Wang, Morgenstern & Dickerson (2025) measure demographic misportrayal and flattening, a different failure axis than respondent coverage, and so do not corroborate the respondent-coverage-gap preprint

llm-evaluationsilicon-samplingsurvey-methodscorroborationfairness

Wang, Morgenstern & Dickerson (Nature Machine Intelligence 7(3), 2025) argue and empirically show — across 3,200 human participants spanning 16 demographic identities — that LLMs prompted to adopt a demographic identity tend to misportray that identity (reproducing out-group stereotypes rather than in-group self-perception) and flatten within-group diversity. Their claim is about what a single simulated persona says about its assigned group: whether the content it produces is a faithful, varied representation of that group's members.

That is a different failure axis from the one Kumagae & Kaneko's respondent coverage gap measures, which asks whether a population of generated respondents collectively covers the distribution of real respondents' complete answer profiles (see claim-silicon-sampling-llms-undercover-human-respondents-despite-robustness-sweep). Reading the paper directly settles a lead the originating question left open: although Kumagae & Kaneko cite Wang et al. as related prior art, it does not independently corroborate respondent-coverage failure — it documents an adjacent problem (stereotyped, flattened content) for the same underlying practice of silicon sampling. Keeping the two distinct matters precisely because they are easy to conflate as "LLMs are bad at simulating people": one is about the believability of a persona's voice, the other about whether synthetic people, taken as a set, reach the real ones at all. This bears on the corroboration question as a resolved non-corroboration — a lead chased down and found to point elsewhere.