What a model can learn about one person

The claim

Thinking Machines calls for AIs “as diverse and distributed as people themselves.” Several 2026 papers measure how much of one person a model can actually learn, on images, on text and on people’s survey answers. Their results agree with each other more than the claims around them do.

Images: two points, and more where people differ

Researchers at Tübingen and KU Leuven collected 70,000 ratings of 5,000 images made by Flux 2 and Nano Banana, 15 raters to an image, from about 200 users. They trained a model on each person’s ratings and tested it on users it had never seen. On how often it picked the image a person preferred, it scored 0.663 against 0.643 for the best shared scorer, HPSv3. Averaged over everyone, the two were level: 0.680 against 0.677.

The gain is largest where people disagree. On pairs where two or more users split, a coin flip scores 50%. The model scored 61.4% for users it had seen and 55.2% for new ones.

The paper also asked six of those users to compare images. Images refined toward a person’s own taste ranked highest, at 1065 on an Elo scale. Images refined toward other participants’ tastes came next at 1038, and the unrefined images were at 1016. Images refined with shared aesthetic scorers ranked below the unrefined ones, at 959 and 922. A shared score used as a target made pictures worse. Part of the gain was shared across people, and tailoring to the viewer ranked 27 points higher still.

Text: a better ranker is not a better writer

Personalized RewardBench built pairs of answers that differ only in whether they follow one user’s stated rubric, both of high general quality. The best reward models peaked at 75.94% accuracy.

Rezk and colleagues showed that accuracy at ranking a person’s preferences is a poor guide to adapting a model to them. Across methods, ranking accuracy tracked downstream accuracy only weakly (Kendall’s tau 0.08 to 0.31), and methods 20 points apart on ranking produced nearly the same text. Simply placing a person’s past examples in the prompt beat every method that guides generation with a reward model, for models of 3 billion parameters and up, by about 3 points of ROUGE-1 at 7 billion.

Descriptions: the average, not the individual

Digital twins are models built from a description of a person, often their answers to a long survey. Wang and colleagues tested them in September 2026 across behavioural experiments and several models. Twins reproduced average human effects but gave little information about which individuals differ from the average. Newer models and richer descriptions improved some measures without reliably saving any human data. An earlier test with more than 500 questions per person found twins only modestly better than a plain model at predicting that person, with an average correlation of 0.20 (see What personal AI can cost).

What our own tests found

The same shape appears in ours. On public joke ratings, a selector that learned one person’s picks beat a shared judge by 2.9 points after 20 ratings, 5.1 after 40 and 6.4 after 80, and 7.4 when the options were all in the crowd’s top third (A selector that learns one person). On artworks a model named a person’s favourite of three 57.5% of the time after 20 picks and 60% after 80, against 52% for shared taste. People agree with their own earlier pick 62% of the time, so the ceiling is near (Taste is mostly personal).

We also tested the claim that a person’s preferred distance from the usual is what to learn. On 549 people rating artworks, we revealed 40 of a person’s ratings, learned their distance, and offered three artworks from the crowd’s top third that they had also rated. The selector picked their favourite 44.7% of the time, against 44.8% for the shared judge. Choosing by distance alone landed at 35%, near the 33% of chance. A second way of measuring distance gained 1.5 points, about twice what a person’s taste in art style added (1.0). Aiming at a person’s own distance rather than the average person’s moved ratings by 0.3 to 1.0 points out of 100. A lean toward the unusual is a small input, not the selector (A lean, not a sweet spot).

What the results have in common

• The gain over a shared model is a few points, and it grows with the number of the person’s own ratings.
• It sits where people disagree, which is also where a shared model has nothing to say.
• What a person chose predicts better than what is said about them.
• A model that ranks well is not thereby a model that writes or draws well for that person.

What it means for our work

This is the design we already use. A framework gives the model a way of thinking to generate from, as a description does. The person’s own keep and pass choices select, as ratings do. The evidence says to put the weight on the second: a description describes, and a choice discriminates. It also says what to ask of a selector: be scored on the pairs where people differ and against that person’s own picks, not against a population average, where a personal model and a shared one look the same.

It makes the distance idea a smaller claim. Where an idea sits relative to the usual answer is useful to show, and each idea carries that score. It does not pick for a person on its own, and the pictures above point to why: shared taste carries a large part of what any person likes, and the personal part is learned from that person’s choices.