022. The two share nothing, neither the people, the medium nor the rating format, and they give almost the same split.
What decides one person’s rating
- 44% stable and their own
- 23% shared with everyone
- 33% noise, the same person disagreeing with themselves
No types, a continuous cloud
People’s tastes do not fall into a handful of kinds. Twenty presets recover only 60 to 80% of what a per-person model gets, and from about ten picks onward the per-person model does as well or better. About three in ten people prefer the more typical artworks of a genre and the rest the less typical ones, which is the distance idea in miniature, but pinning one person’s exact peak takes hundreds of judgements.
How many picks a model needs
About 20 picks buy a quarter of the possible personal gain, and about 80 picks buy 40%. On artworks, a model then names a person’s favourite of three 57.5% of the time after 20 picks and 60% after 80, against 52% for shared taste alone. People agree with their own earlier pick 62% of the time, so that ceiling is close.
Agreement lives at the extremes
Across 14 datasets and about 13,000 people, covering art, songs, jokes, food and products, shared taste sits almost entirely in the items that clearly stand out or fall short. Among middling items it nearly vanishes: 1 to 9% shared, against 8 to 23% for the full set.
Juries are rooms, not samples
Eurovision jurors on the same national jury agree with each other at 0.53, and with jurors from other countries at only 0.22. A jury verdict reflects its room. So when we say the eVolo jury’s awards sit at a distance of 1.43, that is one room’s distance, not a typical person’s.
What it means
Weeding out bad ideas can be generic. Choosing among good ones has to be personal, and it has to learn from that person’s own picks. That is why our selector models each judge separately and is scored only against real human choices.
Sources
- Feathers & Wax, How much creative taste varies between people, Oct 27, 2026, code in
research/cllm-taste. - Feathers & Wax, When taste converges, Oct 27, 2026.
- LAPIS artwork ratings, KU Leuven, 2025. Eurovision Song Contest jury rankings, 2016 to 2022. Jester joke ratings, UC Berkeley.
- Vessel et al., Stronger shared taste for natural aesthetic domains than for artifacts of human culture, Cognition, 2018.