Synthetic respondents have moved from speculation to standard practice. Peer-reviewed studies in Political Analysis, the Journal of Marketing, Psychology & Marketing, and replications by EY, Harvard, MIT Sloan, and Qualtrics show that calibrated synthetic data now matches—and in some cases exceeds—traditional human-only research.
In a double-blind test, professional services firm EY took its annual Global Brand Survey—aimed at CEOs of US companies with $1B+ in revenue—and ran it twice: once through traditional fielding, once through 1,000 synthetic personas built by Aaru.
The synthetic survey returned 95% correlation with the real one. EY also recreated their annual Global Wealth Research Report in a single day, with 90%+ median correlation to the original six-month study.
— Toni Clayton-Hine, EY CMO. Reported in Solomon Partners (Sept 2025).
Read the case study →Four foundational papers from leading journals establish that calibrated synthetic data reproduces human survey responses with rigor.
The AI–human hybrid generates information-rich, coherent data that surpasses human-only data in depth and insightfulness, and matches human performance in theme generation. LLM hybrid recovered 77% of themes identified by human analysts.
DOI: 10.1177/00222429241276529 →Tested against 9,300 human responses across 57 personal-care surveys, the Semantic Similarity Rating method achieved 90% of human test-retest reliability. Distributional similarity to real data exceeded 0.85 (Kolmogorov–Smirnov).
Read on arXiv →The foundational “silicon samples” paper. GPT-3 conditioned on sociodemographic backstories accurately emulates response distributions across human subgroups, successfully replicating real survey results across diverse populations.
DOI: 10.1017/pan.2023.2 →Establishes formal academic guidelines for silicon sampling. Concludes synthetic samples hold particular promise in upstream parts of the research process: qualitative pretesting, pilot studies, and hypothesis generation.
DOI: 10.1002/mar.21982 →Cohen's D deviation between calibrated synthetic respondents and real humans across 11 identical survey questions. The takeaway: calibration is what closes the gap. McLean (Feb 2026) replication of Paxton & Yang.
Read the replication study →Alignment with human survey data and 85% distributional similarity across concept and pricing studies, with carefully calibrated synthetic consumers.
Read the analysis →Faster research timelines. Dollar Shave Club: month-long studies in weeks. Gabb: weeks of work in hours, with rank-order alignment between synthetic and human respondents.
Read the case studies →“LLMs, used carefully, can function as synthetic focus groups—producing realistic and accurate preferences in a fraction of the time and cost.”
“LLM-generated synthetic respondents—digital twins—enable rapid concept testing and AI-moderated interviews for qualitative research at scale.”
“Synthetic data could be better than real data.”
“Synthetic data is as good as real—next comes synthetic strategy.”
Generate personas, run a survey or interview, and see what synthetic respondents reveal — in minutes.