A new open dataset aims to give AI models a more globally representative sense of human values — and the benchmarks to prove it works.
Researchers released PLURAL, a dataset of roughly 500,000 synthetic preference triplets built on top of the Integrated Values Survey, a nationally representative poll covering 92 countries. The team used a two-stage pipeline to convert survey responses into training-ready scenarios that carry the normative value signals of the original data. The initial release covers 20 countries. Training on PLURAL cut mean absolute error against target countries' cultural profiles by up to 27.7% compared to existing baselines. A blind human evaluation involving 176 participants in India, Brazil, and Japan found that PLURAL-aligned model responses felt more representative of their national values.
Most large language models are trained on English-language internet text, which skews heavily toward American and Western European perspectives. The result is models that treat those defaults as neutral — a problem that compounds as AI gets embedded in hiring tools, content moderation, and public services around the world. PLURAL gives researchers a concrete, scalable lever to test and correct that drift.
The dataset is available on Hugging Face, which lowers the barrier to use — though "culturally aligned" AI and "AI that tells people what they want to hear" are distinctions that will need its own rigorous scrutiny as adoption grows.