Errors with a direction are the kind that survive aggregation. Noise washes out when you average over more synthetic respondents; a pull toward the modal answer does not. That is what the paper documents: twin responses came out more homogeneous than the responses of the people they stood in for, and they skewed toward demographic stereotypes [7].
The elicitation ledger is worth spelling out. Roughly 500 questions each for something over 2,000 people is on the order of a million answers, and both figures are stated as minimums, so that is a floor [19]. Many of those questions came from scales that psychology, economics and business research already lean on to characterize a person, and respondents also sat online tests of thought patterns and biases [11]. Toubia says the resulting dataset has been downloaded about 25,000 times [13].
Against that input, the twins were wrong on average about a quarter of the time [4]. Where the profiles did earn something was spread: given two people who rate their self-control 2 and 4, a model holding only demographics tends to answer 3 for both, while the twin answers 3 and 5, wrong in both cases but not flattened [6]. Olivier Toubia, a computational social scientist at Columbia Business School, allows that "there's some promise" while calling the twins' performance "overall a bit disappointing" [12]. For anyone estimating an effect size, recovered variance without accuracy is the more treacherous of the two failures, because a synthetic sample can reproduce the shape of a distribution while placing the wrong people in its tails. That inference is mine, not a result in the paper.
The thing a quarter-wrong average does not tell you is which quarter. It pools 19 experiments spanning questions as different as how people judge donors who give to both Republican and Democratic candidates and what they think about algorithmic hiring [3]. A substitution program does not run into the average hit rate; it runs into whether the surrogate errs the same way inside the one subgroup the study is about.
Hadi Hosseini of Penn State, who has watched LLM agents make health care decisions under scarcity, reports the same directional pull in his own domain: the models distort judgment toward something very rational and more reasonable than people actually are [14]. Two labs finding the same skew in unrelated tasks is the part of this that generalizes. Toubia's own caution points at his field rather than at the models. Social scientists tend to assume their scales capture the full range of human experience, he says, and predicting human behavior with a machine is incredibly hard [18].
