YouGov tests digital twins and calls for validation standards

YouGov partnered with simulation platform OriginalVoices to build AI-based “E-twins” of 1,008 government ID-verified panel members, then asked the simulations and their human counterparts the same new questions. A recontact survey received responses from 693 participants within 48 hours, allowing researchers to assess individual twins rather than relying on aggregate comparisons.

Adding panelists’ ratings of 4,000+ brands and entities increased accuracy from 78% to 90%, measured as agreement within one point on a five-point scale. The twins achieved 90% on questions about fiction books, a topic excluded from their training data. Panelists judged their twins’ open-ended answers as describing their own views 75% of the time, with higher-confidence responses receiving stronger endorsement.

YouGov called for shared verification and validation standards, including testing against individual people, evaluating unseen topics, publishing methods and failures, and predicting real-world outcomes publicly. It warned that public datasets may overlap with language-model training material, while aggregate results can conceal stereotyping or responses that converge toward averages.

The study did not establish how consistently people repeat their own answers, leaving an important benchmark unresolved. YouGov plans to measure that in its next research iteration to better contextualize the twins’ 58% exact-match rate.