Synthetic survey respondents from real census data, with a calibration score (synthetic-audience.noniner.io)

A new method for creating synthetic survey respondents using real US Census and World Bank microdata has been announced, allowing researchers and companies to gather insights without traditional panels. This innovative approach utilizes weighted profiles based on demographics such as age, sex, education, and income, conditioning a language model on each profile to generate realistic responses. Unlike conventional methods that rely on guesses or invented personas, this system presents a way to obtain data that is tailored to represent genuine population structures. The significance of this development lies in its ability to provide accurate, timely feedback with a known calibration score, enhancing the decision-making processes for businesses and researchers. By validating synthetic responses against historical data, including the 2020 US presidential vote, the method has shown a high level of accuracy with an average error within 1-7 points across various public-opinion benchmarks. This technology allows users to test market concepts and gauge audience sentiments effectively, while also acknowledging limitations; it cautions against using synthetic data as definitive measures, especially for high-stakes decisions where diverse views are critical. Overall, this advancement promises a more reliable alternative for data-driven insights in the AI/ML landscape.