A new technique lets AI models guess how different demographic groups within a population would actually answer a question - without ever training on real opinion data.
Researchers introduced Demographic Pluralism, an inference-time framework that skips fine-tuning and skips opinion-distribution training data entirely. It works by generating multiple simulated perspectives within demographically grounded subgroups, then combining them into an estimated population-level opinion distribution. Tested across four different model backbones on two benchmarks, GlobalOpinionQA and VITAL, it cut Jensen-Shannon distance - a measure of how far an estimated distribution strays from the real one - by 8.4% to 26.4% compared to Modular Pluralism, the existing baseline approach. The team also tried three ways of combining subgroup answers (weighted, equal-weighted, inverse-weighted) and found plain equal weighting won out, largely because giving a group more weight also inflated its own error.
This tackles a real gap in AI alignment: most methods treat a demographic bucket, say 'women under 30,' as if everyone inside it thinks alike. A model that can approximate the actual spread of opinion inside a group, without needing a bespoke dataset of real survey responses, would matter for anything from policy simulation to product research in places where that data is scarce or expensive to collect.
Still, it's an estimate of an estimate - synthetic perspectives standing in for people who were never actually asked, which is a comforting habit for an industry to lean on.